diff --git a/.github/workflows/pr_tests.yml b/.github/workflows/pr_tests.yml index dc1c482aa0..55a9bd68de 100644 --- a/.github/workflows/pr_tests.yml +++ b/.github/workflows/pr_tests.yml @@ -60,6 +60,7 @@ jobs: run: | python -m pip install -e .[quality,test] python -m pip install git+https://github.com/huggingface/accelerate + python -m pip install -U git+https://github.com/huggingface/transformers - name: Environment run: | @@ -127,6 +128,7 @@ jobs: ${CONDA_RUN} python -m pip install -e .[quality,test] ${CONDA_RUN} python -m pip install --pre torch==${MPS_TORCH_VERSION} --extra-index-url https://download.pytorch.org/whl/test/cpu ${CONDA_RUN} python -m pip install git+https://github.com/huggingface/accelerate + ${CONDA_RUN} python -m pip install -U git+https://github.com/huggingface/transformers - name: Environment shell: arch -arch arm64 bash {0} diff --git a/.github/workflows/push_tests.yml b/.github/workflows/push_tests.yml index 2beb05e8ea..4bab00b7ee 100644 --- a/.github/workflows/push_tests.yml +++ b/.github/workflows/push_tests.yml @@ -62,6 +62,7 @@ jobs: run: | python -m pip install -e .[quality,test] python -m pip install git+https://github.com/huggingface/accelerate + python -m pip install -U git+https://github.com/huggingface/transformers - name: Environment run: | @@ -131,6 +132,7 @@ jobs: run: | python -m pip install -e .[quality,test,training] python -m pip install git+https://github.com/huggingface/accelerate + python -m pip install -U git+https://github.com/huggingface/transformers - name: Environment run: | diff --git a/docs/source/_toctree.yml b/docs/source/_toctree.yml index c143dab9f5..bf23d363a8 100644 --- a/docs/source/_toctree.yml +++ b/docs/source/_toctree.yml @@ -106,10 +106,14 @@ title: "Score SDE VE" - local: api/pipelines/stable_diffusion title: "Stable Diffusion" + - local: api/pipelines/stable_diffusion_safe + title: "Safe Stable Diffusion" - local: api/pipelines/stochastic_karras_ve title: "Stochastic Karras VE" - local: api/pipelines/dance_diffusion title: "Dance Diffusion" + - local: api/pipelines/versatile_diffusion + title: "Versatile Diffusion" - local: api/pipelines/vq_diffusion title: "VQ Diffusion" - local: api/pipelines/repaint diff --git a/docs/source/api/pipelines/alt_diffusion.mdx b/docs/source/api/pipelines/alt_diffusion.mdx index 84dda88dcb..4a75bc09bf 100644 --- a/docs/source/api/pipelines/alt_diffusion.mdx +++ b/docs/source/api/pipelines/alt_diffusion.mdx @@ -32,7 +32,7 @@ The abstract of the paper is the following: - *Run AltDiffusion* -AltDiffusion can be tested very easily with the [`AltDiffusionPipeline`], [`AltDiffusionImg2ImgPipeline`] and the `"BAAI/AltDiffusion"` checkpoint exactly in the same way it is shown in the [Conditional Image Generation Guide](./using-diffusers/conditional_image_generation) and the [Image-to-Image Generation Guide](./using-diffusers/img2img). +AltDiffusion can be tested very easily with the [`AltDiffusionPipeline`], [`AltDiffusionImg2ImgPipeline`] and the `"BAAI/AltDiffusion-m9"` checkpoint exactly in the same way it is shown in the [Conditional Image Generation Guide](./using-diffusers/conditional_image_generation) and the [Image-to-Image Generation Guide](./using-diffusers/img2img). - *How to load and use different schedulers.* @@ -42,12 +42,12 @@ To use a different scheduler, you can either change it via the [`ConfigMixin.fro ```python >>> from diffusers import AltDiffusionPipeline, EulerDiscreteScheduler ->>> pipeline = AltDiffusionPipeline.from_pretrained("BAAI/AltDiffusion") +>>> pipeline = AltDiffusionPipeline.from_pretrained("BAAI/AltDiffusion-m9") >>> pipeline.scheduler = EulerDiscreteScheduler.from_config(pipeline.scheduler.config) >>> # or ->>> euler_scheduler = EulerDiscreteScheduler.from_pretrained("BAAI/AltDiffusion", subfolder="scheduler") ->>> pipeline = AltDiffusionPipeline.from_pretrained("BAAI/AltDiffusion", scheduler=euler_scheduler) +>>> euler_scheduler = EulerDiscreteScheduler.from_pretrained("BAAI/AltDiffusion-m9", subfolder="scheduler") +>>> pipeline = AltDiffusionPipeline.from_pretrained("BAAI/AltDiffusion-m9", scheduler=euler_scheduler) ``` @@ -61,7 +61,7 @@ If you want to use all possible use cases in a single `DiffusionPipeline` we rec ... AltDiffusionImg2ImgPipeline, ... ) ->>> text2img = AltDiffusionPipeline.from_pretrained("BAAI/AltDiffusion") +>>> text2img = AltDiffusionPipeline.from_pretrained("BAAI/AltDiffusion-m9") >>> img2img = AltDiffusionImg2ImgPipeline(**text2img.components) >>> # now you can use text2img(...) and img2img(...) just like the call methods of each respective pipeline diff --git a/docs/source/api/pipelines/overview.mdx b/docs/source/api/pipelines/overview.mdx index 74c44fbccd..c43f09d66d 100644 --- a/docs/source/api/pipelines/overview.mdx +++ b/docs/source/api/pipelines/overview.mdx @@ -58,7 +58,11 @@ available a colab notebook to directly try them out. | [stable_diffusion](./api/pipelines/stable_diffusion) | [**Stable Diffusion**](https://stability.ai/blog/stable-diffusion-public-release) | Text-to-Image Generation | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/training_example.ipynb) | [stable_diffusion](./api/pipelines/stable_diffusion) | [**Stable Diffusion**](https://stability.ai/blog/stable-diffusion-public-release) | Image-to-Image Text-Guided Generation | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/image_2_image_using_diffusers.ipynb) | [stable_diffusion](./api/pipelines/stable_diffusion) | [**Stable Diffusion**](https://stability.ai/blog/stable-diffusion-public-release) | Text-Guided Image Inpainting | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/in_painting_with_stable_diffusion_using_diffusers.ipynb) -| [stochastic_karras_ve](./api/pipelines/stochastic_karras_ve) | [**Elucidating the Design Space of Diffusion-Based Generative Models**](https://arxiv.org/abs/2206.00364) | Unconditional Image Generation | +| [stable_diffusion_safe](./api/pipelines/stable_diffusion_safe) | [**Safe Stable Diffusion**](https://arxiv.org/abs/2211.05105) | Text-Guided Generation | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/ml-research/safe-latent-diffusion/blob/main/examples/Safe%20Latent%20Diffusion.ipynb) +| [stochastic_karras_ve](./api/pipelines/stochastic_karras_ve) | [**Elucidating the Design Space of Diffusion-Based Generative Models**](https://arxiv.org/abs/2206.00364) | Unconditional Image Generation | +| [versatile_diffusion](./api/pipelines/versatile_diffusion) | [Versatile Diffusion: Text, Images and Variations All in One Diffusion Model](https://arxiv.org/abs/2211.08332) | Text-to-Image Generation | +| [versatile_diffusion](./api/pipelines/versatile_diffusion) | [Versatile Diffusion: Text, Images and Variations All in One Diffusion Model](https://arxiv.org/abs/2211.08332) | Image Variations Generation | +| [versatile_diffusion](./api/pipelines/versatile_diffusion) | [Versatile Diffusion: Text, Images and Variations All in One Diffusion Model](https://arxiv.org/abs/2211.08332) | Dual Image and Text Guided Generation | | [vq_diffusion](./api/pipelines/vq_diffusion) | [Vector Quantized Diffusion Model for Text-to-Image Synthesis](https://arxiv.org/abs/2111.14822) | Text-to-Image Generation | diff --git a/docs/source/api/pipelines/stable_diffusion.mdx b/docs/source/api/pipelines/stable_diffusion.mdx index 8b551f7a3b..9884cbb207 100644 --- a/docs/source/api/pipelines/stable_diffusion.mdx +++ b/docs/source/api/pipelines/stable_diffusion.mdx @@ -88,3 +88,10 @@ If you want to use all possible use cases in a single `DiffusionPipeline` you ca - __call__ - enable_attention_slicing - disable_attention_slicing + + +## StableDiffusionImageVariationPipeline +[[autodoc]] StableDiffusionImageVariationPipeline + - __call__ + - enable_attention_slicing + - disable_attention_slicing diff --git a/docs/source/api/pipelines/stable_diffusion_safe.mdx b/docs/source/api/pipelines/stable_diffusion_safe.mdx new file mode 100644 index 0000000000..81fc59d392 --- /dev/null +++ b/docs/source/api/pipelines/stable_diffusion_safe.mdx @@ -0,0 +1,90 @@ + + +# Safe Stable Diffusion + +Safe Stable Diffusion was proposed in [Safe Latent Diffusion: Mitigating Inappropriate Degeneration in Diffusion Models](https://arxiv.org/abs/2211.05105) and mitigates the well known issue that models like Stable Diffusion that are trained on unfiltered, web-crawled datasets tend to suffer from inappropriate degeneration. For instance Stable Diffusion may unexpectedly generate nudity, violence, images depicting self-harm, or otherwise offensive content. +Safe Stable Diffusion is an extension to the Stable Diffusion that drastically reduces content like this. + +The abstract of the paper is the following: + +*Text-conditioned image generation models have recently achieved astonishing results in image quality and text alignment and are consequently employed in a fast-growing number of applications. Since they are highly data-driven, relying on billion-sized datasets randomly scraped from the internet, they also suffer, as we demonstrate, from degenerated and biased human behavior. In turn, they may even reinforce such biases. To help combat these undesired side effects, we present safe latent diffusion (SLD). Specifically, to measure the inappropriate degeneration due to unfiltered and imbalanced training sets, we establish a novel image generation test bed-inappropriate image prompts (I2P)-containing dedicated, real-world image-to-text prompts covering concepts such as nudity and violence. As our exhaustive empirical evaluation demonstrates, the introduced SLD removes and suppresses inappropriate image parts during the diffusion process, with no additional training required and no adverse effect on overall image quality or text alignment.* + + +*Overview*: + +| Pipeline | Tasks | Colab | Demo +|---|---|:---:|:---:| +| [pipeline_stable_diffusion_safe.py](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/stable_diffusion_safe/pipeline_stable_diffusion_safe.py) | *Text-to-Image Generation* | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/ml-research/safe-latent-diffusion/blob/main/examples/Safe%20Latent%20Diffusion.ipynb) | - + +## Tips + +- Safe Stable Diffusion may also be used with weights of [Stable Diffusion](./api/pipelines/stable_diffusion). + +### Run Safe Stable Diffusion + +Safe Stable Diffusion can be tested very easily with the [`StableDiffusionPipelineSafe`], and the `"AIML-TUDA/stable-diffusion-safe"` checkpoint exactly in the same way it is shown in the [Conditional Image Generation Guide](./using-diffusers/conditional_image_generation). + +### Interacting with the Safety Concept + +To check and edit the currently used safety concept, use the `safety_concept` property of [`StableDiffusionPipelineSafe`] +```python +>>> from diffusers import StableDiffusionPipelineSafe + +>>> pipeline = StableDiffusionPipelineSafe.from_pretrained("AIML-TUDA/stable-diffusion-safe") +>>> pipeline.safety_concept +``` +For each image generation the active concept is also contained in [`StableDiffusionSafePipelineOutput`]. + +### Using pre-defined safety configurations + +You may use the 4 configurations defined in the [Safe Latent Diffusion paper](https://arxiv.org/abs/2211.05105) as follows: + +```python +>>> from diffusers import StableDiffusionPipelineSafe +>>> from diffusers.pipelines.stable_diffusion_safe import SafetyConfig + +>>> pipeline = StableDiffusionPipelineSafe.from_pretrained("AIML-TUDA/stable-diffusion-safe") +>>> prompt = "the four horsewomen of the apocalypse, painting by tom of finland, gaston bussiere, craig mullins, j. c. leyendecker" +>>> out = pipeline(prompt=prompt, **SafetyConfig.MAX) +``` + +The following configurations are available: `SafetyConfig.WEAK`, `SafetyConfig.MEDIUM`, `SafetyConfig.STRONg`, and `SafetyConfig.MAX`. + +### How to load and use different schedulers. + +The safe stable diffusion pipeline uses [`PNDMScheduler`] scheduler by default. But `diffusers` provides many other schedulers that can be used with the stable diffusion pipeline such as [`DDIMScheduler`], [`LMSDiscreteScheduler`], [`EulerDiscreteScheduler`], [`EulerAncestralDiscreteScheduler`] etc. +To use a different scheduler, you can either change it via the [`ConfigMixin.from_config`] method or pass the `scheduler` argument to the `from_pretrained` method of the pipeline. For example, to use the [`EulerDiscreteScheduler`], you can do the following: + +```python +>>> from diffusers import StableDiffusionPipelineSafe, EulerDiscreteScheduler + +>>> pipeline = StableDiffusionPipelineSafe.from_pretrained("AIML-TUDA/stable-diffusion-safe") +>>> pipeline.scheduler = EulerDiscreteScheduler.from_config(pipeline.scheduler.config) + +>>> # or +>>> euler_scheduler = EulerDiscreteScheduler.from_pretrained("AIML-TUDA/stable-diffusion-safe", subfolder="scheduler") +>>> pipeline = StableDiffusionPipelineSafe.from_pretrained( +... "AIML-TUDA/stable-diffusion-safe", scheduler=euler_scheduler +... ) +``` + + +## StableDiffusionSafePipelineOutput +[[autodoc]] pipelines.stable_diffusion_safe.StableDiffusionSafePipelineOutput + +## StableDiffusionPipelineSafe +[[autodoc]] StableDiffusionPipelineSafe + - __call__ + - enable_attention_slicing + - disable_attention_slicing + diff --git a/docs/source/api/pipelines/versatile_diffusion.mdx b/docs/source/api/pipelines/versatile_diffusion.mdx new file mode 100644 index 0000000000..f557c5b0aa --- /dev/null +++ b/docs/source/api/pipelines/versatile_diffusion.mdx @@ -0,0 +1,73 @@ + + +# VersatileDiffusion + +VersatileDiffusion was proposed in [Versatile Diffusion: Text, Images and Variations All in One Diffusion Model](https://arxiv.org/abs/2211.08332) by Xingqian Xu, Zhangyang Wang, Eric Zhang, Kai Wang, Humphrey Shi . + +The abstract of the paper is the following: + +*The recent advances in diffusion models have set an impressive milestone in many generation tasks. Trending works such as DALL-E2, Imagen, and Stable Diffusion have attracted great interest in academia and industry. Despite the rapid landscape changes, recent new approaches focus on extensions and performance rather than capacity, thus requiring separate models for separate tasks. In this work, we expand the existing single-flow diffusion pipeline into a multi-flow network, dubbed Versatile Diffusion (VD), that handles text-to-image, image-to-text, image-variation, and text-variation in one unified model. Moreover, we generalize VD to a unified multi-flow multimodal diffusion framework with grouped layers, swappable streams, and other propositions that can process modalities beyond images and text. Through our experiments, we demonstrate that VD and its underlying framework have the following merits: a) VD handles all subtasks with competitive quality; b) VD initiates novel extensions and applications such as disentanglement of style and semantic, image-text dual-guided generation, etc.; c) Through these experiments and applications, VD provides more semantic insights of the generated outputs.* + +## Tips + +- VersatileDiffusion is conceptually very similar as [Stable Diffusion](./api/pipelines/stable_diffusion), but instead of providing just a image data stream conditioned on text, VersatileDiffusion provides both a image and text data stream and can be conditioned on both text and image. + +### *Run VersatileDiffusion* + +You can both load the memory intensive "all-in-one" [`VersatileDiffusionPipeline`] that can run all tasks +with the same class as shown in [`VersatileDiffusionPipeline.text_to_image`], [`VersatileDiffusionPipeline.image_variation`], and [`VersatileDiffusionPipeline.dual_guided`] + +**or** + +You can run the individual pipelines which are much more memory efficient: + +- *Text-to-Image*: [`VersatileDiffusionTextToImagePipeline.__call__`] +- *Image Variation*: [`VersatileDiffusionImageVariationPipeline.__call__`] +- *Dual Text and Image Guided Generation*: [`VersatileDiffusionDualGuidedPipeline.__call__`] + +### *How to load and use different schedulers.* + +The versatile diffusion pipelines uses [`DDIMScheduler`] scheduler by default. But `diffusers` provides many other schedulers that can be used with the alt diffusion pipeline such as [`PNDMScheduler`], [`LMSDiscreteScheduler`], [`EulerDiscreteScheduler`], [`EulerAncestralDiscreteScheduler`] etc. +To use a different scheduler, you can either change it via the [`ConfigMixin.from_config`] method or pass the `scheduler` argument to the `from_pretrained` method of the pipeline. For example, to use the [`EulerDiscreteScheduler`], you can do the following: + +```python +>>> from diffusers import VersatileDiffusionPipeline, EulerDiscreteScheduler + +>>> pipeline = VersatileDiffusionPipeline.from_pretrained("shi-labs/versatile-diffusion") +>>> pipeline.scheduler = EulerDiscreteScheduler.from_config(pipeline.scheduler.config) + +>>> # or +>>> euler_scheduler = EulerDiscreteScheduler.from_pretrained("shi-labs/versatile-diffusion", subfolder="scheduler") +>>> pipeline = VersatileDiffusionPipeline.from_pretrained("shi-labs/versatile-diffusion", scheduler=euler_scheduler) +``` + +## VersatileDiffusionPipeline +[[autodoc]] VersatileDiffusionPipeline + +## VersatileDiffusionTextToImagePipeline +[[autodoc]] VersatileDiffusionTextToImagePipeline + - __call__ + - enable_attention_slicing + - disable_attention_slicing + +## VersatileDiffusionImageVariationPipeline +[[autodoc]] VersatileDiffusionImageVariationPipeline + - __call__ + - enable_attention_slicing + - disable_attention_slicing + +## VersatileDiffusionDualGuidedPipeline +[[autodoc]] VersatileDiffusionDualGuidedPipeline + - __call__ + - enable_attention_slicing + - disable_attention_slicing diff --git a/docs/source/index.mdx b/docs/source/index.mdx index e4722bec68..09cc59fda9 100644 --- a/docs/source/index.mdx +++ b/docs/source/index.mdx @@ -48,7 +48,11 @@ available a colab notebook to directly try them out. | [stable_diffusion](./api/pipelines/stable_diffusion) | [**Stable Diffusion**](https://stability.ai/blog/stable-diffusion-public-release) | Text-to-Image Generation | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/training_example.ipynb) | [stable_diffusion](./api/pipelines/stable_diffusion) | [**Stable Diffusion**](https://stability.ai/blog/stable-diffusion-public-release) | Image-to-Image Text-Guided Generation | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/image_2_image_using_diffusers.ipynb) | [stable_diffusion](./api/pipelines/stable_diffusion) | [**Stable Diffusion**](https://stability.ai/blog/stable-diffusion-public-release) | Text-Guided Image Inpainting | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/in_painting_with_stable_diffusion_using_diffusers.ipynb) +| [stable_diffusion_safe](./api/pipelines/stable_diffusion_safe) | [**Safe Stable Diffusion**](https://arxiv.org/abs/2211.05105) | Text-Guided Generation | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/ml-research/safe-latent-diffusion/blob/main/examples/Safe%20Latent%20Diffusion.ipynb) | [stochastic_karras_ve](./api/pipelines/stochastic_karras_ve) | [**Elucidating the Design Space of Diffusion-Based Generative Models**](https://arxiv.org/abs/2206.00364) | Unconditional Image Generation | +| [versatile_diffusion](./api/pipelines/versatile_diffusion) | [Versatile Diffusion: Text, Images and Variations All in One Diffusion Model](https://arxiv.org/abs/2211.08332) | Text-to-Image Generation | +| [versatile_diffusion](./api/pipelines/versatile_diffusion) | [Versatile Diffusion: Text, Images and Variations All in One Diffusion Model](https://arxiv.org/abs/2211.08332) | Image Variations Generation | +| [versatile_diffusion](./api/pipelines/versatile_diffusion) | [Versatile Diffusion: Text, Images and Variations All in One Diffusion Model](https://arxiv.org/abs/2211.08332) | Dual Image and Text Guided Generation | | [vq_diffusion](./api/pipelines/vq_diffusion) | [Vector Quantized Diffusion Model for Text-to-Image Synthesis](https://arxiv.org/abs/2111.14822) | Text-to-Image Generation | **Note**: Pipelines are simple examples of how to play around with the diffusion systems as described in the corresponding papers. diff --git a/examples/community/README.md b/examples/community/README.md index dc35d36a95..108f6f95f1 100644 --- a/examples/community/README.md +++ b/examples/community/README.md @@ -21,6 +21,8 @@ If a community doesn't work as expected, please open an issue and ping the autho | Multilingual Stable Diffusion| Stable Diffusion Pipeline that supports prompts in 50 different languages. | [Multilingual Stable Diffusion](#multilingual-stable-diffusion-pipeline) | - | [Juan Carlos Piñeros](https://github.com/juancopi81) | | Image to Image Inpainting Stable Diffusion | Stable Diffusion Pipeline that enables the overlaying of two images and subsequent inpainting| [Image to Image Inpainting Stable Diffusion](#image-to-image-inpainting-stable-diffusion) | - | [Alex McKinney](https://github.com/vvvm23) | | Text Based Inpainting Stable Diffusion | Stable Diffusion Inpainting Pipeline that enables passing a text prompt to generate the mask for inpainting| [Text Based Inpainting Stable Diffusion](#image-to-image-inpainting-stable-diffusion) | - | [Dhruv Karan](https://github.com/unography) | +| Bit Diffusion | Diffusion on discrete data | [Bit Diffusion](#bit-diffusion) | - |[Stuti R.](https://github.com/kingstut) | +| K-Diffusion Stable Diffusion | Run Stable Diffusion with any of [K-Diffusion's samplers](https://github.com/crowsonkb/k-diffusion/blob/master/k_diffusion/sampling.py) | [Stable Diffusion with K Diffusion](#stable-diffusion-with-k-diffusion) | - | [Patrick von Platen](https://github.com/patrickvonplaten/) | @@ -343,7 +345,6 @@ out = pipe( ) ``` - ### Composable Stable diffusion [Composable Stable Diffusion](https://energy-based-model.github.io/Compositional-Visual-Generation-with-Composable-Diffusion-Models/) proposes conjunction and negation (negative prompts) operators for compositional generation with conditional diffusion models. @@ -370,7 +371,7 @@ def dummy(images, **kwargs): pipe.safety_checker = dummy images = [] -generator = th.Generator("cuda").manual_seed(0) +generator = torch.Generator("cuda").manual_seed(0) seed = 0 prompt = "a forest | a camel" @@ -399,6 +400,7 @@ import requests from PIL import Image from io import BytesIO import torch +import os from diffusers import DiffusionPipeline, DDIMScheduler has_cuda = torch.cuda.is_available() device = torch.device('cpu' if not has_cuda else 'cuda') @@ -423,6 +425,7 @@ res = pipe.train( num_inference_steps=50, generator=generator) res = pipe(alpha=1) +os.makedirs("imagic", exist_ok=True) image = res.images[0] image.save('./imagic/imagic_image_alpha_1.png') res = pipe(alpha=1.5) @@ -652,4 +655,74 @@ prompt = "a cup" # the masked out region will be replaced with this with autocast("cuda"): image = pipe(image=image, text=text, prompt=prompt).images[0] -``` \ No newline at end of file +``` + +### Bit Diffusion +Based https://arxiv.org/abs/2208.04202, this is used for diffusion on discrete data - eg, discreate image data, DNA sequence data. An unconditional discreate image can be generated like this: + +```python +from diffusers import DiffusionPipeline +pipe = DiffusionPipeline.from_pretrained("google/ddpm-cifar10-32", custom_pipeline="bit_diffusion") +image = pipe().images[0] + +``` + +### Stable Diffusion with K Diffusion + +Make sure you have @crowsonkb's https://github.com/crowsonkb/k-diffusion installed: + +``` +pip install k-diffusion +``` + +You can use the community pipeline as follows: + +```python +from diffusers import DiffusionPipeline + +pipe = DiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-4", custom_pipeline="sd_text2img_k_diffusion") +pipe = pipe.to("cuda") + +prompt = "an astronaut riding a horse on mars" +pipe.set_sampler("sample_heun") +generator = torch.Generator(device="cuda").manual_seed(seed) +image = pipe(prompt, generator=generator, num_inference_steps=20).images[0] + +image.save("./astronaut_heun_k_diffusion.png") +``` + +To make sure that K Diffusion and `diffusers` yield the same results: + +**Diffusers**: +```python +from diffusers import DiffusionPipeline, EulerDiscreteScheduler + +seed = 33 + +pipe = DiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-4") +pipe.scheduler = EulerDiscreteScheduler.from_config(pipe.scheduler.config) +pipe = pipe.to("cuda") + +generator = torch.Generator(device="cuda").manual_seed(seed) +image = pipe(prompt, generator=generator, num_inference_steps=50).images[0] +``` + +![diffusers_euler](https://huggingface.co/datasets/patrickvonplaten/images/resolve/main/k_diffusion/astronaut_euler.png) + +**K Diffusion**: +```python +from diffusers import DiffusionPipeline, EulerDiscreteScheduler + +seed = 33 + +pipe = DiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-4", custom_pipeline="sd_text2img_k_diffusion") +pipe.scheduler = EulerDiscreteScheduler.from_config(pipe.scheduler.config) +pipe = pipe.to("cuda") + +pipe.set_sampler("sample_euler") +generator = torch.Generator(device="cuda").manual_seed(seed) +image = pipe(prompt, generator=generator, num_inference_steps=50).images[0] +``` + +![diffusers_euler](https://huggingface.co/datasets/patrickvonplaten/images/resolve/main/k_diffusion/astronaut_euler_k_diffusion.png) + diff --git a/examples/community/bit_diffusion.py b/examples/community/bit_diffusion.py new file mode 100644 index 0000000000..c0be3a13ad --- /dev/null +++ b/examples/community/bit_diffusion.py @@ -0,0 +1,263 @@ +from typing import Optional, Tuple, Union + +import torch + +from diffusers import DDIMScheduler, DDPMScheduler, DiffusionPipeline, UNet2DConditionModel +from diffusers.pipeline_utils import ImagePipelineOutput +from diffusers.schedulers.scheduling_ddim import DDIMSchedulerOutput +from diffusers.schedulers.scheduling_ddpm import DDPMSchedulerOutput +from einops import rearrange, reduce + + +BITS = 8 + + +# convert to bit representations and back taken from https://github.com/lucidrains/bit-diffusion/blob/main/bit_diffusion/bit_diffusion.py +def decimal_to_bits(x, bits=BITS): + """expects image tensor ranging from 0 to 1, outputs bit tensor ranging from -1 to 1""" + device = x.device + + x = (x * 255).int().clamp(0, 255) + + mask = 2 ** torch.arange(bits - 1, -1, -1, device=device) + mask = rearrange(mask, "d -> d 1 1") + x = rearrange(x, "b c h w -> b c 1 h w") + + bits = ((x & mask) != 0).float() + bits = rearrange(bits, "b c d h w -> b (c d) h w") + bits = bits * 2 - 1 + return bits + + +def bits_to_decimal(x, bits=BITS): + """expects bits from -1 to 1, outputs image tensor from 0 to 1""" + device = x.device + + x = (x > 0).int() + mask = 2 ** torch.arange(bits - 1, -1, -1, device=device, dtype=torch.int32) + + mask = rearrange(mask, "d -> d 1 1") + x = rearrange(x, "b (c d) h w -> b c d h w", d=8) + dec = reduce(x * mask, "b c d h w -> b c h w", "sum") + return (dec / 255).clamp(0.0, 1.0) + + +# modified scheduler step functions for clamping the predicted x_0 between -bit_scale and +bit_scale +def ddim_bit_scheduler_step( + self, + model_output: torch.FloatTensor, + timestep: int, + sample: torch.FloatTensor, + eta: float = 0.0, + use_clipped_model_output: bool = True, + generator=None, + return_dict: bool = True, +) -> Union[DDIMSchedulerOutput, Tuple]: + """ + Predict the sample at the previous timestep by reversing the SDE. Core function to propagate the diffusion + process from the learned model outputs (most often the predicted noise). + Args: + model_output (`torch.FloatTensor`): direct output from learned diffusion model. + timestep (`int`): current discrete timestep in the diffusion chain. + sample (`torch.FloatTensor`): + current instance of sample being created by diffusion process. + eta (`float`): weight of noise for added noise in diffusion step. + use_clipped_model_output (`bool`): TODO + generator: random number generator. + return_dict (`bool`): option for returning tuple rather than DDIMSchedulerOutput class + Returns: + [`~schedulers.scheduling_utils.DDIMSchedulerOutput`] or `tuple`: + [`~schedulers.scheduling_utils.DDIMSchedulerOutput`] if `return_dict` is True, otherwise a `tuple`. When + returning a tuple, the first element is the sample tensor. + """ + if self.num_inference_steps is None: + raise ValueError( + "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler" + ) + + # See formulas (12) and (16) of DDIM paper https://arxiv.org/pdf/2010.02502.pdf + # Ideally, read DDIM paper in-detail understanding + + # Notation ( -> + # - pred_noise_t -> e_theta(x_t, t) + # - pred_original_sample -> f_theta(x_t, t) or x_0 + # - std_dev_t -> sigma_t + # - eta -> η + # - pred_sample_direction -> "direction pointing to x_t" + # - pred_prev_sample -> "x_t-1" + + # 1. get previous step value (=t-1) + prev_timestep = timestep - self.config.num_train_timesteps // self.num_inference_steps + + # 2. compute alphas, betas + alpha_prod_t = self.alphas_cumprod[timestep] + alpha_prod_t_prev = self.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else self.final_alpha_cumprod + + beta_prod_t = 1 - alpha_prod_t + + # 3. compute predicted original sample from predicted noise also called + # "predicted x_0" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf + pred_original_sample = (sample - beta_prod_t ** (0.5) * model_output) / alpha_prod_t ** (0.5) + + # 4. Clip "predicted x_0" + scale = self.bit_scale + if self.config.clip_sample: + pred_original_sample = torch.clamp(pred_original_sample, -scale, scale) + + # 5. compute variance: "sigma_t(η)" -> see formula (16) + # σ_t = sqrt((1 − α_t−1)/(1 − α_t)) * sqrt(1 − α_t/α_t−1) + variance = self._get_variance(timestep, prev_timestep) + std_dev_t = eta * variance ** (0.5) + + if use_clipped_model_output: + # the model_output is always re-derived from the clipped x_0 in Glide + model_output = (sample - alpha_prod_t ** (0.5) * pred_original_sample) / beta_prod_t ** (0.5) + + # 6. compute "direction pointing to x_t" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf + pred_sample_direction = (1 - alpha_prod_t_prev - std_dev_t**2) ** (0.5) * model_output + + # 7. compute x_t without "random noise" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf + prev_sample = alpha_prod_t_prev ** (0.5) * pred_original_sample + pred_sample_direction + + if eta > 0: + # randn_like does not support generator https://github.com/pytorch/pytorch/issues/27072 + device = model_output.device if torch.is_tensor(model_output) else "cpu" + noise = torch.randn(model_output.shape, dtype=model_output.dtype, generator=generator).to(device) + variance = self._get_variance(timestep, prev_timestep) ** (0.5) * eta * noise + + prev_sample = prev_sample + variance + + if not return_dict: + return (prev_sample,) + + return DDIMSchedulerOutput(prev_sample=prev_sample, pred_original_sample=pred_original_sample) + + +def ddpm_bit_scheduler_step( + self, + model_output: torch.FloatTensor, + timestep: int, + sample: torch.FloatTensor, + predict_epsilon=True, + generator=None, + return_dict: bool = True, +) -> Union[DDPMSchedulerOutput, Tuple]: + """ + Predict the sample at the previous timestep by reversing the SDE. Core function to propagate the diffusion + process from the learned model outputs (most often the predicted noise). + Args: + model_output (`torch.FloatTensor`): direct output from learned diffusion model. + timestep (`int`): current discrete timestep in the diffusion chain. + sample (`torch.FloatTensor`): + current instance of sample being created by diffusion process. + predict_epsilon (`bool`): + optional flag to use when model predicts the samples directly instead of the noise, epsilon. + generator: random number generator. + return_dict (`bool`): option for returning tuple rather than DDPMSchedulerOutput class + Returns: + [`~schedulers.scheduling_utils.DDPMSchedulerOutput`] or `tuple`: + [`~schedulers.scheduling_utils.DDPMSchedulerOutput`] if `return_dict` is True, otherwise a `tuple`. When + returning a tuple, the first element is the sample tensor. + """ + t = timestep + + if model_output.shape[1] == sample.shape[1] * 2 and self.variance_type in ["learned", "learned_range"]: + model_output, predicted_variance = torch.split(model_output, sample.shape[1], dim=1) + else: + predicted_variance = None + + # 1. compute alphas, betas + alpha_prod_t = self.alphas_cumprod[t] + alpha_prod_t_prev = self.alphas_cumprod[t - 1] if t > 0 else self.one + beta_prod_t = 1 - alpha_prod_t + beta_prod_t_prev = 1 - alpha_prod_t_prev + + # 2. compute predicted original sample from predicted noise also called + # "predicted x_0" of formula (15) from https://arxiv.org/pdf/2006.11239.pdf + if predict_epsilon: + pred_original_sample = (sample - beta_prod_t ** (0.5) * model_output) / alpha_prod_t ** (0.5) + else: + pred_original_sample = model_output + + # 3. Clip "predicted x_0" + scale = self.bit_scale + if self.config.clip_sample: + pred_original_sample = torch.clamp(pred_original_sample, -scale, scale) + + # 4. Compute coefficients for pred_original_sample x_0 and current sample x_t + # See formula (7) from https://arxiv.org/pdf/2006.11239.pdf + pred_original_sample_coeff = (alpha_prod_t_prev ** (0.5) * self.betas[t]) / beta_prod_t + current_sample_coeff = self.alphas[t] ** (0.5) * beta_prod_t_prev / beta_prod_t + + # 5. Compute predicted previous sample µ_t + # See formula (7) from https://arxiv.org/pdf/2006.11239.pdf + pred_prev_sample = pred_original_sample_coeff * pred_original_sample + current_sample_coeff * sample + + # 6. Add noise + variance = 0 + if t > 0: + noise = torch.randn( + model_output.size(), dtype=model_output.dtype, layout=model_output.layout, generator=generator + ).to(model_output.device) + variance = (self._get_variance(t, predicted_variance=predicted_variance) ** 0.5) * noise + + pred_prev_sample = pred_prev_sample + variance + + if not return_dict: + return (pred_prev_sample,) + + return DDPMSchedulerOutput(prev_sample=pred_prev_sample, pred_original_sample=pred_original_sample) + + +class BitDiffusion(DiffusionPipeline): + def __init__( + self, + unet: UNet2DConditionModel, + scheduler: Union[DDIMScheduler, DDPMScheduler], + bit_scale: Optional[float] = 1.0, + ): + super().__init__() + self.bit_scale = bit_scale + self.scheduler.step = ( + ddim_bit_scheduler_step if isinstance(scheduler, DDIMScheduler) else ddpm_bit_scheduler_step + ) + + self.register_modules(unet=unet, scheduler=scheduler) + + @torch.no_grad() + def __call__( + self, + height: Optional[int] = 256, + width: Optional[int] = 256, + num_inference_steps: Optional[int] = 50, + generator: Optional[torch.Generator] = None, + batch_size: Optional[int] = 1, + output_type: Optional[str] = "pil", + return_dict: bool = True, + **kwargs, + ) -> Union[Tuple, ImagePipelineOutput]: + latents = torch.randn( + (batch_size, self.unet.in_channels, height, width), + generator=generator, + ) + latents = decimal_to_bits(latents) * self.bit_scale + latents = latents.to(self.device) + + self.scheduler.set_timesteps(num_inference_steps) + + for t in self.progress_bar(self.scheduler.timesteps): + # predict the noise residual + noise_pred = self.unet(latents, t).sample + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents).prev_sample + + image = bits_to_decimal(latents) + + if output_type == "pil": + image = self.numpy_to_pil(image) + + if not return_dict: + return (image,) + + return ImagePipelineOutput(images=image) diff --git a/examples/community/img2img_inpainting.py b/examples/community/img2img_inpainting.py index f7a107136d..3fa7db13a4 100644 --- a/examples/community/img2img_inpainting.py +++ b/examples/community/img2img_inpainting.py @@ -110,7 +110,7 @@ class ImageToImageInpaintingPipeline(DiffusionPipeline): scheduler._internal_dict = FrozenDict(new_config) if safety_checker is None: - logger.warn( + logger.warning( f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" " results in services or applications open to the public. Both the diffusers team and Hugging Face" diff --git a/examples/community/interpolate_stable_diffusion.py b/examples/community/interpolate_stable_diffusion.py index 761aaeca69..4d7a73f5ba 100644 --- a/examples/community/interpolate_stable_diffusion.py +++ b/examples/community/interpolate_stable_diffusion.py @@ -101,7 +101,7 @@ class StableDiffusionWalkPipeline(DiffusionPipeline): scheduler._internal_dict = FrozenDict(new_config) if safety_checker is None: - logger.warn( + logger.warning( f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" " results in services or applications open to the public. Both the diffusers team and Hugging Face" diff --git a/examples/community/lpw_stable_diffusion.py b/examples/community/lpw_stable_diffusion.py index b952ffe76d..0e7dc9e1ed 100644 --- a/examples/community/lpw_stable_diffusion.py +++ b/examples/community/lpw_stable_diffusion.py @@ -469,7 +469,7 @@ class StableDiffusionLongPromptWeightingPipeline(DiffusionPipeline): scheduler._internal_dict = FrozenDict(new_config) if safety_checker is None: - logger.warn( + logger.warning( f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" " results in services or applications open to the public. Both the diffusers team and Hugging Face" diff --git a/examples/community/multilingual_stable_diffusion.py b/examples/community/multilingual_stable_diffusion.py index c71c1f10c5..19974d6df0 100644 --- a/examples/community/multilingual_stable_diffusion.py +++ b/examples/community/multilingual_stable_diffusion.py @@ -113,7 +113,7 @@ class MultilingualStableDiffusion(DiffusionPipeline): scheduler._internal_dict = FrozenDict(new_config) if safety_checker is None: - logger.warn( + logger.warning( f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" " results in services or applications open to the public. Both the diffusers team and Hugging Face" diff --git a/examples/community/sd_text2img_k_diffusion.py b/examples/community/sd_text2img_k_diffusion.py new file mode 100755 index 0000000000..9592f7879f --- /dev/null +++ b/examples/community/sd_text2img_k_diffusion.py @@ -0,0 +1,479 @@ +# Copyright 2022 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import importlib +from typing import Callable, List, Optional, Union + +import torch + +from diffusers import LMSDiscreteScheduler +from diffusers.pipeline_utils import DiffusionPipeline +from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput +from diffusers.utils import is_accelerate_available, logging +from k_diffusion.external import CompVisDenoiser + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +class ModelWrapper: + def __init__(self, model, alphas_cumprod): + self.model = model + self.alphas_cumprod = alphas_cumprod + + def apply_model(self, *args, **kwargs): + return self.model(*args, **kwargs).sample + + +class StableDiffusionPipeline(DiffusionPipeline): + r""" + Pipeline for text-to-image generation using Stable Diffusion. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModel`]): + Frozen text-encoder. Stable Diffusion uses the text portion of + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically + the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + safety_checker ([`StableDiffusionSafetyChecker`]): + Classification module that estimates whether generated images could be considered offensive or harmful. + Please, refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for details. + feature_extractor ([`CLIPFeatureExtractor`]): + Model that extracts features from generated images to be used as inputs for the `safety_checker`. + """ + + def __init__( + self, + vae, + text_encoder, + tokenizer, + unet, + scheduler, + safety_checker, + feature_extractor, + ): + super().__init__() + + if safety_checker is None: + logger.warning( + f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" + " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" + " results in services or applications open to the public. Both the diffusers team and Hugging Face" + " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling" + " it only for use-cases that involve analyzing network behavior or auditing its results. For more" + " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ." + ) + + # get correct sigmas from LMS + scheduler = LMSDiscreteScheduler.from_config(scheduler.config) + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + ) + + model = ModelWrapper(unet, scheduler.alphas_cumprod) + self.k_diffusion_model = CompVisDenoiser(model) + + def set_sampler(self, scheduler_type: str): + library = importlib.import_module("k_diffusion") + sampling = getattr(library, "sampling") + self.sampler = getattr(sampling, scheduler_type) + + def enable_xformers_memory_efficient_attention(self): + r""" + Enable memory efficient attention as implemented in xformers. + + When this option is enabled, you should observe lower GPU memory usage and a potential speed up at inference + time. Speed up at training time is not guaranteed. + + Warning: When Memory Efficient Attention and Sliced attention are both enabled, the Memory Efficient Attention + is used. + """ + self.unet.set_use_memory_efficient_attention_xformers(True) + + def disable_xformers_memory_efficient_attention(self): + r""" + Disable memory efficient attention as implemented in xformers. + """ + self.unet.set_use_memory_efficient_attention_xformers(False) + + def enable_attention_slicing(self, slice_size: Optional[Union[str, int]] = "auto"): + r""" + Enable sliced attention computation. + + When this option is enabled, the attention module will split the input tensor in slices, to compute attention + in several steps. This is useful to save some memory in exchange for a small speed decrease. + + Args: + slice_size (`str` or `int`, *optional*, defaults to `"auto"`): + When `"auto"`, halves the input to the attention heads, so attention will be computed in two steps. If + a number is provided, uses as many slices as `attention_head_dim // slice_size`. In this case, + `attention_head_dim` must be a multiple of `slice_size`. + """ + if slice_size == "auto": + # half the attention head size is usually a good trade-off between + # speed and memory + slice_size = self.unet.config.attention_head_dim // 2 + self.unet.set_attention_slice(slice_size) + + def disable_attention_slicing(self): + r""" + Disable sliced attention computation. If `enable_attention_slicing` was previously invoked, this method will go + back to computing attention in one step. + """ + # set slice_size = `None` to disable `attention slicing` + self.enable_attention_slicing(None) + + def enable_sequential_cpu_offload(self, gpu_id=0): + r""" + Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet, + text_encoder, vae and safety checker have their state dicts saved to CPU and then are moved to a + `torch.device('meta') and loaded to GPU only when their specific submodule has its `forward` method called. + """ + if is_accelerate_available(): + from accelerate import cpu_offload + else: + raise ImportError("Please install accelerate via `pip install accelerate`") + + device = torch.device(f"cuda:{gpu_id}") + + for cpu_offloaded_model in [self.unet, self.text_encoder, self.vae, self.safety_checker]: + if cpu_offloaded_model is not None: + cpu_offload(cpu_offloaded_model, device) + + @property + def _execution_device(self): + r""" + Returns the device on which the pipeline's models will be executed. After calling + `pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module + hooks. + """ + if self.device != torch.device("meta") or not hasattr(self.unet, "_hf_hook"): + return self.device + for module in self.unet.modules(): + if ( + hasattr(module, "_hf_hook") + and hasattr(module._hf_hook, "execution_device") + and module._hf_hook.execution_device is not None + ): + return torch.device(module._hf_hook.execution_device) + return self.device + + def _encode_prompt(self, prompt, device, num_images_per_prompt, do_classifier_free_guidance, negative_prompt): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list(int)`): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `List[str]`): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `guidance_scale` is less than `1`). + """ + batch_size = len(prompt) if isinstance(prompt, list) else 1 + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="max_length", return_tensors="pt").input_ids + + if not torch.equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer.batch_decode(untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = text_inputs.attention_mask.to(device) + else: + attention_mask = None + + text_embeddings = self.text_encoder( + text_input_ids.to(device), + attention_mask=attention_mask, + ) + text_embeddings = text_embeddings[0] + + # duplicate text embeddings for each generation per prompt, using mps friendly method + bs_embed, seq_len, _ = text_embeddings.shape + text_embeddings = text_embeddings.repeat(1, num_images_per_prompt, 1) + text_embeddings = text_embeddings.view(bs_embed * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance: + uncond_tokens: List[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + max_length = text_input_ids.shape[-1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = uncond_input.attention_mask.to(device) + else: + attention_mask = None + + uncond_embeddings = self.text_encoder( + uncond_input.input_ids.to(device), + attention_mask=attention_mask, + ) + uncond_embeddings = uncond_embeddings[0] + + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = uncond_embeddings.shape[1] + uncond_embeddings = uncond_embeddings.repeat(1, num_images_per_prompt, 1) + uncond_embeddings = uncond_embeddings.view(batch_size * num_images_per_prompt, seq_len, -1) + + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + text_embeddings = torch.cat([uncond_embeddings, text_embeddings]) + + return text_embeddings + + def run_safety_checker(self, image, device, dtype): + if self.safety_checker is not None: + safety_checker_input = self.feature_extractor(self.numpy_to_pil(image), return_tensors="pt").to(device) + image, has_nsfw_concept = self.safety_checker( + images=image, clip_input=safety_checker_input.pixel_values.to(dtype) + ) + else: + has_nsfw_concept = None + return image, has_nsfw_concept + + def decode_latents(self, latents): + latents = 1 / 0.18215 * latents + image = self.vae.decode(latents).sample + image = (image / 2 + 0.5).clamp(0, 1) + # we always cast to float32 as this does not cause significant overhead and is compatible with bfloa16 + image = image.cpu().permute(0, 2, 3, 1).float().numpy() + return image + + def check_inputs(self, prompt, height, width, callback_steps): + if not isinstance(prompt, str) and not isinstance(prompt, list): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if (callback_steps is None) or ( + callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0) + ): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = (batch_size, num_channels_latents, height // 8, width // 8) + if latents is None: + if device.type == "mps": + # randn does not work reproducibly on mps + latents = torch.randn(shape, generator=generator, device="cpu", dtype=dtype).to(device) + else: + latents = torch.randn(shape, generator=generator, device=device, dtype=dtype) + else: + if latents.shape != shape: + raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}") + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + return latents + + @torch.no_grad() + def __call__( + self, + prompt: Union[str, List[str]], + height: int = 512, + width: int = 512, + num_inference_steps: int = 50, + guidance_scale: float = 7.5, + negative_prompt: Optional[Union[str, List[str]]] = None, + num_images_per_prompt: Optional[int] = 1, + eta: float = 0.0, + generator: Optional[torch.Generator] = None, + latents: Optional[torch.FloatTensor] = None, + output_type: Optional[str] = "pil", + return_dict: bool = True, + callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None, + callback_steps: Optional[int] = 1, + **kwargs, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `List[str]`): + The prompt or prompts to guide the image generation. + height (`int`, *optional*, defaults to 512): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to 512): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 7.5): + Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598). + `guidance_scale` is defined as `w` of equation 2. of [Imagen + Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale > + 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`, + usually at the expense of lower image quality. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `guidance_scale` is less than `1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to + [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`torch.Generator`, *optional*): + A [torch generator](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make generation + deterministic. + latents (`torch.FloatTensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will ge generated by sampling using the supplied random `generator`. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + callback (`Callable`, *optional*): + A function that will be called every `callback_steps` steps during inference. The function will be + called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function will be called. If not specified, the callback will be + called at every step. + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple. + When returning a tuple, the first element is a list with the generated images, and the second element is a + list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work" + (nsfw) content, according to the `safety_checker`. + """ + + # 1. Check inputs. Raise error if not correct + self.check_inputs(prompt, height, width, callback_steps) + + # 2. Define call parameters + batch_size = 1 if isinstance(prompt, str) else len(prompt) + device = self._execution_device + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = True + if guidance_scale <= 1.0: + raise ValueError("has to use guidance_scale") + + # 3. Encode input prompt + text_embeddings = self._encode_prompt( + prompt, device, num_images_per_prompt, do_classifier_free_guidance, negative_prompt + ) + + # 4. Prepare timesteps + self.scheduler.set_timesteps(num_inference_steps, device=text_embeddings.device) + sigmas = self.scheduler.sigmas + + # 5. Prepare latent variables + num_channels_latents = self.unet.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + text_embeddings.dtype, + device, + generator, + latents, + ) + latents = latents * sigmas[0] + self.k_diffusion_model.sigmas = self.k_diffusion_model.sigmas.to(latents.device) + self.k_diffusion_model.log_sigmas = self.k_diffusion_model.log_sigmas.to(latents.device) + + def model_fn(x, t): + latent_model_input = torch.cat([x] * 2) + + noise_pred = self.k_diffusion_model(latent_model_input, t, encoder_hidden_states=text_embeddings) + + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + return noise_pred + + latents = self.sampler(model_fn, latents, sigmas) + + # 8. Post-processing + image = self.decode_latents(latents) + + # 9. Run safety checker + image, has_nsfw_concept = self.run_safety_checker(image, device, text_embeddings.dtype) + + # 10. Convert to PIL + if output_type == "pil": + image = self.numpy_to_pil(image) + + if not return_dict: + return (image, has_nsfw_concept) + + return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept) diff --git a/examples/community/speech_to_image_diffusion.py b/examples/community/speech_to_image_diffusion.py index 1a9d296e81..17bc08e3c2 100644 --- a/examples/community/speech_to_image_diffusion.py +++ b/examples/community/speech_to_image_diffusion.py @@ -42,7 +42,7 @@ class SpeechToImagePipeline(DiffusionPipeline): super().__init__() if safety_checker is None: - logger.warn( + logger.warning( f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" " results in services or applications open to the public. Both the diffusers team and Hugging Face" diff --git a/examples/community/text_inpainting.py b/examples/community/text_inpainting.py index 38d5e96337..a4368f8b43 100644 --- a/examples/community/text_inpainting.py +++ b/examples/community/text_inpainting.py @@ -99,7 +99,7 @@ class TextInpainting(DiffusionPipeline): scheduler._internal_dict = FrozenDict(new_config) if safety_checker is None: - logger.warn( + logger.warning( f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" " results in services or applications open to the public. Both the diffusers team and Hugging Face" diff --git a/examples/community/wildcard_stable_diffusion.py b/examples/community/wildcard_stable_diffusion.py index 9ad0d8e9fa..282be8e48b 100644 --- a/examples/community/wildcard_stable_diffusion.py +++ b/examples/community/wildcard_stable_diffusion.py @@ -135,7 +135,7 @@ class WildcardStableDiffusionPipeline(DiffusionPipeline): scheduler._internal_dict = FrozenDict(new_config) if safety_checker is None: - logger.warn( + logger.warning( f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" " results in services or applications open to the public. Both the diffusers team and Hugging Face" diff --git a/examples/dreambooth/README.md b/examples/dreambooth/README.md index 2339e2979d..7aaf1bc46c 100644 --- a/examples/dreambooth/README.md +++ b/examples/dreambooth/README.md @@ -141,7 +141,7 @@ export INSTANCE_DIR="path-to-instance-images" export CLASS_DIR="path-to-class-images" export OUTPUT_DIR="path-to-save-model" -accelerate launch train_dreambooth.py \ +accelerate launch --mixed_precision="fp16" train_dreambooth.py \ --pretrained_model_name_or_path=$MODEL_NAME \ --instance_data_dir=$INSTANCE_DIR \ --class_data_dir=$CLASS_DIR \ @@ -157,8 +157,7 @@ accelerate launch train_dreambooth.py \ --lr_scheduler="constant" \ --lr_warmup_steps=0 \ --num_class_images=200 \ - --max_train_steps=800 \ - --mixed_precision=fp16 + --max_train_steps=800 ``` ### Fine-tune text encoder with the UNet. diff --git a/examples/dreambooth/train_dreambooth.py b/examples/dreambooth/train_dreambooth.py index 610c18533b..1f6c730f2b 100644 --- a/examples/dreambooth/train_dreambooth.py +++ b/examples/dreambooth/train_dreambooth.py @@ -187,12 +187,12 @@ def parse_args(input_args=None): parser.add_argument( "--mixed_precision", type=str, - default="no", + default=None, choices=["no", "fp16", "bf16"], help=( - "Whether to use mixed precision. Choose" - "between fp16 and bf16 (bfloat16). Bf16 requires PyTorch >= 1.10." - "and an Nvidia Ampere GPU." + "Whether to use mixed precision. Choose between fp16 and bf16 (bfloat16). Bf16 requires PyTorch >=" + " 1.10.and an Nvidia Ampere GPU. Default to the value of accelerate config of the current system or the" + " flag passed with the `accelerate.launch` command. Use this argument to override the accelerate config." ), ) parser.add_argument("--local_rank", type=int, default=-1, help="For distributed training: local_rank") @@ -472,7 +472,7 @@ def main(args): eps=args.adam_epsilon, ) - noise_scheduler = DDPMScheduler.from_config("CompVis/stable-diffusion-v1-4", subfolder="scheduler") + noise_scheduler = DDPMScheduler.from_config(args.pretrained_model_name_or_path, subfolder="scheduler") train_dataset = DreamBoothDataset( instance_data_root=args.instance_data_dir, @@ -538,9 +538,9 @@ def main(args): ) weight_dtype = torch.float32 - if args.mixed_precision == "fp16": + if accelerator.mixed_precision == "fp16": weight_dtype = torch.float16 - elif args.mixed_precision == "bf16": + elif accelerator.mixed_precision == "bf16": weight_dtype = torch.bfloat16 # Move text_encode and vae to gpu. diff --git a/examples/text_to_image/README.md b/examples/text_to_image/README.md index 170ed384f1..abe2187584 100644 --- a/examples/text_to_image/README.md +++ b/examples/text_to_image/README.md @@ -46,7 +46,7 @@ With `gradient_checkpointing` and `mixed_precision` it should be possible to fin export MODEL_NAME="CompVis/stable-diffusion-v1-4" export dataset_name="lambdalabs/pokemon-blip-captions" -accelerate launch train_text_to_image.py \ +accelerate launch --mixed_precision="fp16" train_text_to_image.py \ --pretrained_model_name_or_path=$MODEL_NAME \ --dataset_name=$dataset_name \ --use_ema \ @@ -54,7 +54,6 @@ accelerate launch train_text_to_image.py \ --train_batch_size=1 \ --gradient_accumulation_steps=4 \ --gradient_checkpointing \ - --mixed_precision="fp16" \ --max_train_steps=15000 \ --learning_rate=1e-05 \ --max_grad_norm=1 \ @@ -70,7 +69,7 @@ If you wish to use custom loading logic, you should modify the script, we have l export MODEL_NAME="CompVis/stable-diffusion-v1-4" export TRAIN_DIR="path_to_your_dataset" -accelerate launch train_text_to_image.py \ +accelerate launch --mixed_precision="fp16" train_text_to_image.py \ --pretrained_model_name_or_path=$MODEL_NAME \ --train_data_dir=$TRAIN_DIR \ --use_ema \ @@ -78,7 +77,6 @@ accelerate launch train_text_to_image.py \ --train_batch_size=1 \ --gradient_accumulation_steps=4 \ --gradient_checkpointing \ - --mixed_precision="fp16" \ --max_train_steps=15000 \ --learning_rate=1e-05 \ --max_grad_norm=1 \ diff --git a/examples/text_to_image/train_text_to_image.py b/examples/text_to_image/train_text_to_image.py index cf7dac8933..88da2a5509 100644 --- a/examples/text_to_image/train_text_to_image.py +++ b/examples/text_to_image/train_text_to_image.py @@ -186,12 +186,12 @@ def parse_args(): parser.add_argument( "--mixed_precision", type=str, - default="no", + default=None, choices=["no", "fp16", "bf16"], help=( - "Whether to use mixed precision. Choose" - "between fp16 and bf16 (bfloat16). Bf16 requires PyTorch >= 1.10." - "and an Nvidia Ampere GPU." + "Whether to use mixed precision. Choose between fp16 and bf16 (bfloat16). Bf16 requires PyTorch >=" + " 1.10.and an Nvidia Ampere GPU. Default to the value of accelerate config of the current system or the" + " flag passed with the `accelerate.launch` command. Use this argument to override the accelerate config." ), ) parser.add_argument( @@ -372,7 +372,7 @@ def main(): weight_decay=args.adam_weight_decay, eps=args.adam_epsilon, ) - noise_scheduler = DDPMScheduler.from_config("CompVis/stable-diffusion-v1-4", subfolder="scheduler") + noise_scheduler = DDPMScheduler.from_config(args.pretrained_model_name_or_path, subfolder="scheduler") # Get the datasets: you can either provide your own training and evaluation files (see below) # or specify a Dataset from the hub (the dataset will be downloaded automatically from the datasets Hub). @@ -496,9 +496,9 @@ def main(): ) weight_dtype = torch.float32 - if args.mixed_precision == "fp16": + if accelerator.mixed_precision == "fp16": weight_dtype = torch.float16 - elif args.mixed_precision == "bf16": + elif accelerator.mixed_precision == "bf16": weight_dtype = torch.bfloat16 # Move text_encode and vae to gpu. @@ -605,7 +605,7 @@ def main(): vae=vae, unet=unet, tokenizer=tokenizer, - scheduler=PNDMScheduler.from_config("CompVis/stable-diffusion-v1-4", subfolder="scheduler"), + scheduler=PNDMScheduler.from_config(args.pretrained_model_name_or_path, subfolder="scheduler"), safety_checker=StableDiffusionSafetyChecker.from_pretrained("CompVis/stable-diffusion-safety-checker"), feature_extractor=CLIPFeatureExtractor.from_pretrained("openai/clip-vit-base-patch32"), ) diff --git a/examples/textual_inversion/textual_inversion.py b/examples/textual_inversion/textual_inversion.py index 380ce90297..7d9fb7c0f1 100644 --- a/examples/textual_inversion/textual_inversion.py +++ b/examples/textual_inversion/textual_inversion.py @@ -441,7 +441,7 @@ def main(): eps=args.adam_epsilon, ) - noise_scheduler = DDPMScheduler.from_config("CompVis/stable-diffusion-v1-4", subfolder="scheduler") + noise_scheduler = DDPMScheduler.from_config(args.pretrained_model_name_or_path, subfolder="scheduler") train_dataset = TextualInversionDataset( data_root=args.train_data_dir, @@ -574,7 +574,7 @@ def main(): vae=vae, unet=unet, tokenizer=tokenizer, - scheduler=PNDMScheduler.from_config("CompVis/stable-diffusion-v1-4", subfolder="scheduler"), + scheduler=PNDMScheduler.from_config(args.pretrained_model_name_or_path, subfolder="scheduler"), safety_checker=StableDiffusionSafetyChecker.from_pretrained("CompVis/stable-diffusion-safety-checker"), feature_extractor=CLIPFeatureExtractor.from_pretrained("openai/clip-vit-base-patch32"), ) diff --git a/scripts/convert_versatile_diffusion_to_diffusers.py b/scripts/convert_versatile_diffusion_to_diffusers.py new file mode 100644 index 0000000000..86fb0e7b4c --- /dev/null +++ b/scripts/convert_versatile_diffusion_to_diffusers.py @@ -0,0 +1,791 @@ +# coding=utf-8 +# Copyright 2022 The HuggingFace Inc. team. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" Conversion script for the Versatile Stable Diffusion checkpoints. """ + +import argparse +from argparse import Namespace + +import torch + +from diffusers import ( + AutoencoderKL, + DDIMScheduler, + DPMSolverMultistepScheduler, + EulerAncestralDiscreteScheduler, + EulerDiscreteScheduler, + LMSDiscreteScheduler, + PNDMScheduler, + UNet2DConditionModel, + VersatileDiffusionPipeline, +) +from diffusers.pipelines.versatile_diffusion.modeling_text_unet import UNetFlatConditionModel +from transformers import ( + CLIPFeatureExtractor, + CLIPTextModelWithProjection, + CLIPTokenizer, + CLIPVisionModelWithProjection, +) + + +SCHEDULER_CONFIG = Namespace( + **{ + "beta_linear_start": 0.00085, + "beta_linear_end": 0.012, + "timesteps": 1000, + "scale_factor": 0.18215, + } +) + +IMAGE_UNET_CONFIG = Namespace( + **{ + "input_channels": 4, + "model_channels": 320, + "output_channels": 4, + "num_noattn_blocks": [2, 2, 2, 2], + "channel_mult": [1, 2, 4, 4], + "with_attn": [True, True, True, False], + "num_heads": 8, + "context_dim": 768, + "use_checkpoint": True, + } +) + +TEXT_UNET_CONFIG = Namespace( + **{ + "input_channels": 768, + "model_channels": 320, + "output_channels": 768, + "num_noattn_blocks": [2, 2, 2, 2], + "channel_mult": [1, 2, 4, 4], + "second_dim": [4, 4, 4, 4], + "with_attn": [True, True, True, False], + "num_heads": 8, + "context_dim": 768, + "use_checkpoint": True, + } +) + +AUTOENCODER_CONFIG = Namespace( + **{ + "double_z": True, + "z_channels": 4, + "resolution": 256, + "in_channels": 3, + "out_ch": 3, + "ch": 128, + "ch_mult": [1, 2, 4, 4], + "num_res_blocks": 2, + "attn_resolutions": [], + "dropout": 0.0, + } +) + + +def shave_segments(path, n_shave_prefix_segments=1): + """ + Removes segments. Positive values shave the first segments, negative shave the last segments. + """ + if n_shave_prefix_segments >= 0: + return ".".join(path.split(".")[n_shave_prefix_segments:]) + else: + return ".".join(path.split(".")[:n_shave_prefix_segments]) + + +def renew_resnet_paths(old_list, n_shave_prefix_segments=0): + """ + Updates paths inside resnets to the new naming scheme (local renaming) + """ + mapping = [] + for old_item in old_list: + new_item = old_item.replace("in_layers.0", "norm1") + new_item = new_item.replace("in_layers.2", "conv1") + + new_item = new_item.replace("out_layers.0", "norm2") + new_item = new_item.replace("out_layers.3", "conv2") + + new_item = new_item.replace("emb_layers.1", "time_emb_proj") + new_item = new_item.replace("skip_connection", "conv_shortcut") + + new_item = shave_segments(new_item, n_shave_prefix_segments=n_shave_prefix_segments) + + mapping.append({"old": old_item, "new": new_item}) + + return mapping + + +def renew_vae_resnet_paths(old_list, n_shave_prefix_segments=0): + """ + Updates paths inside resnets to the new naming scheme (local renaming) + """ + mapping = [] + for old_item in old_list: + new_item = old_item + + new_item = new_item.replace("nin_shortcut", "conv_shortcut") + new_item = shave_segments(new_item, n_shave_prefix_segments=n_shave_prefix_segments) + + mapping.append({"old": old_item, "new": new_item}) + + return mapping + + +def renew_attention_paths(old_list, n_shave_prefix_segments=0): + """ + Updates paths inside attentions to the new naming scheme (local renaming) + """ + mapping = [] + for old_item in old_list: + new_item = old_item + + # new_item = new_item.replace('norm.weight', 'group_norm.weight') + # new_item = new_item.replace('norm.bias', 'group_norm.bias') + + # new_item = new_item.replace('proj_out.weight', 'proj_attn.weight') + # new_item = new_item.replace('proj_out.bias', 'proj_attn.bias') + + # new_item = shave_segments(new_item, n_shave_prefix_segments=n_shave_prefix_segments) + + mapping.append({"old": old_item, "new": new_item}) + + return mapping + + +def renew_vae_attention_paths(old_list, n_shave_prefix_segments=0): + """ + Updates paths inside attentions to the new naming scheme (local renaming) + """ + mapping = [] + for old_item in old_list: + new_item = old_item + + new_item = new_item.replace("norm.weight", "group_norm.weight") + new_item = new_item.replace("norm.bias", "group_norm.bias") + + new_item = new_item.replace("q.weight", "query.weight") + new_item = new_item.replace("q.bias", "query.bias") + + new_item = new_item.replace("k.weight", "key.weight") + new_item = new_item.replace("k.bias", "key.bias") + + new_item = new_item.replace("v.weight", "value.weight") + new_item = new_item.replace("v.bias", "value.bias") + + new_item = new_item.replace("proj_out.weight", "proj_attn.weight") + new_item = new_item.replace("proj_out.bias", "proj_attn.bias") + + new_item = shave_segments(new_item, n_shave_prefix_segments=n_shave_prefix_segments) + + mapping.append({"old": old_item, "new": new_item}) + + return mapping + + +def assign_to_checkpoint( + paths, checkpoint, old_checkpoint, attention_paths_to_split=None, additional_replacements=None, config=None +): + """ + This does the final conversion step: take locally converted weights and apply a global renaming + to them. It splits attention layers, and takes into account additional replacements + that may arise. + + Assigns the weights to the new checkpoint. + """ + assert isinstance(paths, list), "Paths should be a list of dicts containing 'old' and 'new' keys." + + # Splits the attention layers into three variables. + if attention_paths_to_split is not None: + for path, path_map in attention_paths_to_split.items(): + old_tensor = old_checkpoint[path] + channels = old_tensor.shape[0] // 3 + + target_shape = (-1, channels) if len(old_tensor.shape) == 3 else (-1) + + num_heads = old_tensor.shape[0] // config["num_head_channels"] // 3 + + old_tensor = old_tensor.reshape((num_heads, 3 * channels // num_heads) + old_tensor.shape[1:]) + query, key, value = old_tensor.split(channels // num_heads, dim=1) + + checkpoint[path_map["query"]] = query.reshape(target_shape) + checkpoint[path_map["key"]] = key.reshape(target_shape) + checkpoint[path_map["value"]] = value.reshape(target_shape) + + for path in paths: + new_path = path["new"] + + # These have already been assigned + if attention_paths_to_split is not None and new_path in attention_paths_to_split: + continue + + # Global renaming happens here + new_path = new_path.replace("middle_block.0", "mid_block.resnets.0") + new_path = new_path.replace("middle_block.1", "mid_block.attentions.0") + new_path = new_path.replace("middle_block.2", "mid_block.resnets.1") + + if additional_replacements is not None: + for replacement in additional_replacements: + new_path = new_path.replace(replacement["old"], replacement["new"]) + + # proj_attn.weight has to be converted from conv 1D to linear + if "proj_attn.weight" in new_path: + checkpoint[new_path] = old_checkpoint[path["old"]][:, :, 0] + elif path["old"] in old_checkpoint: + checkpoint[new_path] = old_checkpoint[path["old"]] + + +def conv_attn_to_linear(checkpoint): + keys = list(checkpoint.keys()) + attn_keys = ["query.weight", "key.weight", "value.weight"] + for key in keys: + if ".".join(key.split(".")[-2:]) in attn_keys: + if checkpoint[key].ndim > 2: + checkpoint[key] = checkpoint[key][:, :, 0, 0] + elif "proj_attn.weight" in key: + if checkpoint[key].ndim > 2: + checkpoint[key] = checkpoint[key][:, :, 0] + + +def create_image_unet_diffusers_config(unet_params): + """ + Creates a config for the diffusers based on the config of the VD model. + """ + + block_out_channels = [unet_params.model_channels * mult for mult in unet_params.channel_mult] + + down_block_types = [] + resolution = 1 + for i in range(len(block_out_channels)): + block_type = "CrossAttnDownBlock2D" if unet_params.with_attn[i] else "DownBlock2D" + down_block_types.append(block_type) + if i != len(block_out_channels) - 1: + resolution *= 2 + + up_block_types = [] + for i in range(len(block_out_channels)): + block_type = "CrossAttnUpBlock2D" if unet_params.with_attn[-i - 1] else "UpBlock2D" + up_block_types.append(block_type) + resolution //= 2 + + if not all(n == unet_params.num_noattn_blocks[0] for n in unet_params.num_noattn_blocks): + raise ValueError("Not all num_res_blocks are equal, which is not supported in this script.") + + config = dict( + sample_size=None, + in_channels=unet_params.input_channels, + out_channels=unet_params.output_channels, + down_block_types=tuple(down_block_types), + up_block_types=tuple(up_block_types), + block_out_channels=tuple(block_out_channels), + layers_per_block=unet_params.num_noattn_blocks[0], + cross_attention_dim=unet_params.context_dim, + attention_head_dim=unet_params.num_heads, + ) + + return config + + +def create_text_unet_diffusers_config(unet_params): + """ + Creates a config for the diffusers based on the config of the VD model. + """ + + block_out_channels = [unet_params.model_channels * mult for mult in unet_params.channel_mult] + + down_block_types = [] + resolution = 1 + for i in range(len(block_out_channels)): + block_type = "CrossAttnDownBlockFlat" if unet_params.with_attn[i] else "DownBlockFlat" + down_block_types.append(block_type) + if i != len(block_out_channels) - 1: + resolution *= 2 + + up_block_types = [] + for i in range(len(block_out_channels)): + block_type = "CrossAttnUpBlockFlat" if unet_params.with_attn[-i - 1] else "UpBlockFlat" + up_block_types.append(block_type) + resolution //= 2 + + if not all(n == unet_params.num_noattn_blocks[0] for n in unet_params.num_noattn_blocks): + raise ValueError("Not all num_res_blocks are equal, which is not supported in this script.") + + config = dict( + sample_size=None, + in_channels=(unet_params.input_channels, 1, 1), + out_channels=(unet_params.output_channels, 1, 1), + down_block_types=tuple(down_block_types), + up_block_types=tuple(up_block_types), + block_out_channels=tuple(block_out_channels), + layers_per_block=unet_params.num_noattn_blocks[0], + cross_attention_dim=unet_params.context_dim, + attention_head_dim=unet_params.num_heads, + ) + + return config + + +def create_vae_diffusers_config(vae_params): + """ + Creates a config for the diffusers based on the config of the VD model. + """ + + block_out_channels = [vae_params.ch * mult for mult in vae_params.ch_mult] + down_block_types = ["DownEncoderBlock2D"] * len(block_out_channels) + up_block_types = ["UpDecoderBlock2D"] * len(block_out_channels) + + config = dict( + sample_size=vae_params.resolution, + in_channels=vae_params.in_channels, + out_channels=vae_params.out_ch, + down_block_types=tuple(down_block_types), + up_block_types=tuple(up_block_types), + block_out_channels=tuple(block_out_channels), + latent_channels=vae_params.z_channels, + layers_per_block=vae_params.num_res_blocks, + ) + return config + + +def create_diffusers_scheduler(original_config): + schedular = DDIMScheduler( + num_train_timesteps=original_config.model.params.timesteps, + beta_start=original_config.model.params.linear_start, + beta_end=original_config.model.params.linear_end, + beta_schedule="scaled_linear", + ) + return schedular + + +def convert_vd_unet_checkpoint(checkpoint, config, unet_key, extract_ema=False): + """ + Takes a state dict and a config, and returns a converted checkpoint. + """ + + # extract state_dict for UNet + unet_state_dict = {} + keys = list(checkpoint.keys()) + + # at least a 100 parameters have to start with `model_ema` in order for the checkpoint to be EMA + if sum(k.startswith("model_ema") for k in keys) > 100: + print("Checkpoint has both EMA and non-EMA weights.") + if extract_ema: + print( + "In this conversion only the EMA weights are extracted. If you want to instead extract the non-EMA" + " weights (useful to continue fine-tuning), please make sure to remove the `--extract_ema` flag." + ) + for key in keys: + if key.startswith("model.diffusion_model"): + flat_ema_key = "model_ema." + "".join(key.split(".")[1:]) + unet_state_dict[key.replace(unet_key, "")] = checkpoint.pop(flat_ema_key) + else: + print( + "In this conversion only the non-EMA weights are extracted. If you want to instead extract the EMA" + " weights (usually better for inference), please make sure to add the `--extract_ema` flag." + ) + + for key in keys: + if key.startswith(unet_key): + unet_state_dict[key.replace(unet_key, "")] = checkpoint.pop(key) + + new_checkpoint = {} + + new_checkpoint["time_embedding.linear_1.weight"] = checkpoint["model.diffusion_model.time_embed.0.weight"] + new_checkpoint["time_embedding.linear_1.bias"] = checkpoint["model.diffusion_model.time_embed.0.bias"] + new_checkpoint["time_embedding.linear_2.weight"] = checkpoint["model.diffusion_model.time_embed.2.weight"] + new_checkpoint["time_embedding.linear_2.bias"] = checkpoint["model.diffusion_model.time_embed.2.bias"] + + new_checkpoint["conv_in.weight"] = unet_state_dict["input_blocks.0.0.weight"] + new_checkpoint["conv_in.bias"] = unet_state_dict["input_blocks.0.0.bias"] + + new_checkpoint["conv_norm_out.weight"] = unet_state_dict["out.0.weight"] + new_checkpoint["conv_norm_out.bias"] = unet_state_dict["out.0.bias"] + new_checkpoint["conv_out.weight"] = unet_state_dict["out.2.weight"] + new_checkpoint["conv_out.bias"] = unet_state_dict["out.2.bias"] + + # Retrieves the keys for the input blocks only + num_input_blocks = len({".".join(layer.split(".")[:2]) for layer in unet_state_dict if "input_blocks" in layer}) + input_blocks = { + layer_id: [key for key in unet_state_dict if f"input_blocks.{layer_id}" in key] + for layer_id in range(num_input_blocks) + } + + # Retrieves the keys for the middle blocks only + num_middle_blocks = len({".".join(layer.split(".")[:2]) for layer in unet_state_dict if "middle_block" in layer}) + middle_blocks = { + layer_id: [key for key in unet_state_dict if f"middle_block.{layer_id}" in key] + for layer_id in range(num_middle_blocks) + } + + # Retrieves the keys for the output blocks only + num_output_blocks = len({".".join(layer.split(".")[:2]) for layer in unet_state_dict if "output_blocks" in layer}) + output_blocks = { + layer_id: [key for key in unet_state_dict if f"output_blocks.{layer_id}" in key] + for layer_id in range(num_output_blocks) + } + + for i in range(1, num_input_blocks): + block_id = (i - 1) // (config["layers_per_block"] + 1) + layer_in_block_id = (i - 1) % (config["layers_per_block"] + 1) + + resnets = [ + key for key in input_blocks[i] if f"input_blocks.{i}.0" in key and f"input_blocks.{i}.0.op" not in key + ] + attentions = [key for key in input_blocks[i] if f"input_blocks.{i}.1" in key] + + if f"input_blocks.{i}.0.op.weight" in unet_state_dict: + new_checkpoint[f"down_blocks.{block_id}.downsamplers.0.conv.weight"] = unet_state_dict.pop( + f"input_blocks.{i}.0.op.weight" + ) + new_checkpoint[f"down_blocks.{block_id}.downsamplers.0.conv.bias"] = unet_state_dict.pop( + f"input_blocks.{i}.0.op.bias" + ) + elif f"input_blocks.{i}.0.weight" in unet_state_dict: + # text_unet uses linear layers in place of downsamplers + shape = unet_state_dict[f"input_blocks.{i}.0.weight"].shape + if shape[0] != shape[1]: + continue + new_checkpoint[f"down_blocks.{block_id}.downsamplers.0.weight"] = unet_state_dict.pop( + f"input_blocks.{i}.0.weight" + ) + new_checkpoint[f"down_blocks.{block_id}.downsamplers.0.bias"] = unet_state_dict.pop( + f"input_blocks.{i}.0.bias" + ) + + paths = renew_resnet_paths(resnets) + meta_path = {"old": f"input_blocks.{i}.0", "new": f"down_blocks.{block_id}.resnets.{layer_in_block_id}"} + assign_to_checkpoint( + paths, new_checkpoint, unet_state_dict, additional_replacements=[meta_path], config=config + ) + + if len(attentions): + paths = renew_attention_paths(attentions) + meta_path = {"old": f"input_blocks.{i}.1", "new": f"down_blocks.{block_id}.attentions.{layer_in_block_id}"} + assign_to_checkpoint( + paths, new_checkpoint, unet_state_dict, additional_replacements=[meta_path], config=config + ) + + resnet_0 = middle_blocks[0] + attentions = middle_blocks[1] + resnet_1 = middle_blocks[2] + + resnet_0_paths = renew_resnet_paths(resnet_0) + assign_to_checkpoint(resnet_0_paths, new_checkpoint, unet_state_dict, config=config) + + resnet_1_paths = renew_resnet_paths(resnet_1) + assign_to_checkpoint(resnet_1_paths, new_checkpoint, unet_state_dict, config=config) + + attentions_paths = renew_attention_paths(attentions) + meta_path = {"old": "middle_block.1", "new": "mid_block.attentions.0"} + assign_to_checkpoint( + attentions_paths, new_checkpoint, unet_state_dict, additional_replacements=[meta_path], config=config + ) + + for i in range(num_output_blocks): + block_id = i // (config["layers_per_block"] + 1) + layer_in_block_id = i % (config["layers_per_block"] + 1) + output_block_layers = [shave_segments(name, 2) for name in output_blocks[i]] + output_block_list = {} + + for layer in output_block_layers: + layer_id, layer_name = layer.split(".")[0], shave_segments(layer, 1) + if layer_id in output_block_list: + output_block_list[layer_id].append(layer_name) + else: + output_block_list[layer_id] = [layer_name] + + if len(output_block_list) > 1: + resnets = [key for key in output_blocks[i] if f"output_blocks.{i}.0" in key] + attentions = [key for key in output_blocks[i] if f"output_blocks.{i}.1" in key] + + paths = renew_resnet_paths(resnets) + + meta_path = {"old": f"output_blocks.{i}.0", "new": f"up_blocks.{block_id}.resnets.{layer_in_block_id}"} + assign_to_checkpoint( + paths, new_checkpoint, unet_state_dict, additional_replacements=[meta_path], config=config + ) + + if ["conv.weight", "conv.bias"] in output_block_list.values(): + index = list(output_block_list.values()).index(["conv.weight", "conv.bias"]) + new_checkpoint[f"up_blocks.{block_id}.upsamplers.0.conv.weight"] = unet_state_dict[ + f"output_blocks.{i}.{index}.conv.weight" + ] + new_checkpoint[f"up_blocks.{block_id}.upsamplers.0.conv.bias"] = unet_state_dict[ + f"output_blocks.{i}.{index}.conv.bias" + ] + # Clear attentions as they have been attributed above. + if len(attentions) == 2: + attentions = [] + elif f"output_blocks.{i}.1.weight" in unet_state_dict: + # text_unet uses linear layers in place of upsamplers + shape = unet_state_dict[f"output_blocks.{i}.1.weight"].shape + if shape[0] != shape[1]: + continue + new_checkpoint[f"up_blocks.{block_id}.upsamplers.0.weight"] = unet_state_dict.pop( + f"output_blocks.{i}.1.weight" + ) + new_checkpoint[f"up_blocks.{block_id}.upsamplers.0.bias"] = unet_state_dict.pop( + f"output_blocks.{i}.1.bias" + ) + # Clear attentions as they have been attributed above. + if len(attentions) == 2: + attentions = [] + elif f"output_blocks.{i}.2.weight" in unet_state_dict: + # text_unet uses linear layers in place of upsamplers + shape = unet_state_dict[f"output_blocks.{i}.2.weight"].shape + if shape[0] != shape[1]: + continue + new_checkpoint[f"up_blocks.{block_id}.upsamplers.0.weight"] = unet_state_dict.pop( + f"output_blocks.{i}.2.weight" + ) + new_checkpoint[f"up_blocks.{block_id}.upsamplers.0.bias"] = unet_state_dict.pop( + f"output_blocks.{i}.2.bias" + ) + + if len(attentions): + paths = renew_attention_paths(attentions) + meta_path = { + "old": f"output_blocks.{i}.1", + "new": f"up_blocks.{block_id}.attentions.{layer_in_block_id}", + } + assign_to_checkpoint( + paths, new_checkpoint, unet_state_dict, additional_replacements=[meta_path], config=config + ) + else: + resnet_0_paths = renew_resnet_paths(output_block_layers, n_shave_prefix_segments=1) + for path in resnet_0_paths: + old_path = ".".join(["output_blocks", str(i), path["old"]]) + new_path = ".".join(["up_blocks", str(block_id), "resnets", str(layer_in_block_id), path["new"]]) + + new_checkpoint[new_path] = unet_state_dict[old_path] + + return new_checkpoint + + +def convert_vd_vae_checkpoint(checkpoint, config): + # extract state dict for VAE + vae_state_dict = {} + keys = list(checkpoint.keys()) + for key in keys: + vae_state_dict[key] = checkpoint.get(key) + + new_checkpoint = {} + + new_checkpoint["encoder.conv_in.weight"] = vae_state_dict["encoder.conv_in.weight"] + new_checkpoint["encoder.conv_in.bias"] = vae_state_dict["encoder.conv_in.bias"] + new_checkpoint["encoder.conv_out.weight"] = vae_state_dict["encoder.conv_out.weight"] + new_checkpoint["encoder.conv_out.bias"] = vae_state_dict["encoder.conv_out.bias"] + new_checkpoint["encoder.conv_norm_out.weight"] = vae_state_dict["encoder.norm_out.weight"] + new_checkpoint["encoder.conv_norm_out.bias"] = vae_state_dict["encoder.norm_out.bias"] + + new_checkpoint["decoder.conv_in.weight"] = vae_state_dict["decoder.conv_in.weight"] + new_checkpoint["decoder.conv_in.bias"] = vae_state_dict["decoder.conv_in.bias"] + new_checkpoint["decoder.conv_out.weight"] = vae_state_dict["decoder.conv_out.weight"] + new_checkpoint["decoder.conv_out.bias"] = vae_state_dict["decoder.conv_out.bias"] + new_checkpoint["decoder.conv_norm_out.weight"] = vae_state_dict["decoder.norm_out.weight"] + new_checkpoint["decoder.conv_norm_out.bias"] = vae_state_dict["decoder.norm_out.bias"] + + new_checkpoint["quant_conv.weight"] = vae_state_dict["quant_conv.weight"] + new_checkpoint["quant_conv.bias"] = vae_state_dict["quant_conv.bias"] + new_checkpoint["post_quant_conv.weight"] = vae_state_dict["post_quant_conv.weight"] + new_checkpoint["post_quant_conv.bias"] = vae_state_dict["post_quant_conv.bias"] + + # Retrieves the keys for the encoder down blocks only + num_down_blocks = len({".".join(layer.split(".")[:3]) for layer in vae_state_dict if "encoder.down" in layer}) + down_blocks = { + layer_id: [key for key in vae_state_dict if f"down.{layer_id}" in key] for layer_id in range(num_down_blocks) + } + + # Retrieves the keys for the decoder up blocks only + num_up_blocks = len({".".join(layer.split(".")[:3]) for layer in vae_state_dict if "decoder.up" in layer}) + up_blocks = { + layer_id: [key for key in vae_state_dict if f"up.{layer_id}" in key] for layer_id in range(num_up_blocks) + } + + for i in range(num_down_blocks): + resnets = [key for key in down_blocks[i] if f"down.{i}" in key and f"down.{i}.downsample" not in key] + + if f"encoder.down.{i}.downsample.conv.weight" in vae_state_dict: + new_checkpoint[f"encoder.down_blocks.{i}.downsamplers.0.conv.weight"] = vae_state_dict.pop( + f"encoder.down.{i}.downsample.conv.weight" + ) + new_checkpoint[f"encoder.down_blocks.{i}.downsamplers.0.conv.bias"] = vae_state_dict.pop( + f"encoder.down.{i}.downsample.conv.bias" + ) + + paths = renew_vae_resnet_paths(resnets) + meta_path = {"old": f"down.{i}.block", "new": f"down_blocks.{i}.resnets"} + assign_to_checkpoint(paths, new_checkpoint, vae_state_dict, additional_replacements=[meta_path], config=config) + + mid_resnets = [key for key in vae_state_dict if "encoder.mid.block" in key] + num_mid_res_blocks = 2 + for i in range(1, num_mid_res_blocks + 1): + resnets = [key for key in mid_resnets if f"encoder.mid.block_{i}" in key] + + paths = renew_vae_resnet_paths(resnets) + meta_path = {"old": f"mid.block_{i}", "new": f"mid_block.resnets.{i - 1}"} + assign_to_checkpoint(paths, new_checkpoint, vae_state_dict, additional_replacements=[meta_path], config=config) + + mid_attentions = [key for key in vae_state_dict if "encoder.mid.attn" in key] + paths = renew_vae_attention_paths(mid_attentions) + meta_path = {"old": "mid.attn_1", "new": "mid_block.attentions.0"} + assign_to_checkpoint(paths, new_checkpoint, vae_state_dict, additional_replacements=[meta_path], config=config) + conv_attn_to_linear(new_checkpoint) + + for i in range(num_up_blocks): + block_id = num_up_blocks - 1 - i + resnets = [ + key for key in up_blocks[block_id] if f"up.{block_id}" in key and f"up.{block_id}.upsample" not in key + ] + + if f"decoder.up.{block_id}.upsample.conv.weight" in vae_state_dict: + new_checkpoint[f"decoder.up_blocks.{i}.upsamplers.0.conv.weight"] = vae_state_dict[ + f"decoder.up.{block_id}.upsample.conv.weight" + ] + new_checkpoint[f"decoder.up_blocks.{i}.upsamplers.0.conv.bias"] = vae_state_dict[ + f"decoder.up.{block_id}.upsample.conv.bias" + ] + + paths = renew_vae_resnet_paths(resnets) + meta_path = {"old": f"up.{block_id}.block", "new": f"up_blocks.{i}.resnets"} + assign_to_checkpoint(paths, new_checkpoint, vae_state_dict, additional_replacements=[meta_path], config=config) + + mid_resnets = [key for key in vae_state_dict if "decoder.mid.block" in key] + num_mid_res_blocks = 2 + for i in range(1, num_mid_res_blocks + 1): + resnets = [key for key in mid_resnets if f"decoder.mid.block_{i}" in key] + + paths = renew_vae_resnet_paths(resnets) + meta_path = {"old": f"mid.block_{i}", "new": f"mid_block.resnets.{i - 1}"} + assign_to_checkpoint(paths, new_checkpoint, vae_state_dict, additional_replacements=[meta_path], config=config) + + mid_attentions = [key for key in vae_state_dict if "decoder.mid.attn" in key] + paths = renew_vae_attention_paths(mid_attentions) + meta_path = {"old": "mid.attn_1", "new": "mid_block.attentions.0"} + assign_to_checkpoint(paths, new_checkpoint, vae_state_dict, additional_replacements=[meta_path], config=config) + conv_attn_to_linear(new_checkpoint) + return new_checkpoint + + +if __name__ == "__main__": + parser = argparse.ArgumentParser() + + parser.add_argument( + "--unet_checkpoint_path", default=None, type=str, required=False, help="Path to the checkpoint to convert." + ) + parser.add_argument( + "--vae_checkpoint_path", default=None, type=str, required=False, help="Path to the checkpoint to convert." + ) + parser.add_argument( + "--optimus_checkpoint_path", default=None, type=str, required=False, help="Path to the checkpoint to convert." + ) + parser.add_argument( + "--scheduler_type", + default="pndm", + type=str, + help="Type of scheduler to use. Should be one of ['pndm', 'lms', 'ddim', 'euler', 'euler-ancest', 'dpm']", + ) + parser.add_argument( + "--extract_ema", + action="store_true", + help=( + "Only relevant for checkpoints that have both EMA and non-EMA weights. Whether to extract the EMA weights" + " or not. Defaults to `False`. Add `--extract_ema` to extract the EMA weights. EMA weights usually yield" + " higher quality images for inference. Non-EMA weights are usually better to continue fine-tuning." + ), + ) + parser.add_argument("--dump_path", default=None, type=str, required=True, help="Path to the output model.") + + args = parser.parse_args() + + scheduler_config = SCHEDULER_CONFIG + + num_train_timesteps = scheduler_config.timesteps + beta_start = scheduler_config.beta_linear_start + beta_end = scheduler_config.beta_linear_end + if args.scheduler_type == "pndm": + scheduler = PNDMScheduler( + beta_end=beta_end, + beta_schedule="scaled_linear", + beta_start=beta_start, + num_train_timesteps=num_train_timesteps, + skip_prk_steps=True, + steps_offset=1, + ) + elif args.scheduler_type == "lms": + scheduler = LMSDiscreteScheduler(beta_start=beta_start, beta_end=beta_end, beta_schedule="scaled_linear") + elif args.scheduler_type == "euler": + scheduler = EulerDiscreteScheduler(beta_start=beta_start, beta_end=beta_end, beta_schedule="scaled_linear") + elif args.scheduler_type == "euler-ancestral": + scheduler = EulerAncestralDiscreteScheduler( + beta_start=beta_start, beta_end=beta_end, beta_schedule="scaled_linear" + ) + elif args.scheduler_type == "dpm": + scheduler = DPMSolverMultistepScheduler( + beta_start=beta_start, beta_end=beta_end, beta_schedule="scaled_linear" + ) + elif args.scheduler_type == "ddim": + scheduler = DDIMScheduler( + beta_start=beta_start, + beta_end=beta_end, + beta_schedule="scaled_linear", + clip_sample=False, + set_alpha_to_one=False, + steps_offset=1, + ) + else: + raise ValueError(f"Scheduler of type {args.scheduler_type} doesn't exist!") + + # Convert the UNet2DConditionModel models. + if args.unet_checkpoint_path is not None: + # image UNet + image_unet_config = create_image_unet_diffusers_config(IMAGE_UNET_CONFIG) + checkpoint = torch.load(args.unet_checkpoint_path) + converted_image_unet_checkpoint = convert_vd_unet_checkpoint( + checkpoint, image_unet_config, unet_key="model.diffusion_model.unet_image.", extract_ema=args.extract_ema + ) + image_unet = UNet2DConditionModel(**image_unet_config) + image_unet.load_state_dict(converted_image_unet_checkpoint) + + # text UNet + text_unet_config = create_text_unet_diffusers_config(TEXT_UNET_CONFIG) + converted_text_unet_checkpoint = convert_vd_unet_checkpoint( + checkpoint, text_unet_config, unet_key="model.diffusion_model.unet_text.", extract_ema=args.extract_ema + ) + text_unet = UNetFlatConditionModel(**text_unet_config) + text_unet.load_state_dict(converted_text_unet_checkpoint) + + # Convert the VAE model. + if args.vae_checkpoint_path is not None: + vae_config = create_vae_diffusers_config(AUTOENCODER_CONFIG) + checkpoint = torch.load(args.vae_checkpoint_path) + converted_vae_checkpoint = convert_vd_vae_checkpoint(checkpoint, vae_config) + + vae = AutoencoderKL(**vae_config) + vae.load_state_dict(converted_vae_checkpoint) + + tokenizer = CLIPTokenizer.from_pretrained("openai/clip-vit-large-patch14") + image_feature_extractor = CLIPFeatureExtractor.from_pretrained("openai/clip-vit-large-patch14") + text_encoder = CLIPTextModelWithProjection.from_pretrained("openai/clip-vit-large-patch14") + image_encoder = CLIPVisionModelWithProjection.from_pretrained("openai/clip-vit-large-patch14") + + pipe = VersatileDiffusionPipeline( + scheduler=scheduler, + tokenizer=tokenizer, + image_feature_extractor=image_feature_extractor, + text_encoder=text_encoder, + image_encoder=image_encoder, + image_unet=image_unet, + text_unet=text_unet, + vae=vae, + ) + pipe.save_pretrained(args.dump_path) diff --git a/setup.py b/setup.py index d0aff10da6..a4c336669e 100644 --- a/setup.py +++ b/setup.py @@ -212,7 +212,7 @@ install_requires = [ setup( name="diffusers", - version="0.8.0.dev0", # expected format is one of x.y.z.dev0, or x.y.z.rc1 or x.y.z (no to dashes, yes to dots) + version="0.8.0", # expected format is one of x.y.z.dev0, or x.y.z.rc1 or x.y.z (no to dashes, yes to dots) description="Diffusers", long_description=open("README.md", "r", encoding="utf-8").read(), long_description_content_type="text/markdown", diff --git a/src/diffusers/__init__.py b/src/diffusers/__init__.py index 42cb2cb585..a1faa28000 100644 --- a/src/diffusers/__init__.py +++ b/src/diffusers/__init__.py @@ -9,7 +9,7 @@ from .utils import ( ) -__version__ = "0.8.0.dev0" +__version__ = "0.8.0" from .configuration_utils import ConfigMixin from .onnx_utils import OnnxRuntimeModel @@ -69,10 +69,16 @@ if is_torch_available() and is_transformers_available(): AltDiffusionPipeline, CycleDiffusionPipeline, LDMTextToImagePipeline, + StableDiffusionImageVariationPipeline, StableDiffusionImg2ImgPipeline, StableDiffusionInpaintPipeline, StableDiffusionInpaintPipelineLegacy, StableDiffusionPipeline, + StableDiffusionPipelineSafe, + VersatileDiffusionDualGuidedPipeline, + VersatileDiffusionImageVariationPipeline, + VersatileDiffusionPipeline, + VersatileDiffusionTextToImagePipeline, VQDiffusionPipeline, ) else: @@ -82,6 +88,7 @@ if is_torch_available() and is_transformers_available() and is_onnx_available(): from .pipelines import ( OnnxStableDiffusionImg2ImgPipeline, OnnxStableDiffusionInpaintPipeline, + OnnxStableDiffusionInpaintPipelineLegacy, OnnxStableDiffusionPipeline, StableDiffusionOnnxPipeline, ) diff --git a/src/diffusers/modeling_utils.py b/src/diffusers/modeling_utils.py index 1e91ccd56a..704ba00cad 100644 --- a/src/diffusers/modeling_utils.py +++ b/src/diffusers/modeling_utils.py @@ -332,7 +332,7 @@ class ModelMixin(torch.nn.Module): if low_cpu_mem_usage and not is_accelerate_available(): low_cpu_mem_usage = False - logger.warn( + logger.warning( "Cannot initialize model with low cpu memory usage because `accelerate` was not found in the" " environment. Defaulting to `low_cpu_mem_usage=False`. It is strongly recommended to install" " `accelerate` for faster and less memory-intense model loading. You can do so with: \n```\npip" diff --git a/src/diffusers/models/attention.py b/src/diffusers/models/attention.py index be9203b4d6..0aacddf34d 100644 --- a/src/diffusers/models/attention.py +++ b/src/diffusers/models/attention.py @@ -12,6 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. import math +import warnings from dataclasses import dataclass from typing import Optional @@ -98,8 +99,10 @@ class Transformer2DModel(ModelMixin, ConfigMixin): num_vector_embeds: Optional[int] = None, activation_fn: str = "geglu", num_embeds_ada_norm: Optional[int] = None, + use_linear_projection: bool = False, ): super().__init__() + self.use_linear_projection = use_linear_projection self.num_attention_heads = num_attention_heads self.attention_head_dim = attention_head_dim inner_dim = num_attention_heads * attention_head_dim @@ -125,7 +128,10 @@ class Transformer2DModel(ModelMixin, ConfigMixin): self.in_channels = in_channels self.norm = torch.nn.GroupNorm(num_groups=norm_num_groups, num_channels=in_channels, eps=1e-6, affine=True) - self.proj_in = nn.Conv2d(in_channels, inner_dim, kernel_size=1, stride=1, padding=0) + if use_linear_projection: + self.proj_in = nn.Linear(in_channels, inner_dim) + else: + self.proj_in = nn.Conv2d(in_channels, inner_dim, kernel_size=1, stride=1, padding=0) elif self.is_input_vectorized: assert sample_size is not None, "Transformer2DModel over discrete input must provide sample_size" assert num_vector_embeds is not None, "Transformer2DModel over discrete input must provide num_embed" @@ -158,7 +164,10 @@ class Transformer2DModel(ModelMixin, ConfigMixin): # 4. Define output layers if self.is_input_continuous: - self.proj_out = nn.Conv2d(inner_dim, in_channels, kernel_size=1, stride=1, padding=0) + if use_linear_projection: + self.proj_out = nn.Linear(in_channels, inner_dim) + else: + self.proj_out = nn.Conv2d(inner_dim, in_channels, kernel_size=1, stride=1, padding=0) elif self.is_input_vectorized: self.norm_out = nn.LayerNorm(inner_dim) self.out = nn.Linear(inner_dim, self.num_vector_embeds - 1) @@ -190,10 +199,16 @@ class Transformer2DModel(ModelMixin, ConfigMixin): if self.is_input_continuous: batch, channel, height, weight = hidden_states.shape residual = hidden_states + hidden_states = self.norm(hidden_states) - hidden_states = self.proj_in(hidden_states) - inner_dim = hidden_states.shape[1] - hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(batch, height * weight, inner_dim) + if not self.use_linear_projection: + hidden_states = self.proj_in(hidden_states) + inner_dim = hidden_states.shape[1] + hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(batch, height * weight, inner_dim) + else: + inner_dim = hidden_states.shape[1] + hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(batch, height * weight, inner_dim) + hidden_states = self.proj_in(hidden_states) elif self.is_input_vectorized: hidden_states = self.latent_image_embedding(hidden_states) @@ -203,8 +218,13 @@ class Transformer2DModel(ModelMixin, ConfigMixin): # 3. Output if self.is_input_continuous: - hidden_states = hidden_states.reshape(batch, height, weight, inner_dim).permute(0, 3, 1, 2) - hidden_states = self.proj_out(hidden_states) + if not self.use_linear_projection: + hidden_states = hidden_states.reshape(batch, height, weight, inner_dim).permute(0, 3, 1, 2) + hidden_states = self.proj_out(hidden_states) + else: + hidden_states = self.proj_out(hidden_states) + hidden_states = hidden_states.reshape(batch, height, weight, inner_dim).permute(0, 3, 1, 2) + output = hidden_states + residual elif self.is_input_vectorized: hidden_states = self.norm_out(hidden_states) @@ -284,22 +304,52 @@ class AttentionBlock(nn.Module): key_proj = self.key(hidden_states) value_proj = self.value(hidden_states) - # transpose - query_states = self.transpose_for_scores(query_proj) - key_states = self.transpose_for_scores(key_proj) - value_states = self.transpose_for_scores(value_proj) + scale = 1 / math.sqrt(self.channels / self.num_heads) # get scores - scale = 1 / math.sqrt(math.sqrt(self.channels / self.num_heads)) - attention_scores = torch.matmul(query_states * scale, key_states.transpose(-1, -2) * scale) # TODO: use baddmm + if self.num_heads > 1: + query_states = self.transpose_for_scores(query_proj) + key_states = self.transpose_for_scores(key_proj) + value_states = self.transpose_for_scores(value_proj) + + # TODO: is there a way to perform batched matmul (e.g. baddbmm) on 4D tensors? + # or reformulate this into a 3D problem? + # TODO: measure whether on MPS device it would be faster to do this matmul via einsum + # as some matmuls can be 1.94x slower than an equivalent einsum on MPS + # https://gist.github.com/Birch-san/cba16789ec27bb20996a4b4831b13ce0 + attention_scores = torch.matmul(query_states, key_states.transpose(-1, -2)) * scale + else: + query_states, key_states, value_states = query_proj, key_proj, value_proj + + attention_scores = torch.baddbmm( + torch.empty( + query_states.shape[0], + query_states.shape[1], + key_states.shape[1], + dtype=query_states.dtype, + device=query_states.device, + ), + query_states, + key_states.transpose(-1, -2), + beta=0, + alpha=scale, + ) + attention_probs = torch.softmax(attention_scores.float(), dim=-1).type(attention_scores.dtype) # compute attention output - hidden_states = torch.matmul(attention_probs, value_states) - - hidden_states = hidden_states.permute(0, 2, 1, 3).contiguous() - new_hidden_states_shape = hidden_states.size()[:-2] + (self.channels,) - hidden_states = hidden_states.view(new_hidden_states_shape) + if self.num_heads > 1: + # TODO: is there a way to perform batched matmul (e.g. bmm) on 4D tensors? + # or reformulate this into a 3D problem? + # TODO: measure whether on MPS device it would be faster to do this matmul via einsum + # as some matmuls can be 1.94x slower than an equivalent einsum on MPS + # https://gist.github.com/Birch-san/cba16789ec27bb20996a4b4831b13ce0 + hidden_states = torch.matmul(attention_probs, value_states) + hidden_states = hidden_states.permute(0, 2, 1, 3).contiguous() + new_hidden_states_shape = hidden_states.size()[:-2] + (self.channels,) + hidden_states = hidden_states.view(new_hidden_states_shape) + else: + hidden_states = torch.bmm(attention_probs, value_states) # compute next hidden_states hidden_states = self.proj_attn(hidden_states) @@ -366,6 +416,16 @@ class BasicTransformerBlock(nn.Module): self.norm2 = nn.LayerNorm(dim) self.norm3 = nn.LayerNorm(dim) + # if xformers is installed try to use memory_efficient_attention by default + if is_xformers_available(): + try: + self._set_use_memory_efficient_attention_xformers(True) + except Exception as e: + warnings.warn( + "Could not enable memory efficient attention. Make sure xformers is installed" + f" correctly and a GPU is available: {e}" + ) + def _set_attention_slice(self, slice_size): self.attn1._slice_size = slice_size self.attn2._slice_size = slice_size @@ -507,19 +567,17 @@ class CrossAttention(nn.Module): return hidden_states def _attention(self, query, key, value): - # TODO: use baddbmm for better performance - if query.device.type == "mps": - # Better performance on mps (~20-25%) - attention_scores = torch.einsum("b i d, b j d -> b i j", query, key) * self.scale - else: - attention_scores = torch.matmul(query, key.transpose(-1, -2)) * self.scale + attention_scores = torch.baddbmm( + torch.empty(query.shape[0], query.shape[1], key.shape[1], dtype=query.dtype, device=query.device), + query, + key.transpose(-1, -2), + beta=0, + alpha=self.scale, + ) attention_probs = attention_scores.softmax(dim=-1) # compute attention output - if query.device.type == "mps": - hidden_states = torch.einsum("b i j, b j d -> b i d", attention_probs, value) - else: - hidden_states = torch.matmul(attention_probs, value) + hidden_states = torch.bmm(attention_probs, value) # reshape hidden_states hidden_states = self.reshape_batch_dim_to_heads(hidden_states) @@ -534,21 +592,15 @@ class CrossAttention(nn.Module): for i in range(hidden_states.shape[0] // slice_size): start_idx = i * slice_size end_idx = (i + 1) * slice_size - if query.device.type == "mps": - # Better performance on mps (~20-25%) - attn_slice = ( - torch.einsum("b i d, b j d -> b i j", query[start_idx:end_idx], key[start_idx:end_idx]) - * self.scale - ) - else: - attn_slice = ( - torch.matmul(query[start_idx:end_idx], key[start_idx:end_idx].transpose(1, 2)) * self.scale - ) # TODO: use baddbmm for better performance + attn_slice = torch.baddbmm( + torch.empty(slice_size, query.shape[1], key.shape[1], dtype=query.dtype, device=query.device), + query[start_idx:end_idx], + key[start_idx:end_idx].transpose(-1, -2), + beta=0, + alpha=self.scale, + ) attn_slice = attn_slice.softmax(dim=-1) - if query.device.type == "mps": - attn_slice = torch.einsum("b i j, b j d -> b i d", attn_slice, value[start_idx:end_idx]) - else: - attn_slice = torch.matmul(attn_slice, value[start_idx:end_idx]) + attn_slice = torch.bmm(attn_slice, value[start_idx:end_idx]) hidden_states[start_idx:end_idx] = attn_slice @@ -666,3 +718,129 @@ class AdaLayerNorm(nn.Module): scale, shift = torch.chunk(emb, 2) x = self.norm(x) * (1 + scale) + shift return x + + +class DualTransformer2DModel(nn.Module): + """ + Dual transformer wrapper that combines two `Transformer2DModel`s for mixed inference. + + Parameters: + num_attention_heads (`int`, *optional*, defaults to 16): The number of heads to use for multi-head attention. + attention_head_dim (`int`, *optional*, defaults to 88): The number of channels in each head. + in_channels (`int`, *optional*): + Pass if the input is continuous. The number of channels in the input and output. + num_layers (`int`, *optional*, defaults to 1): The number of layers of Transformer blocks to use. + dropout (`float`, *optional*, defaults to 0.1): The dropout probability to use. + cross_attention_dim (`int`, *optional*): The number of context dimensions to use. + sample_size (`int`, *optional*): Pass if the input is discrete. The width of the latent images. + Note that this is fixed at training time as it is used for learning a number of position embeddings. See + `ImagePositionalEmbeddings`. + num_vector_embeds (`int`, *optional*): + Pass if the input is discrete. The number of classes of the vector embeddings of the latent pixels. + Includes the class for the masked latent pixel. + activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward. + num_embeds_ada_norm ( `int`, *optional*): Pass if at least one of the norm_layers is `AdaLayerNorm`. + The number of diffusion steps used during training. Note that this is fixed at training time as it is used + to learn a number of embeddings that are added to the hidden states. During inference, you can denoise for + up to but not more than steps than `num_embeds_ada_norm`. + attention_bias (`bool`, *optional*): + Configure if the TransformerBlocks' attention should contain a bias parameter. + """ + + def __init__( + self, + num_attention_heads: int = 16, + attention_head_dim: int = 88, + in_channels: Optional[int] = None, + num_layers: int = 1, + dropout: float = 0.0, + norm_num_groups: int = 32, + cross_attention_dim: Optional[int] = None, + attention_bias: bool = False, + sample_size: Optional[int] = None, + num_vector_embeds: Optional[int] = None, + activation_fn: str = "geglu", + num_embeds_ada_norm: Optional[int] = None, + ): + super().__init__() + self.transformers = nn.ModuleList( + [ + Transformer2DModel( + num_attention_heads=num_attention_heads, + attention_head_dim=attention_head_dim, + in_channels=in_channels, + num_layers=num_layers, + dropout=dropout, + norm_num_groups=norm_num_groups, + cross_attention_dim=cross_attention_dim, + attention_bias=attention_bias, + sample_size=sample_size, + num_vector_embeds=num_vector_embeds, + activation_fn=activation_fn, + num_embeds_ada_norm=num_embeds_ada_norm, + ) + for _ in range(2) + ] + ) + + # Variables that can be set by a pipeline: + + # The ratio of transformer1 to transformer2's output states to be combined during inference + self.mix_ratio = 0.5 + + # The shape of `encoder_hidden_states` is expected to be + # `(batch_size, condition_lengths[0]+condition_lengths[1], num_features)` + self.condition_lengths = [77, 257] + + # Which transformer to use to encode which condition. + # E.g. `(1, 0)` means that we'll use `transformers[1](conditions[0])` and `transformers[0](conditions[1])` + self.transformer_index_for_condition = [1, 0] + + def forward(self, hidden_states, encoder_hidden_states, timestep=None, return_dict: bool = True): + """ + Args: + hidden_states ( When discrete, `torch.LongTensor` of shape `(batch size, num latent pixels)`. + When continuous, `torch.FloatTensor` of shape `(batch size, channel, height, width)`): Input + hidden_states + encoder_hidden_states ( `torch.LongTensor` of shape `(batch size, context dim)`, *optional*): + Conditional embeddings for cross attention layer. If not given, cross-attention defaults to + self-attention. + timestep ( `torch.long`, *optional*): + Optional timestep to be applied as an embedding in AdaLayerNorm's. Used to indicate denoising step. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`models.unet_2d_condition.UNet2DConditionOutput`] instead of a plain tuple. + + Returns: + [`~models.attention.Transformer2DModelOutput`] or `tuple`: [`~models.attention.Transformer2DModelOutput`] + if `return_dict` is True, otherwise a `tuple`. When returning a tuple, the first element is the sample + tensor. + """ + input_states = hidden_states + + encoded_states = [] + tokens_start = 0 + for i in range(2): + # for each of the two transformers, pass the corresponding condition tokens + condition_state = encoder_hidden_states[:, tokens_start : tokens_start + self.condition_lengths[i]] + transformer_index = self.transformer_index_for_condition[i] + encoded_state = self.transformers[transformer_index](input_states, condition_state, timestep, return_dict)[ + 0 + ] + encoded_states.append(encoded_state - input_states) + tokens_start += self.condition_lengths[i] + + output_states = encoded_states[0] * self.mix_ratio + encoded_states[1] * (1 - self.mix_ratio) + output_states = output_states + input_states + + if not return_dict: + return (output_states,) + + return Transformer2DModelOutput(sample=output_states) + + def _set_attention_slice(self, slice_size): + for transformer in self.transformers: + transformer._set_attention_slice(slice_size) + + def _set_use_memory_efficient_attention_xformers(self, use_memory_efficient_attention_xformers: bool): + for transformer in self.transformers: + transformer._set_use_memory_efficient_attention_xformers(use_memory_efficient_attention_xformers) diff --git a/src/diffusers/models/unet_2d.py b/src/diffusers/models/unet_2d.py index 0432405760..5b337f482c 100644 --- a/src/diffusers/models/unet_2d.py +++ b/src/diffusers/models/unet_2d.py @@ -43,8 +43,8 @@ class UNet2DModel(ModelMixin, ConfigMixin): implements for all the model (such as downloading or saving, etc.) Parameters: - sample_size (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`, *optional*): - Input sample size. + sample_size (`int` or `Tuple[int, int]`, *optional*, defaults to `None`): + Height and width of input/output sample. in_channels (`int`, *optional*, defaults to 3): Number of channels in the input image. out_channels (`int`, *optional*, defaults to 3): Number of channels in the output. center_input_sample (`bool`, *optional*, defaults to `False`): Whether to center the input sample. @@ -71,7 +71,7 @@ class UNet2DModel(ModelMixin, ConfigMixin): @register_to_config def __init__( self, - sample_size: Optional[int] = None, + sample_size: Optional[Union[int, Tuple[int, int]]] = None, in_channels: int = 3, out_channels: int = 3, center_input_sample: bool = False, @@ -175,7 +175,7 @@ class UNet2DModel(ModelMixin, ConfigMixin): num_groups_out = norm_num_groups if norm_num_groups is not None else min(block_out_channels[0] // 4, 32) self.conv_norm_out = nn.GroupNorm(num_channels=block_out_channels[0], num_groups=num_groups_out, eps=norm_eps) self.conv_act = nn.SiLU() - self.conv_out = nn.Conv2d(block_out_channels[0], out_channels, 3, padding=1) + self.conv_out = nn.Conv2d(block_out_channels[0], out_channels, kernel_size=3, padding=1) def forward( self, @@ -209,6 +209,11 @@ class UNet2DModel(ModelMixin, ConfigMixin): timesteps = timesteps * torch.ones(sample.shape[0], dtype=timesteps.dtype, device=timesteps.device) t_emb = self.time_proj(timesteps) + + # timesteps does not contain any weights and will always return f32 tensors + # but time_embedding might actually be running in fp16. so we need to cast here. + # there might be better ways to encapsulate this. + t_emb = t_emb.to(dtype=self.dtype) emb = self.time_embedding(t_emb) # 2. pre-process @@ -242,9 +247,7 @@ class UNet2DModel(ModelMixin, ConfigMixin): sample = upsample_block(sample, res_samples, emb) # 6. post-process - # make sure hidden states is in float32 - # when running in half-precision - sample = self.conv_norm_out(sample.float()).type(sample.dtype) + sample = self.conv_norm_out(sample) sample = self.conv_act(sample) sample = self.conv_out(sample) diff --git a/src/diffusers/models/unet_2d_blocks.py b/src/diffusers/models/unet_2d_blocks.py index 770043f053..5a8a97187f 100644 --- a/src/diffusers/models/unet_2d_blocks.py +++ b/src/diffusers/models/unet_2d_blocks.py @@ -15,7 +15,7 @@ import numpy as np import torch from torch import nn -from .attention import AttentionBlock, Transformer2DModel +from .attention import AttentionBlock, DualTransformer2DModel, Transformer2DModel from .resnet import Downsample2D, FirDownsample2D, FirUpsample2D, ResnetBlock2D, Upsample2D @@ -32,6 +32,8 @@ def get_down_block( resnet_groups=None, cross_attention_dim=None, downsample_padding=None, + dual_cross_attention=False, + use_linear_projection=False, ): down_block_type = down_block_type[7:] if down_block_type.startswith("UNetRes") else down_block_type if down_block_type == "DownBlock2D": @@ -74,6 +76,8 @@ def get_down_block( downsample_padding=downsample_padding, cross_attention_dim=cross_attention_dim, attn_num_head_channels=attn_num_head_channels, + dual_cross_attention=dual_cross_attention, + use_linear_projection=use_linear_projection, ) elif down_block_type == "SkipDownBlock2D": return SkipDownBlock2D( @@ -137,6 +141,8 @@ def get_up_block( attn_num_head_channels, resnet_groups=None, cross_attention_dim=None, + dual_cross_attention=False, + use_linear_projection=False, ): up_block_type = up_block_type[7:] if up_block_type.startswith("UNetRes") else up_block_type if up_block_type == "UpBlock2D": @@ -166,6 +172,8 @@ def get_up_block( resnet_groups=resnet_groups, cross_attention_dim=cross_attention_dim, attn_num_head_channels=attn_num_head_channels, + dual_cross_attention=dual_cross_attention, + use_linear_projection=use_linear_projection, ) elif up_block_type == "AttnUpBlock2D": return AttnUpBlock2D( @@ -322,6 +330,8 @@ class UNetMidBlock2DCrossAttn(nn.Module): attention_type="default", output_scale_factor=1.0, cross_attention_dim=1280, + dual_cross_attention=False, + use_linear_projection=False, **kwargs, ): super().__init__() @@ -348,16 +358,29 @@ class UNetMidBlock2DCrossAttn(nn.Module): attentions = [] for _ in range(num_layers): - attentions.append( - Transformer2DModel( - attn_num_head_channels, - in_channels // attn_num_head_channels, - in_channels=in_channels, - num_layers=1, - cross_attention_dim=cross_attention_dim, - norm_num_groups=resnet_groups, + if not dual_cross_attention: + attentions.append( + Transformer2DModel( + attn_num_head_channels, + in_channels // attn_num_head_channels, + in_channels=in_channels, + num_layers=1, + cross_attention_dim=cross_attention_dim, + norm_num_groups=resnet_groups, + use_linear_projection=use_linear_projection, + ) + ) + else: + attentions.append( + DualTransformer2DModel( + attn_num_head_channels, + in_channels // attn_num_head_channels, + in_channels=in_channels, + num_layers=1, + cross_attention_dim=cross_attention_dim, + norm_num_groups=resnet_groups, + ) ) - ) resnets.append( ResnetBlock2D( in_channels=in_channels, @@ -505,6 +528,8 @@ class CrossAttnDownBlock2D(nn.Module): output_scale_factor=1.0, downsample_padding=1, add_downsample=True, + dual_cross_attention=False, + use_linear_projection=False, ): super().__init__() resnets = [] @@ -529,16 +554,29 @@ class CrossAttnDownBlock2D(nn.Module): pre_norm=resnet_pre_norm, ) ) - attentions.append( - Transformer2DModel( - attn_num_head_channels, - out_channels // attn_num_head_channels, - in_channels=out_channels, - num_layers=1, - cross_attention_dim=cross_attention_dim, - norm_num_groups=resnet_groups, + if not dual_cross_attention: + attentions.append( + Transformer2DModel( + attn_num_head_channels, + out_channels // attn_num_head_channels, + in_channels=out_channels, + num_layers=1, + cross_attention_dim=cross_attention_dim, + norm_num_groups=resnet_groups, + use_linear_projection=use_linear_projection, + ) + ) + else: + attentions.append( + DualTransformer2DModel( + attn_num_head_channels, + out_channels // attn_num_head_channels, + in_channels=out_channels, + num_layers=1, + cross_attention_dim=cross_attention_dim, + norm_num_groups=resnet_groups, + ) ) - ) self.attentions = nn.ModuleList(attentions) self.resnets = nn.ModuleList(resnets) @@ -1089,6 +1127,8 @@ class CrossAttnUpBlock2D(nn.Module): attention_type="default", output_scale_factor=1.0, add_upsample=True, + dual_cross_attention=False, + use_linear_projection=False, ): super().__init__() resnets = [] @@ -1115,16 +1155,29 @@ class CrossAttnUpBlock2D(nn.Module): pre_norm=resnet_pre_norm, ) ) - attentions.append( - Transformer2DModel( - attn_num_head_channels, - out_channels // attn_num_head_channels, - in_channels=out_channels, - num_layers=1, - cross_attention_dim=cross_attention_dim, - norm_num_groups=resnet_groups, + if not dual_cross_attention: + attentions.append( + Transformer2DModel( + attn_num_head_channels, + out_channels // attn_num_head_channels, + in_channels=out_channels, + num_layers=1, + cross_attention_dim=cross_attention_dim, + norm_num_groups=resnet_groups, + use_linear_projection=use_linear_projection, + ) + ) + else: + attentions.append( + DualTransformer2DModel( + attn_num_head_channels, + out_channels // attn_num_head_channels, + in_channels=out_channels, + num_layers=1, + cross_attention_dim=cross_attention_dim, + norm_num_groups=resnet_groups, + ) ) - ) self.attentions = nn.ModuleList(attentions) self.resnets = nn.ModuleList(resnets) diff --git a/src/diffusers/models/unet_2d_condition.py b/src/diffusers/models/unet_2d_condition.py index c3f2fb87b6..2060971493 100644 --- a/src/diffusers/models/unet_2d_condition.py +++ b/src/diffusers/models/unet_2d_condition.py @@ -56,11 +56,12 @@ class UNet2DConditionModel(ModelMixin, ConfigMixin): implements for all the models (such as downloading or saving, etc.) Parameters: - sample_size (`int`, *optional*): The size of the input sample. + sample_size (`int` or `Tuple[int, int]`, *optional*, defaults to `None`): + Height and width of input/output sample. in_channels (`int`, *optional*, defaults to 4): The number of channels in the input sample. out_channels (`int`, *optional*, defaults to 4): The number of channels in the output. center_input_sample (`bool`, *optional*, defaults to `False`): Whether to center the input sample. - flip_sin_to_cos (`bool`, *optional*, defaults to `True`): + flip_sin_to_cos (`bool`, *optional*, defaults to `False`): Whether to flip the sin to cos in the time embedding. freq_shift (`int`, *optional*, defaults to 0): The frequency shift to apply to the time embedding. down_block_types (`Tuple[str]`, *optional*, defaults to `("CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "DownBlock2D")`): @@ -105,7 +106,9 @@ class UNet2DConditionModel(ModelMixin, ConfigMixin): norm_num_groups: int = 32, norm_eps: float = 1e-5, cross_attention_dim: int = 1280, - attention_head_dim: int = 8, + attention_head_dim: Union[int, Tuple[int]] = 8, + dual_cross_attention: bool = False, + use_linear_projection: bool = False, ): super().__init__() @@ -125,6 +128,9 @@ class UNet2DConditionModel(ModelMixin, ConfigMixin): self.mid_block = None self.up_blocks = nn.ModuleList([]) + if isinstance(attention_head_dim, int): + attention_head_dim = (attention_head_dim,) * len(down_block_types) + # down output_channel = block_out_channels[0] for i, down_block_type in enumerate(down_block_types): @@ -143,8 +149,10 @@ class UNet2DConditionModel(ModelMixin, ConfigMixin): resnet_act_fn=act_fn, resnet_groups=norm_num_groups, cross_attention_dim=cross_attention_dim, - attn_num_head_channels=attention_head_dim, + attn_num_head_channels=attention_head_dim[i], downsample_padding=downsample_padding, + dual_cross_attention=dual_cross_attention, + use_linear_projection=use_linear_projection, ) self.down_blocks.append(down_block) @@ -157,8 +165,10 @@ class UNet2DConditionModel(ModelMixin, ConfigMixin): output_scale_factor=mid_block_scale_factor, resnet_time_scale_shift="default", cross_attention_dim=cross_attention_dim, - attn_num_head_channels=attention_head_dim, + attn_num_head_channels=attention_head_dim[-1], resnet_groups=norm_num_groups, + dual_cross_attention=dual_cross_attention, + use_linear_projection=use_linear_projection, ) # count how many layers upsample the images @@ -166,6 +176,7 @@ class UNet2DConditionModel(ModelMixin, ConfigMixin): # up reversed_block_out_channels = list(reversed(block_out_channels)) + reversed_attention_head_dim = list(reversed(attention_head_dim)) output_channel = reversed_block_out_channels[0] for i, up_block_type in enumerate(up_block_types): is_final_block = i == len(block_out_channels) - 1 @@ -193,7 +204,9 @@ class UNet2DConditionModel(ModelMixin, ConfigMixin): resnet_act_fn=act_fn, resnet_groups=norm_num_groups, cross_attention_dim=cross_attention_dim, - attn_num_head_channels=attention_head_dim, + attn_num_head_channels=reversed_attention_head_dim[i], + dual_cross_attention=dual_cross_attention, + use_linear_projection=use_linear_projection, ) self.up_blocks.append(up_block) prev_output_channel = output_channel @@ -201,7 +214,7 @@ class UNet2DConditionModel(ModelMixin, ConfigMixin): # out self.conv_norm_out = nn.GroupNorm(num_channels=block_out_channels[0], num_groups=norm_num_groups, eps=norm_eps) self.conv_act = nn.SiLU() - self.conv_out = nn.Conv2d(block_out_channels[0], out_channels, 3, padding=1) + self.conv_out = nn.Conv2d(block_out_channels[0], out_channels, kernel_size=3, padding=1) def set_attention_slice(self, slice_size): if slice_size is not None and self.config.attention_head_dim % slice_size != 0: @@ -251,8 +264,7 @@ class UNet2DConditionModel(ModelMixin, ConfigMixin): Args: sample (`torch.FloatTensor`): (batch, channel, height, width) noisy inputs tensor timestep (`torch.FloatTensor` or `float` or `int`): (batch) timesteps - encoder_hidden_states (`torch.FloatTensor`): - (batch_size, sequence_length, hidden_size) encoder hidden states + encoder_hidden_states (`torch.FloatTensor`): (batch, channel, height, width) encoder hidden states return_dict (`bool`, *optional*, defaults to `True`): Whether or not to return a [`models.unet_2d_condition.UNet2DConditionOutput`] instead of a plain tuple. diff --git a/src/diffusers/pipeline_flax_utils.py b/src/diffusers/pipeline_flax_utils.py index 54bb028139..bf2e259ea1 100644 --- a/src/diffusers/pipeline_flax_utils.py +++ b/src/diffusers/pipeline_flax_utils.py @@ -411,13 +411,13 @@ class FlaxDiffusionPipeline(ConfigMixin): f" {expected_class_obj}" ) elif passed_class_obj[name] is None: - logger.warn( + logger.warning( f"You have passed `None` for {name} to disable its functionality in {pipeline_class}. Note" f" that this might lead to problems when using {pipeline_class} and is not recommended." ) sub_model_should_be_defined = False else: - logger.warn( + logger.warning( f"You have passed a non-standard module {passed_class_obj[name]}. We cannot verify whether it" " has the correct type" ) diff --git a/src/diffusers/pipeline_utils.py b/src/diffusers/pipeline_utils.py index cf2bbb980e..3f2857fa4f 100644 --- a/src/diffusers/pipeline_utils.py +++ b/src/diffusers/pipeline_utils.py @@ -405,7 +405,7 @@ class DiffusionPipeline(ConfigMixin): if low_cpu_mem_usage and not is_accelerate_available(): low_cpu_mem_usage = False - logger.warn( + logger.warning( "Cannot initialize model with low cpu memory usage because `accelerate` was not found in the" " environment. Defaulting to `low_cpu_mem_usage=False`. It is strongly recommended to install" " `accelerate` for faster and less memory-intense model loading. You can do so with: \n```\npip" @@ -571,13 +571,13 @@ class DiffusionPipeline(ConfigMixin): f" {expected_class_obj}" ) elif passed_class_obj[name] is None: - logger.warn( + logger.warning( f"You have passed `None` for {name} to disable its functionality in {pipeline_class}. Note" f" that this might lead to problems when using {pipeline_class} and is not recommended." ) sub_model_should_be_defined = False else: - logger.warn( + logger.warning( f"You have passed a non-standard module {passed_class_obj[name]}. We cannot verify whether it" " has the correct type" ) diff --git a/src/diffusers/pipelines/__init__.py b/src/diffusers/pipelines/__init__.py index 3ca66b28b5..9f4cef4b73 100644 --- a/src/diffusers/pipelines/__init__.py +++ b/src/diffusers/pipelines/__init__.py @@ -19,17 +19,26 @@ if is_torch_available() and is_transformers_available(): from .latent_diffusion import LDMTextToImagePipeline from .stable_diffusion import ( CycleDiffusionPipeline, + StableDiffusionImageVariationPipeline, StableDiffusionImg2ImgPipeline, StableDiffusionInpaintPipeline, StableDiffusionInpaintPipelineLegacy, StableDiffusionPipeline, ) + from .stable_diffusion_safe import StableDiffusionPipelineSafe + from .versatile_diffusion import ( + VersatileDiffusionDualGuidedPipeline, + VersatileDiffusionImageVariationPipeline, + VersatileDiffusionPipeline, + VersatileDiffusionTextToImagePipeline, + ) from .vq_diffusion import VQDiffusionPipeline if is_transformers_available() and is_onnx_available(): from .stable_diffusion import ( OnnxStableDiffusionImg2ImgPipeline, OnnxStableDiffusionInpaintPipeline, + OnnxStableDiffusionInpaintPipelineLegacy, OnnxStableDiffusionPipeline, StableDiffusionOnnxPipeline, ) diff --git a/src/diffusers/pipelines/alt_diffusion/pipeline_alt_diffusion.py b/src/diffusers/pipelines/alt_diffusion/pipeline_alt_diffusion.py index afb2c52886..246f2b8720 100644 --- a/src/diffusers/pipelines/alt_diffusion/pipeline_alt_diffusion.py +++ b/src/diffusers/pipelines/alt_diffusion/pipeline_alt_diffusion.py @@ -115,7 +115,7 @@ class AltDiffusionPipeline(DiffusionPipeline): scheduler._internal_dict = FrozenDict(new_config) if safety_checker is None: - logger.warn( + logger.warning( f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" " that you abide to the conditions of the Alt Diffusion license and do not expose unfiltered" " results in services or applications open to the public. Both the diffusers team and Hugging Face" diff --git a/src/diffusers/pipelines/alt_diffusion/pipeline_alt_diffusion_img2img.py b/src/diffusers/pipelines/alt_diffusion/pipeline_alt_diffusion_img2img.py index fc530603ee..7fc1658ea0 100644 --- a/src/diffusers/pipelines/alt_diffusion/pipeline_alt_diffusion_img2img.py +++ b/src/diffusers/pipelines/alt_diffusion/pipeline_alt_diffusion_img2img.py @@ -81,7 +81,6 @@ class AltDiffusionImg2ImgPipeline(DiffusionPipeline): Model that extracts features from generated images to be used as inputs for the `safety_checker`. """ - # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.AltDiffusionPipeline.__init__ def __init__( self, vae: AutoencoderKL, @@ -129,7 +128,7 @@ class AltDiffusionImg2ImgPipeline(DiffusionPipeline): scheduler._internal_dict = FrozenDict(new_config) if safety_checker is None: - logger.warn( + logger.warning( f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" " that you abide to the conditions of the Alt Diffusion license and do not expose unfiltered" " results in services or applications open to the public. Both the diffusers team and Hugging Face" @@ -148,7 +147,6 @@ class AltDiffusionImg2ImgPipeline(DiffusionPipeline): feature_extractor=feature_extractor, ) - # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.AltDiffusionPipeline.enable_attention_slicing def enable_attention_slicing(self, slice_size: Optional[Union[str, int]] = "auto"): r""" Enable sliced attention computation. @@ -168,7 +166,6 @@ class AltDiffusionImg2ImgPipeline(DiffusionPipeline): slice_size = self.unet.config.attention_head_dim // 2 self.unet.set_attention_slice(slice_size) - # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.AltDiffusionPipeline.disable_attention_slicing def disable_attention_slicing(self): r""" Disable sliced attention computation. If `enable_attention_slicing` was previously invoked, this method will go @@ -177,7 +174,6 @@ class AltDiffusionImg2ImgPipeline(DiffusionPipeline): # set slice_size = `None` to disable `attention slicing` self.enable_attention_slicing(None) - # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.AltDiffusionPipeline.enable_sequential_cpu_offload def enable_sequential_cpu_offload(self, gpu_id=0): r""" Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet, @@ -196,7 +192,6 @@ class AltDiffusionImg2ImgPipeline(DiffusionPipeline): cpu_offload(cpu_offloaded_model, device) @property - # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.AltDiffusionPipeline._execution_device def _execution_device(self): r""" Returns the device on which the pipeline's models will be executed. After calling @@ -214,7 +209,6 @@ class AltDiffusionImg2ImgPipeline(DiffusionPipeline): return torch.device(module._hf_hook.execution_device) return self.device - # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.AltDiffusionPipeline.enable_xformers_memory_efficient_attention def enable_xformers_memory_efficient_attention(self): r""" Enable memory efficient attention as implemented in xformers. @@ -227,14 +221,12 @@ class AltDiffusionImg2ImgPipeline(DiffusionPipeline): """ self.unet.set_use_memory_efficient_attention_xformers(True) - # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.AltDiffusionPipeline.disable_xformers_memory_efficient_attention def disable_xformers_memory_efficient_attention(self): r""" Disable memory efficient attention as implemented in xformers. """ self.unet.set_use_memory_efficient_attention_xformers(False) - # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.AltDiffusionPipeline._encode_prompt def _encode_prompt(self, prompt, device, num_images_per_prompt, do_classifier_free_guidance, negative_prompt): r""" Encodes the prompt into text encoder hidden states. @@ -340,7 +332,6 @@ class AltDiffusionImg2ImgPipeline(DiffusionPipeline): return text_embeddings - # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.AltDiffusionPipeline.run_safety_checker def run_safety_checker(self, image, device, dtype): if self.safety_checker is not None: safety_checker_input = self.feature_extractor(self.numpy_to_pil(image), return_tensors="pt").to(device) @@ -351,7 +342,6 @@ class AltDiffusionImg2ImgPipeline(DiffusionPipeline): has_nsfw_concept = None return image, has_nsfw_concept - # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.AltDiffusionPipeline.decode_latents def decode_latents(self, latents): latents = 1 / 0.18215 * latents image = self.vae.decode(latents).sample @@ -360,7 +350,6 @@ class AltDiffusionImg2ImgPipeline(DiffusionPipeline): image = image.cpu().permute(0, 2, 3, 1).float().numpy() return image - # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.AltDiffusionPipeline.prepare_extra_step_kwargs def prepare_extra_step_kwargs(self, generator, eta): # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. diff --git a/src/diffusers/pipelines/ddim/pipeline_ddim.py b/src/diffusers/pipelines/ddim/pipeline_ddim.py index 6db6298329..b9e590dea6 100644 --- a/src/diffusers/pipelines/ddim/pipeline_ddim.py +++ b/src/diffusers/pipelines/ddim/pipeline_ddim.py @@ -89,7 +89,11 @@ class DDIMPipeline(DiffusionPipeline): generator = None # Sample gaussian noise to begin loop - image_shape = (batch_size, self.unet.in_channels, self.unet.sample_size, self.unet.sample_size) + if isinstance(self.unet.sample_size, int): + image_shape = (batch_size, self.unet.in_channels, self.unet.sample_size, self.unet.sample_size) + else: + image_shape = (batch_size, self.unet.in_channels, *self.unet.sample_size) + if self.device.type == "mps": # randn does not work reproducibly on mps image = torch.randn(image_shape, generator=generator) diff --git a/src/diffusers/pipelines/ddpm/pipeline_ddpm.py b/src/diffusers/pipelines/ddpm/pipeline_ddpm.py index c937a23003..634e1c0f99 100644 --- a/src/diffusers/pipelines/ddpm/pipeline_ddpm.py +++ b/src/diffusers/pipelines/ddpm/pipeline_ddpm.py @@ -94,7 +94,11 @@ class DDPMPipeline(DiffusionPipeline): generator = None # Sample gaussian noise to begin loop - image_shape = (batch_size, self.unet.in_channels, self.unet.sample_size, self.unet.sample_size) + if isinstance(self.unet.sample_size, int): + image_shape = (batch_size, self.unet.in_channels, self.unet.sample_size, self.unet.sample_size) + else: + image_shape = (batch_size, self.unet.in_channels, *self.unet.sample_size) + if self.device.type == "mps": # randn does not work reproducibly on mps image = torch.randn(image_shape, generator=generator) diff --git a/src/diffusers/pipelines/stable_diffusion/__init__.py b/src/diffusers/pipelines/stable_diffusion/__init__.py index 6623929f86..3c012dbab8 100644 --- a/src/diffusers/pipelines/stable_diffusion/__init__.py +++ b/src/diffusers/pipelines/stable_diffusion/__init__.py @@ -6,7 +6,14 @@ import numpy as np import PIL from PIL import Image -from ...utils import BaseOutput, is_flax_available, is_onnx_available, is_torch_available, is_transformers_available +from ...utils import ( + BaseOutput, + is_flax_available, + is_onnx_available, + is_torch_available, + is_transformers_available, + is_transformers_version, +) @dataclass @@ -35,10 +42,16 @@ if is_transformers_available() and is_torch_available(): from .pipeline_stable_diffusion_inpaint_legacy import StableDiffusionInpaintPipelineLegacy from .safety_checker import StableDiffusionSafetyChecker +if is_transformers_available() and is_torch_available() and is_transformers_version(">=", "4.25.0.dev0"): + from .pipeline_stable_diffusion_image_variation import StableDiffusionImageVariationPipeline +else: + from ...utils.dummy_torch_and_transformers_objects import StableDiffusionImageVariationPipeline + if is_transformers_available() and is_onnx_available(): from .pipeline_onnx_stable_diffusion import OnnxStableDiffusionPipeline, StableDiffusionOnnxPipeline from .pipeline_onnx_stable_diffusion_img2img import OnnxStableDiffusionImg2ImgPipeline from .pipeline_onnx_stable_diffusion_inpaint import OnnxStableDiffusionInpaintPipeline + from .pipeline_onnx_stable_diffusion_inpaint_legacy import OnnxStableDiffusionInpaintPipelineLegacy if is_transformers_available() and is_flax_available(): import flax diff --git a/src/diffusers/pipelines/stable_diffusion/pipeline_cycle_diffusion.py b/src/diffusers/pipelines/stable_diffusion/pipeline_cycle_diffusion.py index 2b3cf8fa95..8d702b1b02 100644 --- a/src/diffusers/pipelines/stable_diffusion/pipeline_cycle_diffusion.py +++ b/src/diffusers/pipelines/stable_diffusion/pipeline_cycle_diffusion.py @@ -160,7 +160,7 @@ class CycleDiffusionPipeline(DiffusionPipeline): scheduler._internal_dict = FrozenDict(new_config) if safety_checker is None: - logger.warn( + logger.warning( f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" " results in services or applications open to the public. Both the diffusers team and Hugging Face" diff --git a/src/diffusers/pipelines/stable_diffusion/pipeline_flax_stable_diffusion.py b/src/diffusers/pipelines/stable_diffusion/pipeline_flax_stable_diffusion.py index 02943997d9..9c668d5e51 100644 --- a/src/diffusers/pipelines/stable_diffusion/pipeline_flax_stable_diffusion.py +++ b/src/diffusers/pipelines/stable_diffusion/pipeline_flax_stable_diffusion.py @@ -88,7 +88,7 @@ class FlaxStableDiffusionPipeline(FlaxDiffusionPipeline): self.dtype = dtype if safety_checker is None: - logger.warn( + logger.warning( f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" " results in services or applications open to the public. Both the diffusers team and Hugging Face" @@ -165,6 +165,7 @@ class FlaxStableDiffusionPipeline(FlaxDiffusionPipeline): guidance_scale: float = 7.5, latents: Optional[jnp.array] = None, debug: bool = False, + neg_prompt_ids: jnp.array = None, ): if height % 8 != 0 or width % 8 != 0: raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") @@ -177,10 +178,14 @@ class FlaxStableDiffusionPipeline(FlaxDiffusionPipeline): batch_size = prompt_ids.shape[0] max_length = prompt_ids.shape[-1] - uncond_input = self.tokenizer( - [""] * batch_size, padding="max_length", max_length=max_length, return_tensors="np" - ) - uncond_embeddings = self.text_encoder(uncond_input.input_ids, params=params["text_encoder"])[0] + + if neg_prompt_ids is None: + uncond_input = self.tokenizer( + [""] * batch_size, padding="max_length", max_length=max_length, return_tensors="np" + ).input_ids + else: + uncond_input = neg_prompt_ids + uncond_embeddings = self.text_encoder(uncond_input, params=params["text_encoder"])[0] context = jnp.concatenate([uncond_embeddings, text_embeddings]) latents_shape = (batch_size, self.unet.in_channels, height // 8, width // 8) @@ -251,6 +256,7 @@ class FlaxStableDiffusionPipeline(FlaxDiffusionPipeline): return_dict: bool = True, jit: bool = False, debug: bool = False, + neg_prompt_ids: jnp.array = None, **kwargs, ): r""" @@ -298,11 +304,30 @@ class FlaxStableDiffusionPipeline(FlaxDiffusionPipeline): """ if jit: images = _p_generate( - self, prompt_ids, params, prng_seed, num_inference_steps, height, width, guidance_scale, latents, debug + self, + prompt_ids, + params, + prng_seed, + num_inference_steps, + height, + width, + guidance_scale, + latents, + debug, + neg_prompt_ids, ) else: images = self._generate( - prompt_ids, params, prng_seed, num_inference_steps, height, width, guidance_scale, latents, debug + prompt_ids, + params, + prng_seed, + num_inference_steps, + height, + width, + guidance_scale, + latents, + debug, + neg_prompt_ids, ) if self.safety_checker is not None: @@ -333,10 +358,29 @@ class FlaxStableDiffusionPipeline(FlaxDiffusionPipeline): # TODO: maybe use a config dict instead of so many static argnums @partial(jax.pmap, static_broadcasted_argnums=(0, 4, 5, 6, 7, 9)) def _p_generate( - pipe, prompt_ids, params, prng_seed, num_inference_steps, height, width, guidance_scale, latents, debug + pipe, + prompt_ids, + params, + prng_seed, + num_inference_steps, + height, + width, + guidance_scale, + latents, + debug, + neg_prompt_ids, ): return pipe._generate( - prompt_ids, params, prng_seed, num_inference_steps, height, width, guidance_scale, latents, debug + prompt_ids, + params, + prng_seed, + num_inference_steps, + height, + width, + guidance_scale, + latents, + debug, + neg_prompt_ids, ) diff --git a/src/diffusers/pipelines/stable_diffusion/pipeline_onnx_stable_diffusion_inpaint.py b/src/diffusers/pipelines/stable_diffusion/pipeline_onnx_stable_diffusion_inpaint.py index ede30b5563..b933c52bf6 100644 --- a/src/diffusers/pipelines/stable_diffusion/pipeline_onnx_stable_diffusion_inpaint.py +++ b/src/diffusers/pipelines/stable_diffusion/pipeline_onnx_stable_diffusion_inpaint.py @@ -408,9 +408,9 @@ class OnnxStableDiffusionInpaintPipeline(DiffusionPipeline): # expand the latents if we are doing classifier free guidance latent_model_input = np.concatenate([latents] * 2) if do_classifier_free_guidance else latents # concat latents, mask, masked_image_latnets in the channel dimension - latent_model_input = np.concatenate([latent_model_input, mask, masked_image_latents], axis=1) latent_model_input = self.scheduler.scale_model_input(torch.from_numpy(latent_model_input), t) latent_model_input = latent_model_input.cpu().numpy() + latent_model_input = np.concatenate([latent_model_input, mask, masked_image_latents], axis=1) # predict the noise residual timestep = np.array([t], dtype=timestep_dtype) diff --git a/src/diffusers/pipelines/stable_diffusion/pipeline_onnx_stable_diffusion_inpaint_legacy.py b/src/diffusers/pipelines/stable_diffusion/pipeline_onnx_stable_diffusion_inpaint_legacy.py new file mode 100644 index 0000000000..34f1d0e95d --- /dev/null +++ b/src/diffusers/pipelines/stable_diffusion/pipeline_onnx_stable_diffusion_inpaint_legacy.py @@ -0,0 +1,447 @@ +import inspect +from typing import Callable, List, Optional, Union + +import numpy as np +import torch + +import PIL +from transformers import CLIPFeatureExtractor, CLIPTokenizer + +from ...configuration_utils import FrozenDict +from ...onnx_utils import OnnxRuntimeModel +from ...pipeline_utils import DiffusionPipeline +from ...schedulers import DDIMScheduler, LMSDiscreteScheduler, PNDMScheduler +from ...utils import deprecate, logging +from . import StableDiffusionPipelineOutput + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +def preprocess(image): + w, h = image.size + w, h = map(lambda x: x - x % 32, (w, h)) # resize to integer multiple of 32 + image = image.resize((w, h), resample=PIL.Image.LANCZOS) + image = np.array(image).astype(np.float32) / 255.0 + image = image[None].transpose(0, 3, 1, 2) + return 2.0 * image - 1.0 + + +def preprocess_mask(mask): + mask = mask.convert("L") + w, h = mask.size + w, h = map(lambda x: x - x % 32, (w, h)) # resize to integer multiple of 32 + mask = mask.resize((w // 8, h // 8), resample=PIL.Image.NEAREST) + mask = np.array(mask).astype(np.float32) / 255.0 + mask = np.tile(mask, (4, 1, 1)) + mask = mask[None].transpose(0, 1, 2, 3) # what does this step do? + mask = 1 - mask # repaint white, keep black + return mask + + +class OnnxStableDiffusionInpaintPipelineLegacy(DiffusionPipeline): + r""" + Pipeline for text-guided image inpainting using Stable Diffusion. This is a *legacy feature* for Onnx pipelines to + provide compatibility with StableDiffusionInpaintPipelineLegacy and may be removed in the future. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModel`]): + Frozen text-encoder. Stable Diffusion uses the text portion of + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically + the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + safety_checker ([`StableDiffusionSafetyChecker`]): + Classification module that estimates whether generated images could be considered offensive or harmful. + Please, refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for details. + feature_extractor ([`CLIPFeatureExtractor`]): + Model that extracts features from generated images to be used as inputs for the `safety_checker`. + """ + vae_encoder: OnnxRuntimeModel + vae_decoder: OnnxRuntimeModel + text_encoder: OnnxRuntimeModel + tokenizer: CLIPTokenizer + unet: OnnxRuntimeModel + scheduler: Union[DDIMScheduler, PNDMScheduler, LMSDiscreteScheduler] + safety_checker: OnnxRuntimeModel + feature_extractor: CLIPFeatureExtractor + + def __init__( + self, + vae_encoder: OnnxRuntimeModel, + vae_decoder: OnnxRuntimeModel, + text_encoder: OnnxRuntimeModel, + tokenizer: CLIPTokenizer, + unet: OnnxRuntimeModel, + scheduler: Union[DDIMScheduler, PNDMScheduler, LMSDiscreteScheduler], + safety_checker: OnnxRuntimeModel, + feature_extractor: CLIPFeatureExtractor, + ): + super().__init__() + + if hasattr(scheduler.config, "steps_offset") and scheduler.config.steps_offset != 1: + deprecation_message = ( + f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`" + f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure " + "to update the config accordingly as leaving `steps_offset` might led to incorrect results" + " in future versions. If you have downloaded this checkpoint from the Hugging Face Hub," + " it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`" + " file" + ) + deprecate("steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(scheduler.config) + new_config["steps_offset"] = 1 + scheduler._internal_dict = FrozenDict(new_config) + + if hasattr(scheduler.config, "clip_sample") and scheduler.config.clip_sample is True: + deprecation_message = ( + f"The configuration file of this scheduler: {scheduler} has not set the configuration `clip_sample`." + " `clip_sample` should be set to False in the configuration file. Please make sure to update the" + " config accordingly as not setting `clip_sample` in the config might lead to incorrect results in" + " future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it would be very" + " nice if you could open a Pull request for the `scheduler/scheduler_config.json` file" + ) + deprecate("clip_sample not set", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(scheduler.config) + new_config["clip_sample"] = False + scheduler._internal_dict = FrozenDict(new_config) + + if safety_checker is None: + logger.warning( + f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" + " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" + " results in services or applications open to the public. Both the diffusers team and Hugging Face" + " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling" + " it only for use-cases that involve analyzing network behavior or auditing its results. For more" + " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ." + ) + + self.register_modules( + vae_encoder=vae_encoder, + vae_decoder=vae_decoder, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + ) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_onnx_stable_diffusion.OnnxStableDiffusionPipeline._encode_prompt + def _encode_prompt(self, prompt, num_images_per_prompt, do_classifier_free_guidance, negative_prompt): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list(int)`): + prompt to be encoded + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `List[str]`): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `guidance_scale` is less than `1`). + """ + batch_size = len(prompt) if isinstance(prompt, list) else 1 + + # get prompt text embeddings + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="np", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="max_length", return_tensors="np").input_ids + + if not np.array_equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer.batch_decode(untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + text_embeddings = self.text_encoder(input_ids=text_input_ids.astype(np.int32))[0] + text_embeddings = np.repeat(text_embeddings, num_images_per_prompt, axis=0) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance: + uncond_tokens: List[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] * batch_size + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + max_length = text_input_ids.shape[-1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="np", + ) + uncond_embeddings = self.text_encoder(input_ids=uncond_input.input_ids.astype(np.int32))[0] + uncond_embeddings = np.repeat(uncond_embeddings, num_images_per_prompt, axis=0) + + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + text_embeddings = np.concatenate([uncond_embeddings, text_embeddings]) + + return text_embeddings + + def __call__( + self, + prompt: Union[str, List[str]], + init_image: Union[np.ndarray, PIL.Image.Image], + mask_image: Union[np.ndarray, PIL.Image.Image], + strength: float = 0.8, + num_inference_steps: Optional[int] = 50, + guidance_scale: Optional[float] = 7.5, + negative_prompt: Optional[Union[str, List[str]]] = None, + num_images_per_prompt: Optional[int] = 1, + eta: Optional[float] = 0.0, + generator: Optional[np.random.RandomState] = None, + output_type: Optional[str] = "pil", + return_dict: bool = True, + callback: Optional[Callable[[int, int, np.ndarray], None]] = None, + callback_steps: Optional[int] = 1, + **kwargs, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `List[str]`): + The prompt or prompts to guide the image generation. + init_image (`nd.ndarray` or `PIL.Image.Image`): + `Image`, or tensor representing an image batch, that will be used as the starting point for the + process. This is the image whose masked region will be inpainted. + mask_image (`nd.ndarray` or `PIL.Image.Image`): + `Image`, or tensor representing an image batch, to mask `init_image`. White pixels in the mask will be + replaced by noise and therefore repainted, while black pixels will be preserved. If `mask_image` is a + PIL image, it will be converted to a single channel (luminance) before use. If it's a tensor, it should + contain one color channel (L) instead of 3, so the expected shape would be `(B, H, W, 1)`.uu + strength (`float`, *optional*, defaults to 0.8): + Conceptually, indicates how much to transform the reference `init_image`. Must be between 0 and 1. + `init_image` will be used as a starting point, adding more noise to it the larger the `strength`. The + number of denoising steps depends on the amount of noise initially added. When `strength` is 1, added + noise will be maximum and the denoising process will run for the full number of iterations specified in + `num_inference_steps`. A value of 1, therefore, essentially ignores `init_image`. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. This parameter will be modulated by `strength`. + guidance_scale (`float`, *optional*, defaults to 7.5): + Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598). + `guidance_scale` is defined as `w` of equation 2. of [Imagen + Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale > + 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`, + usually at the expense of lower image quality. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `guidance_scale` is less than `1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (?) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to + [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`np.random.RandomState`, *optional*): + A np.random.RandomState to make generation deterministic. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + callback (`Callable`, *optional*): + A function that will be called every `callback_steps` steps during inference. The function will be + called with the following arguments: `callback(step: int, timestep: int, latents: np.ndarray)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function will be called. If not specified, the callback will be + called at every step. + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple. + When returning a tuple, the first element is a list with the generated images, and the second element is a + list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work" + (nsfw) content, according to the `safety_checker`. + """ + if isinstance(prompt, str): + batch_size = 1 + elif isinstance(prompt, list): + batch_size = len(prompt) + else: + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if strength < 0 or strength > 1: + raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}") + + if (callback_steps is None) or ( + callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0) + ): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + if generator is None: + generator = np.random + + # set timesteps + self.scheduler.set_timesteps(num_inference_steps) + + if isinstance(init_image, PIL.Image.Image): + init_image = preprocess(init_image) + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = guidance_scale > 1.0 + + text_embeddings = self._encode_prompt( + prompt, num_images_per_prompt, do_classifier_free_guidance, negative_prompt + ) + + latents_dtype = text_embeddings.dtype + init_image = init_image.astype(latents_dtype) + + # encode the init image into latents and scale the latents + init_latents = self.vae_encoder(sample=init_image)[0] + init_latents = 0.18215 * init_latents + + # Expand init_latents for batch_size and num_images_per_prompt + init_latents = np.concatenate([init_latents] * num_images_per_prompt, axis=0) + init_latents_orig = init_latents + + # preprocess mask + if not isinstance(mask_image, np.ndarray): + mask_image = preprocess_mask(mask_image) + mask_image = mask_image.astype(latents_dtype) + mask = np.concatenate([mask_image] * num_images_per_prompt, axis=0) + + # check sizes + if not mask.shape == init_latents.shape: + raise ValueError("The mask and init_image should be the same size!") + + # get the original timestep using init_timestep + offset = self.scheduler.config.get("steps_offset", 0) + init_timestep = int(num_inference_steps * strength) + offset + init_timestep = min(init_timestep, num_inference_steps) + + timesteps = self.scheduler.timesteps.numpy()[-init_timestep] + timesteps = np.array([timesteps] * batch_size * num_images_per_prompt) + + # add noise to latents using the timesteps + noise = generator.randn(*init_latents.shape).astype(latents_dtype) + init_latents = self.scheduler.add_noise( + torch.from_numpy(init_latents), torch.from_numpy(noise), torch.from_numpy(timesteps) + ) + init_latents = init_latents.numpy() + + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (?) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to ? in DDIM paper: https://arxiv.org/abs/2010.02502 + # and should be between [0, 1] + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + latents = init_latents + + t_start = max(num_inference_steps - init_timestep + offset, 0) + timesteps = self.scheduler.timesteps[t_start:].numpy() + + for i, t in enumerate(self.progress_bar(timesteps)): + # expand the latents if we are doing classifier free guidance + latent_model_input = np.concatenate([latents] * 2) if do_classifier_free_guidance else latents + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # predict the noise residual + noise_pred = self.unet( + sample=latent_model_input, timestep=np.array([t]), encoder_hidden_states=text_embeddings + )[0] + + # perform guidance + if do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = np.split(noise_pred, 2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step( + torch.from_numpy(noise_pred), t, torch.from_numpy(latents), **extra_step_kwargs + ).prev_sample + + latents = latents.numpy() + + init_latents_proper = self.scheduler.add_noise( + torch.from_numpy(init_latents_orig), torch.from_numpy(noise), torch.from_numpy(np.array([t])) + ) + + init_latents_proper = init_latents_proper.numpy() + + latents = (init_latents_proper * mask) + (latents * (1 - mask)) + + # call the callback, if provided + if callback is not None and i % callback_steps == 0: + callback(i, t, latents) + + latents = 1 / 0.18215 * latents + # image = self.vae_decoder(latent_sample=latents)[0] + # it seems likes there is a strange result for using half-precision vae decoder if batchsize>1 + image = np.concatenate( + [self.vae_decoder(latent_sample=latents[i : i + 1])[0] for i in range(latents.shape[0])] + ) + + image = np.clip(image / 2 + 0.5, 0, 1) + image = image.transpose((0, 2, 3, 1)) + + if self.safety_checker is not None: + safety_checker_input = self.feature_extractor( + self.numpy_to_pil(image), return_tensors="np" + ).pixel_values.astype(image.dtype) + # There will throw an error if use safety_checker batchsize>1 + images, has_nsfw_concept = [], [] + for i in range(image.shape[0]): + image_i, has_nsfw_concept_i = self.safety_checker( + clip_input=safety_checker_input[i : i + 1], images=image[i : i + 1] + ) + images.append(image_i) + has_nsfw_concept.append(has_nsfw_concept_i[0]) + image = np.concatenate(images) + else: + has_nsfw_concept = None + + if output_type == "pil": + image = self.numpy_to_pil(image) + + if not return_dict: + return (image, has_nsfw_concept) + + return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept) diff --git a/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion.py b/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion.py index 963d75c58b..fbfac6b5a0 100644 --- a/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion.py +++ b/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion.py @@ -114,7 +114,7 @@ class StableDiffusionPipeline(DiffusionPipeline): scheduler._internal_dict = FrozenDict(new_config) if safety_checker is None: - logger.warn( + logger.warning( f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" " results in services or applications open to the public. Both the diffusers team and Hugging Face" diff --git a/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_image_variation.py b/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_image_variation.py new file mode 100644 index 0000000000..4cfa5817af --- /dev/null +++ b/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_image_variation.py @@ -0,0 +1,437 @@ +# Copyright 2022 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Callable, List, Optional, Union + +import torch + +import PIL +from diffusers.utils import is_accelerate_available +from transformers import CLIPFeatureExtractor, CLIPVisionModelWithProjection + +from ...models import AutoencoderKL, UNet2DConditionModel +from ...pipeline_utils import DiffusionPipeline +from ...schedulers import ( + DDIMScheduler, + DPMSolverMultistepScheduler, + EulerAncestralDiscreteScheduler, + EulerDiscreteScheduler, + LMSDiscreteScheduler, + PNDMScheduler, +) +from ...utils import logging +from . import StableDiffusionPipelineOutput +from .safety_checker import StableDiffusionSafetyChecker + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +class StableDiffusionImageVariationPipeline(DiffusionPipeline): + r""" + Pipeline to generate variations from an input image using Stable Diffusion. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + image_encoder ([`CLIPVisionModelWithProjection`]): + Frozen CLIP image-encoder. Stable Diffusion Image Variation uses the vision portion of + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPVisionModelWithProjection), + specifically the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. + unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + safety_checker ([`StableDiffusionSafetyChecker`]): + Classification module that estimates whether generated images could be considered offensive or harmful. + Please, refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for details. + feature_extractor ([`CLIPFeatureExtractor`]): + Model that extracts features from generated images to be used as inputs for the `safety_checker`. + """ + + def __init__( + self, + vae: AutoencoderKL, + image_encoder: CLIPVisionModelWithProjection, + unet: UNet2DConditionModel, + scheduler: Union[ + DDIMScheduler, + PNDMScheduler, + LMSDiscreteScheduler, + EulerDiscreteScheduler, + EulerAncestralDiscreteScheduler, + DPMSolverMultistepScheduler, + ], + safety_checker: StableDiffusionSafetyChecker, + feature_extractor: CLIPFeatureExtractor, + ): + super().__init__() + + if safety_checker is None: + logger.warn( + f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" + " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" + " results in services or applications open to the public. Both the diffusers team and Hugging Face" + " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling" + " it only for use-cases that involve analyzing network behavior or auditing its results. For more" + " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ." + ) + + self.register_modules( + vae=vae, + image_encoder=image_encoder, + unet=unet, + scheduler=scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + ) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_xformers_memory_efficient_attention + def enable_xformers_memory_efficient_attention(self): + r""" + Enable memory efficient attention as implemented in xformers. + + When this option is enabled, you should observe lower GPU memory usage and a potential speed up at inference + time. Speed up at training time is not guaranteed. + + Warning: When Memory Efficient Attention and Sliced attention are both enabled, the Memory Efficient Attention + is used. + """ + self.unet.set_use_memory_efficient_attention_xformers(True) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_xformers_memory_efficient_attention + def disable_xformers_memory_efficient_attention(self): + r""" + Disable memory efficient attention as implemented in xformers. + """ + self.unet.set_use_memory_efficient_attention_xformers(False) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_attention_slicing + def enable_attention_slicing(self, slice_size: Optional[Union[str, int]] = "auto"): + r""" + Enable sliced attention computation. + + When this option is enabled, the attention module will split the input tensor in slices, to compute attention + in several steps. This is useful to save some memory in exchange for a small speed decrease. + + Args: + slice_size (`str` or `int`, *optional*, defaults to `"auto"`): + When `"auto"`, halves the input to the attention heads, so attention will be computed in two steps. If + a number is provided, uses as many slices as `attention_head_dim // slice_size`. In this case, + `attention_head_dim` must be a multiple of `slice_size`. + """ + if slice_size == "auto": + # half the attention head size is usually a good trade-off between + # speed and memory + slice_size = self.unet.config.attention_head_dim // 2 + self.unet.set_attention_slice(slice_size) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_attention_slicing + def disable_attention_slicing(self): + r""" + Disable sliced attention computation. If `enable_attention_slicing` was previously invoked, this method will go + back to computing attention in one step. + """ + # set slice_size = `None` to disable `attention slicing` + self.enable_attention_slicing(None) + + def enable_sequential_cpu_offload(self, gpu_id=0): + r""" + Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet, + text_encoder, vae and safety checker have their state dicts saved to CPU and then are moved to a + `torch.device('meta') and loaded to GPU only when their specific submodule has its `forward` method called. + """ + if is_accelerate_available(): + from accelerate import cpu_offload + else: + raise ImportError("Please install accelerate via `pip install accelerate`") + + device = torch.device(f"cuda:{gpu_id}") + + for cpu_offloaded_model in [self.unet, self.image_encoder, self.vae, self.safety_checker]: + if cpu_offloaded_model is not None: + cpu_offload(cpu_offloaded_model, device) + + @property + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._execution_device + def _execution_device(self): + r""" + Returns the device on which the pipeline's models will be executed. After calling + `pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module + hooks. + """ + if self.device != torch.device("meta") or not hasattr(self.unet, "_hf_hook"): + return self.device + for module in self.unet.modules(): + if ( + hasattr(module, "_hf_hook") + and hasattr(module._hf_hook, "execution_device") + and module._hf_hook.execution_device is not None + ): + return torch.device(module._hf_hook.execution_device) + return self.device + + def _encode_image(self, image, device, num_images_per_prompt, do_classifier_free_guidance): + dtype = next(self.image_encoder.parameters()).dtype + + if not isinstance(image, torch.Tensor): + image = self.feature_extractor(images=image, return_tensors="pt").pixel_values + + image = image.to(device=device, dtype=dtype) + image_embeddings = self.image_encoder(image).image_embeds + image_embeddings = image_embeddings.unsqueeze(1) + + # duplicate image embeddings for each generation per prompt, using mps friendly method + bs_embed, seq_len, _ = image_embeddings.shape + image_embeddings = image_embeddings.repeat(1, num_images_per_prompt, 1) + image_embeddings = image_embeddings.view(bs_embed * num_images_per_prompt, seq_len, -1) + + if do_classifier_free_guidance: + uncond_embeddings = torch.zeros_like(image_embeddings) + + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + image_embeddings = torch.cat([uncond_embeddings, image_embeddings]) + + return image_embeddings + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker + def run_safety_checker(self, image, device, dtype): + if self.safety_checker is not None: + safety_checker_input = self.feature_extractor(self.numpy_to_pil(image), return_tensors="pt").to(device) + image, has_nsfw_concept = self.safety_checker( + images=image, clip_input=safety_checker_input.pixel_values.to(dtype) + ) + else: + has_nsfw_concept = None + return image, has_nsfw_concept + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents + def decode_latents(self, latents): + latents = 1 / 0.18215 * latents + image = self.vae.decode(latents).sample + image = (image / 2 + 0.5).clamp(0, 1) + # we always cast to float32 as this does not cause significant overhead and is compatible with bfloa16 + image = image.cpu().permute(0, 2, 3, 1).float().numpy() + return image + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def check_inputs(self, image, height, width, callback_steps): + if ( + not isinstance(image, torch.Tensor) + and not isinstance(image, PIL.Image.Image) + and not isinstance(image, list) + ): + raise ValueError( + f"`image` has to be of type `torch.FloatTensor` or `PIL.Image.Image` or `list` but is {type(image)}" + ) + + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if (callback_steps is None) or ( + callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0) + ): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = (batch_size, num_channels_latents, height // 8, width // 8) + if latents is None: + if device.type == "mps": + # randn does not work reproducibly on mps + latents = torch.randn(shape, generator=generator, device="cpu", dtype=dtype).to(device) + else: + latents = torch.randn(shape, generator=generator, device=device, dtype=dtype) + else: + if latents.shape != shape: + raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}") + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + @torch.no_grad() + def __call__( + self, + image: Union[PIL.Image.Image, List[PIL.Image.Image], torch.FloatTensor], + height: int = 512, + width: int = 512, + num_inference_steps: int = 50, + guidance_scale: float = 7.5, + num_images_per_prompt: Optional[int] = 1, + eta: float = 0.0, + generator: Optional[torch.Generator] = None, + latents: Optional[torch.FloatTensor] = None, + output_type: Optional[str] = "pil", + return_dict: bool = True, + callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None, + callback_steps: Optional[int] = 1, + **kwargs, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + image (`PIL.Image.Image` or `List[PIL.Image.Image]` or `torch.FloatTensor`): + The image or images to guide the image generation. If you provide a tensor, it needs to comply with the + configuration of + [this](https://huggingface.co/lambdalabs/sd-image-variations-diffusers/blob/main/feature_extractor/preprocessor_config.json) + `CLIPFeatureExtractor` + height (`int`, *optional*, defaults to 512): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to 512): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 7.5): + Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598). + `guidance_scale` is defined as `w` of equation 2. of [Imagen + Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale > + 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`, + usually at the expense of lower image quality. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to + [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`torch.Generator`, *optional*): + A [torch generator](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make generation + deterministic. + latents (`torch.FloatTensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will ge generated by sampling using the supplied random `generator`. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + callback (`Callable`, *optional*): + A function that will be called every `callback_steps` steps during inference. The function will be + called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function will be called. If not specified, the callback will be + called at every step. + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple. + When returning a tuple, the first element is a list with the generated images, and the second element is a + list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work" + (nsfw) content, according to the `safety_checker`. + """ + + # 1. Check inputs. Raise error if not correct + self.check_inputs(image, height, width, callback_steps) + + # 2. Define call parameters + if isinstance(image, PIL.Image.Image): + batch_size = 1 + elif isinstance(image, list): + batch_size = len(image) + else: + batch_size = image.shape[0] + device = self._execution_device + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = guidance_scale > 1.0 + + # 3. Encode input image + image_embeddings = self._encode_image(image, device, num_images_per_prompt, do_classifier_free_guidance) + + # 4. Prepare timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + # 5. Prepare latent variables + num_channels_latents = self.unet.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + image_embeddings.dtype, + device, + generator, + latents, + ) + + # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 7. Denoising loop + for i, t in enumerate(self.progress_bar(timesteps)): + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # predict the noise residual + noise_pred = self.unet(latent_model_input, t, encoder_hidden_states=image_embeddings).sample + + # perform guidance + if do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample + + # call the callback, if provided + if callback is not None and i % callback_steps == 0: + callback(i, t, latents) + + # 8. Post-processing + image = self.decode_latents(latents) + + # 9. Run safety checker + image, has_nsfw_concept = self.run_safety_checker(image, device, image_embeddings.dtype) + + # 10. Convert to PIL + if output_type == "pil": + image = self.numpy_to_pil(image) + + if not return_dict: + return (image, has_nsfw_concept) + + return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept) diff --git a/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_img2img.py b/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_img2img.py index f543d564fe..7efd39e726 100644 --- a/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_img2img.py +++ b/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_img2img.py @@ -127,7 +127,7 @@ class StableDiffusionImg2ImgPipeline(DiffusionPipeline): scheduler._internal_dict = FrozenDict(new_config) if safety_checker is None: - logger.warn( + logger.warning( f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" " results in services or applications open to the public. Both the diffusers team and Hugging Face" diff --git a/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_inpaint.py b/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_inpaint.py index e85e238699..9eb8de2482 100644 --- a/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_inpaint.py +++ b/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_inpaint.py @@ -35,16 +35,88 @@ logger = logging.get_logger(__name__) # pylint: disable=invalid-name def prepare_mask_and_masked_image(image, mask): - image = np.array(image.convert("RGB")) - image = image[None].transpose(0, 3, 1, 2) - image = torch.from_numpy(image).to(dtype=torch.float32) / 127.5 - 1.0 + """ + Prepares a pair (image, mask) to be consumed by the Stable Diffusion pipeline. This means that those inputs will be + converted to ``torch.Tensor`` with shapes ``batch x channels x height x width`` where ``channels`` is ``3`` for the + ``image`` and ``1`` for the ``mask``. - mask = np.array(mask.convert("L")) - mask = mask.astype(np.float32) / 255.0 - mask = mask[None, None] - mask[mask < 0.5] = 0 - mask[mask >= 0.5] = 1 - mask = torch.from_numpy(mask) + The ``image`` will be converted to ``torch.float32`` and normalized to be in ``[-1, 1]``. The ``mask`` will be + binarized (``mask > 0.5``) and cast to ``torch.float32`` too. + + Args: + image (Union[np.array, PIL.Image, torch.Tensor]): The image to inpaint. + It can be a ``PIL.Image``, or a ``height x width x 3`` ``np.array`` or a ``channels x height x width`` + ``torch.Tensor`` or a ``batch x channels x height x width`` ``torch.Tensor``. + mask (_type_): The mask to apply to the image, i.e. regions to inpaint. + It can be a ``PIL.Image``, or a ``height x width`` ``np.array`` or a ``1 x height x width`` + ``torch.Tensor`` or a ``batch x 1 x height x width`` ``torch.Tensor``. + + + Raises: + ValueError: ``torch.Tensor`` images should be in the ``[-1, 1]`` range. ValueError: ``torch.Tensor`` mask + should be in the ``[0, 1]`` range. ValueError: ``mask`` and ``image`` should have the same spatial dimensions. + TypeError: ``mask`` is a ``torch.Tensor`` but ``image`` is not + (ot the other way around). + + Returns: + tuple[torch.Tensor]: The pair (mask, masked_image) as ``torch.Tensor`` with 4 + dimensions: ``batch x channels x height x width``. + """ + if isinstance(image, torch.Tensor): + if not isinstance(mask, torch.Tensor): + raise TypeError(f"`image` is a torch.Tensor but `mask` (type: {type(mask)} is not") + + # Batch single image + if image.ndim == 3: + assert image.shape[0] == 3, "Image outside a batch should be of shape (3, H, W)" + image = image.unsqueeze(0) + + # Batch and add channel dim for single mask + if mask.ndim == 2: + mask = mask.unsqueeze(0).unsqueeze(0) + + # Batch single mask or add channel dim + if mask.ndim == 3: + # Single batched mask, no channel dim or single mask not batched but channel dim + if mask.shape[0] == 1: + mask = mask.unsqueeze(0) + + # Batched masks no channel dim + else: + mask = mask.unsqueeze(1) + + assert image.ndim == 4 and mask.ndim == 4, "Image and Mask must have 4 dimensions" + assert image.shape[-2:] == mask.shape[-2:], "Image and Mask must have the same spatial dimensions" + assert image.shape[0] == mask.shape[0], "Image and Mask must have the same batch size" + + # Check image is in [-1, 1] + if image.min() < -1 or image.max() > 1: + raise ValueError("Image should be in [-1, 1] range") + + # Check mask is in [0, 1] + if mask.min() < 0 or mask.max() > 1: + raise ValueError("Mask should be in [0, 1] range") + + # Binarize mask + mask[mask < 0.5] = 0 + mask[mask >= 0.5] = 1 + + # Image as float32 + image = image.to(dtype=torch.float32) + elif isinstance(mask, torch.Tensor): + raise TypeError(f"`mask` is a torch.Tensor but `image` (type: {type(image)} is not") + else: + if isinstance(image, PIL.Image.Image): + image = np.array(image.convert("RGB")) + image = image[None].transpose(0, 3, 1, 2) + image = torch.from_numpy(image).to(dtype=torch.float32) / 127.5 - 1.0 + if isinstance(mask, PIL.Image.Image): + mask = np.array(mask.convert("L")) + mask = mask.astype(np.float32) / 255.0 + mask = mask[None, None] + mask[mask < 0.5] = 0 + mask[mask >= 0.5] = 1 + mask = torch.from_numpy(mask) masked_image = image * (mask < 0.5) @@ -120,7 +192,7 @@ class StableDiffusionInpaintPipeline(DiffusionPipeline): scheduler._internal_dict = FrozenDict(new_config) if safety_checker is None: - logger.warn( + logger.warning( f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" " results in services or applications open to the public. Both the diffusers team and Hugging Face" @@ -586,9 +658,8 @@ class StableDiffusionInpaintPipeline(DiffusionPipeline): latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents # concat latents, mask, masked_image_latents in the channel dimension - latent_model_input = torch.cat([latent_model_input, mask, masked_image_latents], dim=1) - latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + latent_model_input = torch.cat([latent_model_input, mask, masked_image_latents], dim=1) # predict the noise residual noise_pred = self.unet(latent_model_input, t, encoder_hidden_states=text_embeddings).sample diff --git a/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_inpaint_legacy.py b/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_inpaint_legacy.py index 77e903ff68..003b2668e7 100644 --- a/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_inpaint_legacy.py +++ b/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_inpaint_legacy.py @@ -140,7 +140,7 @@ class StableDiffusionInpaintPipelineLegacy(DiffusionPipeline): scheduler._internal_dict = FrozenDict(new_config) if safety_checker is None: - logger.warn( + logger.warning( f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" " results in services or applications open to the public. Both the diffusers team and Hugging Face" diff --git a/src/diffusers/pipelines/stable_diffusion_safe/__init__.py b/src/diffusers/pipelines/stable_diffusion_safe/__init__.py new file mode 100644 index 0000000000..59ff61fa3b --- /dev/null +++ b/src/diffusers/pipelines/stable_diffusion_safe/__init__.py @@ -0,0 +1,72 @@ +from dataclasses import dataclass +from enum import Enum +from typing import List, Optional, Union + +import numpy as np + +import PIL +from PIL import Image + +from ...utils import BaseOutput, is_torch_available, is_transformers_available + + +@dataclass +class SafetyConfig(object): + WEAK = { + "sld_warmup_steps": 15, + "sld_guidance_scale": 20, + "sld_threshold": 0.0, + "sld_momentum_scale": 0.0, + "sld_mom_beta": 0.0, + } + MEDIUM = { + "sld_warmup_steps": 10, + "sld_guidance_scale": 1000, + "sld_threshold": 0.01, + "sld_momentum_scale": 0.3, + "sld_mom_beta": 0.4, + } + STRONG = { + "sld_warmup_steps": 7, + "sld_guidance_scale": 2000, + "sld_threshold": 0.025, + "sld_momentum_scale": 0.5, + "sld_mom_beta": 0.7, + } + MAX = { + "sld_warmup_steps": 0, + "sld_guidance_scale": 5000, + "sld_threshold": 1.0, + "sld_momentum_scale": 0.5, + "sld_mom_beta": 0.7, + } + + +@dataclass +class StableDiffusionSafePipelineOutput(BaseOutput): + """ + Output class for Safe Stable Diffusion pipelines. + + Args: + images (`List[PIL.Image.Image]` or `np.ndarray`) + List of denoised PIL images of length `batch_size` or numpy array of shape `(batch_size, height, width, + num_channels)`. PIL images or numpy array present the denoised images of the diffusion pipeline. + nsfw_content_detected (`List[bool]`) + List of flags denoting whether the corresponding generated image likely represents "not-safe-for-work" + (nsfw) content, or `None` if safety checking could not be performed. + images (`List[PIL.Image.Image]` or `np.ndarray`) + List of denoised PIL images that were flagged by the safety checker any may contain "not-safe-for-work" + (nsfw) content, or `None` if no safety check was performed or no images were flagged. + applied_safety_concept (`str`) + The safety concept that was applied for safety guidance, or `None` if safety guidance was disabled + """ + + images: Union[List[PIL.Image.Image], np.ndarray] + nsfw_content_detected: Optional[List[bool]] + unsafe_images: Optional[Union[List[PIL.Image.Image], np.ndarray]] + applied_safety_concept: Optional[str] + + +if is_transformers_available() and is_torch_available(): + from .pipeline_stable_diffusion_safe import StableDiffusionPipelineSafe + from .safety_checker import SafeStableDiffusionSafetyChecker diff --git a/src/diffusers/pipelines/stable_diffusion_safe/pipeline_stable_diffusion_safe.py b/src/diffusers/pipelines/stable_diffusion_safe/pipeline_stable_diffusion_safe.py new file mode 100644 index 0000000000..cfa71b9242 --- /dev/null +++ b/src/diffusers/pipelines/stable_diffusion_safe/pipeline_stable_diffusion_safe.py @@ -0,0 +1,721 @@ +import inspect +import warnings +from typing import Callable, List, Optional, Union + +import numpy as np +import torch + +from transformers import CLIPFeatureExtractor, CLIPTextModel, CLIPTokenizer + +from ...configuration_utils import FrozenDict +from ...models import AutoencoderKL, UNet2DConditionModel +from ...pipeline_utils import DiffusionPipeline +from ...schedulers import ( + DDIMScheduler, + DPMSolverMultistepScheduler, + EulerAncestralDiscreteScheduler, + EulerDiscreteScheduler, + LMSDiscreteScheduler, + PNDMScheduler, +) +from ...utils import deprecate, is_accelerate_available, logging +from . import StableDiffusionSafePipelineOutput +from .safety_checker import SafeStableDiffusionSafetyChecker + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +class StableDiffusionPipelineSafe(DiffusionPipeline): + r""" + Pipeline for text-to-image generation using Safe Latent Diffusion. + + The implementation is based on the [`StableDiffusionPipeline`] + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModel`]): + Frozen text-encoder. Stable Diffusion uses the text portion of + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically + the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + safety_checker ([`StableDiffusionSafetyChecker`]): + Classification module that estimates whether generated images could be considered offensive or harmful. + Please, refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for details. + feature_extractor ([`CLIPFeatureExtractor`]): + Model that extracts features from generated images to be used as inputs for the `safety_checker`. + """ + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + unet: UNet2DConditionModel, + scheduler: Union[ + DDIMScheduler, + DPMSolverMultistepScheduler, + EulerAncestralDiscreteScheduler, + EulerDiscreteScheduler, + LMSDiscreteScheduler, + PNDMScheduler, + ], + safety_checker: SafeStableDiffusionSafetyChecker, + feature_extractor: CLIPFeatureExtractor, + ): + super().__init__() + safety_concept: Optional[str] = ( + "an image showing hate, harassment, violence, suffering, humiliation, harm, suicide, sexual, nudity," + " bodily fluids, blood, obscene gestures, illegal activity, drug use, theft, vandalism, weapons, child" + " abuse, brutality, cruelty" + ) + + if hasattr(scheduler.config, "steps_offset") and scheduler.config.steps_offset != 1: + deprecation_message = ( + f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`" + f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure " + "to update the config accordingly as leaving `steps_offset` might led to incorrect results" + " in future versions. If you have downloaded this checkpoint from the Hugging Face Hub," + " it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`" + " file" + ) + deprecate("steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(scheduler.config) + new_config["steps_offset"] = 1 + scheduler._internal_dict = FrozenDict(new_config) + + if hasattr(scheduler.config, "clip_sample") and scheduler.config.clip_sample is True: + deprecation_message = ( + f"The configuration file of this scheduler: {scheduler} has not set the configuration `clip_sample`." + " `clip_sample` should be set to False in the configuration file. Please make sure to update the" + " config accordingly as not setting `clip_sample` in the config might lead to incorrect results in" + " future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it would be very" + " nice if you could open a Pull request for the `scheduler/scheduler_config.json` file" + ) + deprecate("clip_sample not set", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(scheduler.config) + new_config["clip_sample"] = False + scheduler._internal_dict = FrozenDict(new_config) + + if safety_checker is None: + logger.warning( + f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" + " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" + " results in services or applications open to the public. Both the diffusers team and Hugging Face" + " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling" + " it only for use-cases that involve analyzing network behavior or auditing its results. For more" + " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ." + ) + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + ) + self._safety_text_concept = safety_concept + + @property + def safety_concept(self): + r""" + Getter method for the safety concept used with SLD + + Returns: + `str`: The text describing the safety concept + """ + return self._safety_text_concept + + @safety_concept.setter + def safety_concept(self, concept): + r""" + Setter method for the safety concept used with SLD + + Args: + concept (`str`): + The text of the new safety concept + """ + self._safety_text_concept = concept + + def enable_xformers_memory_efficient_attention(self): + r""" + Enable memory efficient attention as implemented in xformers. + + When this option is enabled, you should observe lower GPU memory usage and a potential speed up at inference + time. Speed up at training time is not guaranteed. + + Warning: When Memory Efficient Attention and Sliced attention are both enabled, the Memory Efficient Attention + is used. + """ + self.unet.set_use_memory_efficient_attention_xformers(True) + + def disable_xformers_memory_efficient_attention(self): + r""" + Disable memory efficient attention as implemented in xformers. + """ + self.unet.set_use_memory_efficient_attention_xformers(False) + + def enable_attention_slicing(self, slice_size: Optional[Union[str, int]] = "auto"): + r""" + Enable sliced attention computation. + + When this option is enabled, the attention module will split the input tensor in slices, to compute attention + in several steps. This is useful to save some memory in exchange for a small speed decrease. + + Args: + slice_size (`str` or `int`, *optional*, defaults to `"auto"`): + When `"auto"`, halves the input to the attention heads, so attention will be computed in two steps. If + a number is provided, uses as many slices as `attention_head_dim // slice_size`. In this case, + `attention_head_dim` must be a multiple of `slice_size`. + """ + if slice_size == "auto": + # half the attention head size is usually a good trade-off between + # speed and memory + slice_size = self.unet.config.attention_head_dim // 2 + self.unet.set_attention_slice(slice_size) + + def disable_attention_slicing(self): + r""" + Disable sliced attention computation. If `enable_attention_slicing` was previously invoked, this method will go + back to computing attention in one step. + """ + # set slice_size = `None` to disable `attention slicing` + self.enable_attention_slicing(None) + + def enable_sequential_cpu_offload(self): + r""" + Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet, + text_encoder, vae and safety checker have their state dicts saved to CPU and then are moved to a + `torch.device('meta') and loaded to GPU only when their specific submodule has its `forward` method called. + """ + if is_accelerate_available(): + from accelerate import cpu_offload + else: + raise ImportError("Please install accelerate via `pip install accelerate`") + + device = torch.device("cuda") + + for cpu_offloaded_model in [self.unet, self.text_encoder, self.vae, self.safety_checker]: + if cpu_offloaded_model is not None: + cpu_offload(cpu_offloaded_model, device) + + @property + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._execution_device + def _execution_device(self): + r""" + Returns the device on which the pipeline's models will be executed. After calling + `pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module + hooks. + """ + if self.device != torch.device("meta") or not hasattr(self.unet, "_hf_hook"): + return self.device + for module in self.unet.modules(): + if ( + hasattr(module, "_hf_hook") + and hasattr(module._hf_hook, "execution_device") + and module._hf_hook.execution_device is not None + ): + return torch.device(module._hf_hook.execution_device) + return self.device + + def _encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt, + enable_safety_guidance, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list(int)`): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `List[str]`): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `guidance_scale` is less than `1`). + """ + batch_size = len(prompt) if isinstance(prompt, list) else 1 + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="max_length", return_tensors="pt").input_ids + + if not torch.equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer.batch_decode(untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = text_inputs.attention_mask.to(device) + else: + attention_mask = None + + text_embeddings = self.text_encoder( + text_input_ids.to(device), + attention_mask=attention_mask, + ) + text_embeddings = text_embeddings[0] + + # duplicate text embeddings for each generation per prompt, using mps friendly method + bs_embed, seq_len, _ = text_embeddings.shape + text_embeddings = text_embeddings.repeat(1, num_images_per_prompt, 1) + text_embeddings = text_embeddings.view(bs_embed * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance: + uncond_tokens: List[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + max_length = text_input_ids.shape[-1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = uncond_input.attention_mask.to(device) + else: + attention_mask = None + + uncond_embeddings = self.text_encoder( + uncond_input.input_ids.to(device), + attention_mask=attention_mask, + ) + uncond_embeddings = uncond_embeddings[0] + + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = uncond_embeddings.shape[1] + uncond_embeddings = uncond_embeddings.repeat(1, num_images_per_prompt, 1) + uncond_embeddings = uncond_embeddings.view(batch_size * num_images_per_prompt, seq_len, -1) + + # Encode the safety concept text + if enable_safety_guidance: + safety_concept_input = self.tokenizer( + [self._safety_text_concept], + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + safety_embeddings = self.text_encoder(safety_concept_input.input_ids.to(self.device))[0] + + # duplicate safety embeddings for each generation per prompt, using mps friendly method + seq_len = safety_embeddings.shape[1] + safety_embeddings = safety_embeddings.repeat(batch_size, num_images_per_prompt, 1) + safety_embeddings = safety_embeddings.view(batch_size * num_images_per_prompt, seq_len, -1) + + # For classifier free guidance + sld, we need to do three forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing three forward passes + text_embeddings = torch.cat([uncond_embeddings, text_embeddings, safety_embeddings]) + + else: + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + text_embeddings = torch.cat([uncond_embeddings, text_embeddings]) + + return text_embeddings + + def run_safety_checker(self, image, device, dtype, enable_safety_guidance): + if self.safety_checker is not None: + safety_checker_input = self.feature_extractor(self.numpy_to_pil(image), return_tensors="pt").to(device) + image, has_nsfw_concept = self.safety_checker( + images=image, clip_input=safety_checker_input.pixel_values.to(dtype) + ) + flagged_images = None + if any(has_nsfw_concept): + logger.warning( + "Potential NSFW content was detected in one or more images. A black image will be returned" + " instead." + f" {'You may look at this images in the `unsafe_images` variable of the output at your own discretion.' if enable_safety_guidance else 'Try again with a different prompt and/or seed.'} " + ) + flagged_images = np.zeros((2, *image.shape[1:])) + for idx, has_nsfw_concept in enumerate(has_nsfw_concept): + if has_nsfw_concept: + flagged_images[idx] = image[idx] + image[idx] = np.zeros(image[idx].shape) # black image + else: + has_nsfw_concept = None + flagged_images = None + return image, has_nsfw_concept, flagged_images + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents + def decode_latents(self, latents): + latents = 1 / 0.18215 * latents + image = self.vae.decode(latents).sample + image = (image / 2 + 0.5).clamp(0, 1) + # we always cast to float32 as this does not cause significant overhead and is compatible with bfloa16 + image = image.cpu().permute(0, 2, 3, 1).float().numpy() + return image + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.check_inputs + def check_inputs(self, prompt, height, width, callback_steps): + if not isinstance(prompt, str) and not isinstance(prompt, list): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if (callback_steps is None) or ( + callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0) + ): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = (batch_size, num_channels_latents, height // 8, width // 8) + if latents is None: + if device.type == "mps": + # randn does not work reproducibly on mps + latents = torch.randn(shape, generator=generator, device="cpu", dtype=dtype).to(device) + else: + latents = torch.randn(shape, generator=generator, device=device, dtype=dtype) + else: + if latents.shape != shape: + raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}") + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + def perform_safety_guidance( + self, + enable_safety_guidance, + safety_momentum, + noise_guidance, + noise_pred_out, + i, + sld_guidance_scale, + sld_warmup_steps, + sld_threshold, + sld_momentum_scale, + sld_mom_beta, + ): + # Perform SLD guidance + if enable_safety_guidance: + if safety_momentum is None: + safety_momentum = torch.zeros_like(noise_guidance) + noise_pred_text, noise_pred_uncond = noise_pred_out[0], noise_pred_out[1] + noise_pred_safety_concept = noise_pred_out[2] + + # Equation 6 + scale = torch.clamp(torch.abs((noise_pred_text - noise_pred_safety_concept)) * sld_guidance_scale, max=1.0) + + # Equation 6 + safety_concept_scale = torch.where( + (noise_pred_text - noise_pred_safety_concept) >= sld_threshold, torch.zeros_like(scale), scale + ) + + # Equation 4 + noise_guidance_safety = torch.mul((noise_pred_safety_concept - noise_pred_uncond), safety_concept_scale) + + # Equation 7 + noise_guidance_safety = noise_guidance_safety + sld_momentum_scale * safety_momentum + + # Equation 8 + safety_momentum = sld_mom_beta * safety_momentum + (1 - sld_mom_beta) * noise_guidance_safety + + if i >= sld_warmup_steps: # Warmup + # Equation 3 + noise_guidance = noise_guidance - noise_guidance_safety + return noise_guidance, safety_momentum + + @torch.no_grad() + def __call__( + self, + prompt: Union[str, List[str]], + height: int = 512, + width: int = 512, + num_inference_steps: int = 50, + guidance_scale: float = 7.5, + negative_prompt: Optional[Union[str, List[str]]] = None, + num_images_per_prompt: Optional[int] = 1, + eta: float = 0.0, + generator: Optional[torch.Generator] = None, + latents: Optional[torch.FloatTensor] = None, + output_type: Optional[str] = "pil", + return_dict: bool = True, + callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None, + callback_steps: Optional[int] = 1, + sld_guidance_scale: Optional[float] = 1000, + sld_warmup_steps: Optional[int] = 10, + sld_threshold: Optional[float] = 0.01, + sld_momentum_scale: Optional[float] = 0.3, + sld_mom_beta: Optional[float] = 0.4, + **kwargs, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `List[str]`): + The prompt or prompts to guide the image generation. + height (`int`, *optional*, defaults to 512): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to 512): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 7.5): + Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598). + `guidance_scale` is defined as `w` of equation 2. of [Imagen + Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale > + 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`, + usually at the expense of lower image quality. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `guidance_scale` is less than `1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to + [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`torch.Generator`, *optional*): + A [torch generator](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make generation + deterministic. + latents (`torch.FloatTensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will ge generated by sampling using the supplied random `generator`. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + callback (`Callable`, *optional*): + A function that will be called every `callback_steps` steps during inference. The function will be + called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function will be called. If not specified, the callback will be + called at every step. + sld_guidance_scale (`float`, *optional*, defaults to 1000): + Safe latent guidance as defined in [Safe Latent Diffusion](https://arxiv.org/abs/2211.05105). + `sld_guidance_scale` is defined as sS of Eq. 6. If set to be less than 1, safety guidance will be + disabled. + sld_warmup_steps (`int`, *optional*, defaults to 10): + Number of warmup steps for safety guidance. SLD will only be applied for diffusion steps greater than + `sld_warmup_steps`. `sld_warmup_steps` is defined as `delta` of [Safe Latent + Diffusion](https://arxiv.org/abs/2211.05105). + sld_threshold (`float`, *optional*, defaults to 0.01): + Threshold that separates the hyperplane between appropriate and inappropriate images. `sld_threshold` + is defined as `lamda` of Eq. 5 in [Safe Latent Diffusion](https://arxiv.org/abs/2211.05105). + sld_momentum_scale (`float`, *optional*, defaults to 0.3): + Scale of the SLD momentum to be added to the safety guidance at each diffusion step. If set to 0.0 + momentum will be disabled. Momentum is already built up during warmup, i.e. for diffusion steps smaller + than `sld_warmup_steps`. `sld_momentum_scale` is defined as `sm` of Eq. 7 in [Safe Latent + Diffusion](https://arxiv.org/abs/2211.05105). + sld_mom_beta (`float`, *optional*, defaults to 0.4): + Defines how safety guidance momentum builds up. `sld_mom_beta` indicates how much of the previous + momentum will be kept. Momentum is already built up during warmup, i.e. for diffusion steps smaller + than `sld_warmup_steps`. `sld_mom_beta` is defined as `beta m` of Eq. 8 in [Safe Latent + Diffusion](https://arxiv.org/abs/2211.05105). + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple. + When returning a tuple, the first element is a list with the generated images, and the second element is a + list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work" + (nsfw) content, according to the `safety_checker`. + """ + + # 1. Check inputs. Raise error if not correct + self.check_inputs(prompt, height, width, callback_steps) + + # 2. Define call parameters + batch_size = 1 if isinstance(prompt, str) else len(prompt) + device = self._execution_device + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = guidance_scale > 1.0 + + enable_safety_guidance = sld_guidance_scale > 1.0 and do_classifier_free_guidance + if not enable_safety_guidance: + warnings.warn("Safety checker disabled!") + + # 3. Encode input prompt + text_embeddings = self._encode_prompt( + prompt, device, num_images_per_prompt, do_classifier_free_guidance, negative_prompt, enable_safety_guidance + ) + + # 4. Prepare timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + # 5. Prepare latent variables + num_channels_latents = self.unet.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + text_embeddings.dtype, + device, + generator, + latents, + ) + + # 6. Prepare extra step kwargs. + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + safety_momentum = None + + for i, t in enumerate(self.progress_bar(timesteps)): + # expand the latents if we are doing classifier free guidance + latent_model_input = ( + torch.cat([latents] * (3 if enable_safety_guidance else 2)) if do_classifier_free_guidance else latents + ) + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # predict the noise residual + noise_pred = self.unet(latent_model_input, t, encoder_hidden_states=text_embeddings).sample + + # perform guidance + if do_classifier_free_guidance: + noise_pred_out = noise_pred.chunk((3 if enable_safety_guidance else 2)) + noise_pred_uncond, noise_pred_text = noise_pred_out[0], noise_pred_out[1] + + # default classifier free guidance + noise_guidance = noise_pred_text - noise_pred_uncond + + # Perform SLD guidance + if enable_safety_guidance: + if safety_momentum is None: + safety_momentum = torch.zeros_like(noise_guidance) + noise_pred_safety_concept = noise_pred_out[2] + + # Equation 6 + scale = torch.clamp( + torch.abs((noise_pred_text - noise_pred_safety_concept)) * sld_guidance_scale, max=1.0 + ) + + # Equation 6 + safety_concept_scale = torch.where( + (noise_pred_text - noise_pred_safety_concept) >= sld_threshold, torch.zeros_like(scale), scale + ) + + # Equation 4 + noise_guidance_safety = torch.mul( + (noise_pred_safety_concept - noise_pred_uncond), safety_concept_scale + ) + + # Equation 7 + noise_guidance_safety = noise_guidance_safety + sld_momentum_scale * safety_momentum + + # Equation 8 + safety_momentum = sld_mom_beta * safety_momentum + (1 - sld_mom_beta) * noise_guidance_safety + + if i >= sld_warmup_steps: # Warmup + # Equation 3 + noise_guidance = noise_guidance - noise_guidance_safety + + noise_pred = noise_pred_uncond + guidance_scale * noise_guidance + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample + + # call the callback, if provided + if callback is not None and i % callback_steps == 0: + callback(i, t, latents) + + # 8. Post-processing + image = self.decode_latents(latents) + + # 9. Run safety checker + image, has_nsfw_concept, flagged_images = self.run_safety_checker( + image, device, text_embeddings.dtype, enable_safety_guidance + ) + + # 10. Convert to PIL + if output_type == "pil": + image = self.numpy_to_pil(image) + if flagged_images is not None: + flagged_images = self.numpy_to_pil(flagged_images) + + if not return_dict: + return ( + image, + has_nsfw_concept, + self._safety_text_concept if enable_safety_guidance else None, + flagged_images, + ) + + return StableDiffusionSafePipelineOutput( + images=image, + nsfw_content_detected=has_nsfw_concept, + applied_safety_concept=self._safety_text_concept if enable_safety_guidance else None, + unsafe_images=flagged_images, + ) diff --git a/src/diffusers/pipelines/stable_diffusion_safe/safety_checker.py b/src/diffusers/pipelines/stable_diffusion_safe/safety_checker.py new file mode 100644 index 0000000000..f9dbf51e86 --- /dev/null +++ b/src/diffusers/pipelines/stable_diffusion_safe/safety_checker.py @@ -0,0 +1,110 @@ +# Copyright 2022 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import torch +import torch.nn as nn + +from transformers import CLIPConfig, CLIPVisionModel, PreTrainedModel + +from ...utils import logging + + +logger = logging.get_logger(__name__) + + +def cosine_distance(image_embeds, text_embeds): + normalized_image_embeds = nn.functional.normalize(image_embeds) + normalized_text_embeds = nn.functional.normalize(text_embeds) + return torch.mm(normalized_image_embeds, normalized_text_embeds.t()) + + +class SafeStableDiffusionSafetyChecker(PreTrainedModel): + config_class = CLIPConfig + + _no_split_modules = ["CLIPEncoderLayer"] + + def __init__(self, config: CLIPConfig): + super().__init__(config) + + self.vision_model = CLIPVisionModel(config.vision_config) + self.visual_projection = nn.Linear(config.vision_config.hidden_size, config.projection_dim, bias=False) + + self.concept_embeds = nn.Parameter(torch.ones(17, config.projection_dim), requires_grad=False) + self.special_care_embeds = nn.Parameter(torch.ones(3, config.projection_dim), requires_grad=False) + + self.concept_embeds_weights = nn.Parameter(torch.ones(17), requires_grad=False) + self.special_care_embeds_weights = nn.Parameter(torch.ones(3), requires_grad=False) + + @torch.no_grad() + def forward(self, clip_input, images): + pooled_output = self.vision_model(clip_input)[1] # pooled_output + image_embeds = self.visual_projection(pooled_output) + + # we always cast to float32 as this does not cause significant overhead and is compatible with bfloa16 + special_cos_dist = cosine_distance(image_embeds, self.special_care_embeds).cpu().float().numpy() + cos_dist = cosine_distance(image_embeds, self.concept_embeds).cpu().float().numpy() + + result = [] + batch_size = image_embeds.shape[0] + for i in range(batch_size): + result_img = {"special_scores": {}, "special_care": [], "concept_scores": {}, "bad_concepts": []} + + # increase this value to create a stronger `nfsw` filter + # at the cost of increasing the possibility of filtering benign images + adjustment = 0.0 + + for concept_idx in range(len(special_cos_dist[0])): + concept_cos = special_cos_dist[i][concept_idx] + concept_threshold = self.special_care_embeds_weights[concept_idx].item() + result_img["special_scores"][concept_idx] = round(concept_cos - concept_threshold + adjustment, 3) + if result_img["special_scores"][concept_idx] > 0: + result_img["special_care"].append({concept_idx, result_img["special_scores"][concept_idx]}) + adjustment = 0.01 + + for concept_idx in range(len(cos_dist[0])): + concept_cos = cos_dist[i][concept_idx] + concept_threshold = self.concept_embeds_weights[concept_idx].item() + result_img["concept_scores"][concept_idx] = round(concept_cos - concept_threshold + adjustment, 3) + if result_img["concept_scores"][concept_idx] > 0: + result_img["bad_concepts"].append(concept_idx) + + result.append(result_img) + + has_nsfw_concepts = [len(res["bad_concepts"]) > 0 for res in result] + + return images, has_nsfw_concepts + + @torch.no_grad() + def forward_onnx(self, clip_input: torch.FloatTensor, images: torch.FloatTensor): + pooled_output = self.vision_model(clip_input)[1] # pooled_output + image_embeds = self.visual_projection(pooled_output) + + special_cos_dist = cosine_distance(image_embeds, self.special_care_embeds) + cos_dist = cosine_distance(image_embeds, self.concept_embeds) + + # increase this value to create a stronger `nsfw` filter + # at the cost of increasing the possibility of filtering benign images + adjustment = 0.0 + + special_scores = special_cos_dist - self.special_care_embeds_weights + adjustment + # special_scores = special_scores.round(decimals=3) + special_care = torch.any(special_scores > 0, dim=1) + special_adjustment = special_care * 0.01 + special_adjustment = special_adjustment.unsqueeze(1).expand(-1, cos_dist.shape[1]) + + concept_scores = (cos_dist - self.concept_embeds_weights) + special_adjustment + # concept_scores = concept_scores.round(decimals=3) + has_nsfw_concepts = torch.any(concept_scores > 0, dim=1) + + return images, has_nsfw_concepts diff --git a/src/diffusers/pipelines/versatile_diffusion/__init__.py b/src/diffusers/pipelines/versatile_diffusion/__init__.py new file mode 100644 index 0000000000..1d2caa7e23 --- /dev/null +++ b/src/diffusers/pipelines/versatile_diffusion/__init__.py @@ -0,0 +1,16 @@ +from ...utils import is_torch_available, is_transformers_available, is_transformers_version + + +if is_transformers_available() and is_torch_available() and is_transformers_version(">=", "4.25.0.dev0"): + from .modeling_text_unet import UNetFlatConditionModel + from .pipeline_versatile_diffusion import VersatileDiffusionPipeline + from .pipeline_versatile_diffusion_dual_guided import VersatileDiffusionDualGuidedPipeline + from .pipeline_versatile_diffusion_image_variation import VersatileDiffusionImageVariationPipeline + from .pipeline_versatile_diffusion_text_to_image import VersatileDiffusionTextToImagePipeline +else: + from ...utils.dummy_torch_and_transformers_objects import ( + VersatileDiffusionDualGuidedPipeline, + VersatileDiffusionImageVariationPipeline, + VersatileDiffusionPipeline, + VersatileDiffusionTextToImagePipeline, + ) diff --git a/src/diffusers/pipelines/versatile_diffusion/modeling_text_unet.py b/src/diffusers/pipelines/versatile_diffusion/modeling_text_unet.py new file mode 100644 index 0000000000..6d521228e3 --- /dev/null +++ b/src/diffusers/pipelines/versatile_diffusion/modeling_text_unet.py @@ -0,0 +1,1095 @@ +from typing import Optional, Tuple, Union + +import numpy as np +import torch +import torch.nn as nn + +from ...configuration_utils import ConfigMixin, register_to_config +from ...modeling_utils import ModelMixin +from ...models.attention import DualTransformer2DModel, Transformer2DModel +from ...models.embeddings import TimestepEmbedding, Timesteps +from ...models.unet_2d_condition import UNet2DConditionOutput +from ...utils import logging + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +def get_down_block( + down_block_type, + num_layers, + in_channels, + out_channels, + temb_channels, + add_downsample, + resnet_eps, + resnet_act_fn, + attn_num_head_channels, + resnet_groups=None, + cross_attention_dim=None, + downsample_padding=None, + dual_cross_attention=None, +): + down_block_type = down_block_type[7:] if down_block_type.startswith("UNetRes") else down_block_type + if down_block_type == "DownBlockFlat": + return DownBlockFlat( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + add_downsample=add_downsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + downsample_padding=downsample_padding, + ) + elif down_block_type == "CrossAttnDownBlockFlat": + if cross_attention_dim is None: + raise ValueError("cross_attention_dim must be specified for CrossAttnDownBlockFlat") + return CrossAttnDownBlockFlat( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + add_downsample=add_downsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + downsample_padding=downsample_padding, + cross_attention_dim=cross_attention_dim, + attn_num_head_channels=attn_num_head_channels, + ) + raise ValueError(f"{down_block_type} is not supported.") + + +def get_up_block( + up_block_type, + num_layers, + in_channels, + out_channels, + prev_output_channel, + temb_channels, + add_upsample, + resnet_eps, + resnet_act_fn, + attn_num_head_channels, + resnet_groups=None, + cross_attention_dim=None, + dual_cross_attention=None, +): + up_block_type = up_block_type[7:] if up_block_type.startswith("UNetRes") else up_block_type + if up_block_type == "UpBlockFlat": + return UpBlockFlat( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + prev_output_channel=prev_output_channel, + temb_channels=temb_channels, + add_upsample=add_upsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + ) + elif up_block_type == "CrossAttnUpBlockFlat": + if cross_attention_dim is None: + raise ValueError("cross_attention_dim must be specified for CrossAttnUpBlockFlat") + return CrossAttnUpBlockFlat( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + prev_output_channel=prev_output_channel, + temb_channels=temb_channels, + add_upsample=add_upsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + cross_attention_dim=cross_attention_dim, + attn_num_head_channels=attn_num_head_channels, + ) + raise ValueError(f"{up_block_type} is not supported.") + + +# Copied from diffusers.models.unet_2d_condition.UNet2DConditionModel with UNet2DConditionModel->UNetFlatConditionModel, nn.Conv2d->LinearMultiDim, Block2D->BlockFlat +class UNetFlatConditionModel(ModelMixin, ConfigMixin): + r""" + UNetFlatConditionModel is a conditional 2D UNet model that takes in a noisy sample, conditional state, and a + timestep and returns sample shaped output. + + This model inherits from [`ModelMixin`]. Check the superclass documentation for the generic methods the library + implements for all the models (such as downloading or saving, etc.) + + Parameters: + sample_size (`int` or `Tuple[int, int]`, *optional*, defaults to `None`): + Height and width of input/output sample. + in_channels (`int`, *optional*, defaults to 4): The number of channels in the input sample. + out_channels (`int`, *optional*, defaults to 4): The number of channels in the output. + center_input_sample (`bool`, *optional*, defaults to `False`): Whether to center the input sample. + flip_sin_to_cos (`bool`, *optional*, defaults to `False`): + Whether to flip the sin to cos in the time embedding. + freq_shift (`int`, *optional*, defaults to 0): The frequency shift to apply to the time embedding. + down_block_types (`Tuple[str]`, *optional*, defaults to `("CrossAttnDownBlockFlat", "CrossAttnDownBlockFlat", "CrossAttnDownBlockFlat", "DownBlockFlat")`): + The tuple of downsample blocks to use. + up_block_types (`Tuple[str]`, *optional*, defaults to `("UpBlockFlat", "CrossAttnUpBlockFlat", "CrossAttnUpBlockFlat", "CrossAttnUpBlockFlat",)`): + The tuple of upsample blocks to use. + block_out_channels (`Tuple[int]`, *optional*, defaults to `(320, 640, 1280, 1280)`): + The tuple of output channels for each block. + layers_per_block (`int`, *optional*, defaults to 2): The number of layers per block. + downsample_padding (`int`, *optional*, defaults to 1): The padding to use for the downsampling convolution. + mid_block_scale_factor (`float`, *optional*, defaults to 1.0): The scale factor to use for the mid block. + act_fn (`str`, *optional*, defaults to `"silu"`): The activation function to use. + norm_num_groups (`int`, *optional*, defaults to 32): The number of groups to use for the normalization. + norm_eps (`float`, *optional*, defaults to 1e-5): The epsilon to use for the normalization. + cross_attention_dim (`int`, *optional*, defaults to 1280): The dimension of the cross attention features. + attention_head_dim (`int`, *optional*, defaults to 8): The dimension of the attention heads. + """ + + _supports_gradient_checkpointing = True + + @register_to_config + def __init__( + self, + sample_size: Optional[int] = None, + in_channels: int = 4, + out_channels: int = 4, + center_input_sample: bool = False, + flip_sin_to_cos: bool = True, + freq_shift: int = 0, + down_block_types: Tuple[str] = ( + "CrossAttnDownBlockFlat", + "CrossAttnDownBlockFlat", + "CrossAttnDownBlockFlat", + "DownBlockFlat", + ), + up_block_types: Tuple[str] = ( + "UpBlockFlat", + "CrossAttnUpBlockFlat", + "CrossAttnUpBlockFlat", + "CrossAttnUpBlockFlat", + ), + block_out_channels: Tuple[int] = (320, 640, 1280, 1280), + layers_per_block: int = 2, + downsample_padding: int = 1, + mid_block_scale_factor: float = 1, + act_fn: str = "silu", + norm_num_groups: int = 32, + norm_eps: float = 1e-5, + cross_attention_dim: int = 1280, + attention_head_dim: Union[int, Tuple[int]] = 8, + dual_cross_attention: bool = False, + use_linear_projection: bool = False, + ): + super().__init__() + + self.sample_size = sample_size + time_embed_dim = block_out_channels[0] * 4 + + # input + self.conv_in = LinearMultiDim(in_channels, block_out_channels[0], kernel_size=3, padding=(1, 1)) + + # time + self.time_proj = Timesteps(block_out_channels[0], flip_sin_to_cos, freq_shift) + timestep_input_dim = block_out_channels[0] + + self.time_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim) + + self.down_blocks = nn.ModuleList([]) + self.mid_block = None + self.up_blocks = nn.ModuleList([]) + + if isinstance(attention_head_dim, int): + attention_head_dim = (attention_head_dim,) * len(down_block_types) + + # down + output_channel = block_out_channels[0] + for i, down_block_type in enumerate(down_block_types): + input_channel = output_channel + output_channel = block_out_channels[i] + is_final_block = i == len(block_out_channels) - 1 + + down_block = get_down_block( + down_block_type, + num_layers=layers_per_block, + in_channels=input_channel, + out_channels=output_channel, + temb_channels=time_embed_dim, + add_downsample=not is_final_block, + resnet_eps=norm_eps, + resnet_act_fn=act_fn, + resnet_groups=norm_num_groups, + cross_attention_dim=cross_attention_dim, + attn_num_head_channels=attention_head_dim[i], + downsample_padding=downsample_padding, + dual_cross_attention=dual_cross_attention, + use_linear_projection=use_linear_projection, + ) + self.down_blocks.append(down_block) + + # mid + self.mid_block = UNetMidBlockFlatCrossAttn( + in_channels=block_out_channels[-1], + temb_channels=time_embed_dim, + resnet_eps=norm_eps, + resnet_act_fn=act_fn, + output_scale_factor=mid_block_scale_factor, + resnet_time_scale_shift="default", + cross_attention_dim=cross_attention_dim, + attn_num_head_channels=attention_head_dim[-1], + resnet_groups=norm_num_groups, + dual_cross_attention=dual_cross_attention, + use_linear_projection=use_linear_projection, + ) + + # count how many layers upsample the images + self.num_upsamplers = 0 + + # up + reversed_block_out_channels = list(reversed(block_out_channels)) + reversed_attention_head_dim = list(reversed(attention_head_dim)) + output_channel = reversed_block_out_channels[0] + for i, up_block_type in enumerate(up_block_types): + is_final_block = i == len(block_out_channels) - 1 + + prev_output_channel = output_channel + output_channel = reversed_block_out_channels[i] + input_channel = reversed_block_out_channels[min(i + 1, len(block_out_channels) - 1)] + + # add upsample block for all BUT final layer + if not is_final_block: + add_upsample = True + self.num_upsamplers += 1 + else: + add_upsample = False + + up_block = get_up_block( + up_block_type, + num_layers=layers_per_block + 1, + in_channels=input_channel, + out_channels=output_channel, + prev_output_channel=prev_output_channel, + temb_channels=time_embed_dim, + add_upsample=add_upsample, + resnet_eps=norm_eps, + resnet_act_fn=act_fn, + resnet_groups=norm_num_groups, + cross_attention_dim=cross_attention_dim, + attn_num_head_channels=reversed_attention_head_dim[i], + dual_cross_attention=dual_cross_attention, + use_linear_projection=use_linear_projection, + ) + self.up_blocks.append(up_block) + prev_output_channel = output_channel + + # out + self.conv_norm_out = nn.GroupNorm(num_channels=block_out_channels[0], num_groups=norm_num_groups, eps=norm_eps) + self.conv_act = nn.SiLU() + self.conv_out = LinearMultiDim(block_out_channels[0], out_channels, kernel_size=3, padding=1) + + def set_attention_slice(self, slice_size): + if slice_size is not None and self.config.attention_head_dim % slice_size != 0: + raise ValueError( + f"Make sure slice_size {slice_size} is a divisor of " + f"the number of heads used in cross_attention {self.config.attention_head_dim}" + ) + if slice_size is not None and slice_size > self.config.attention_head_dim: + raise ValueError( + f"Chunk_size {slice_size} has to be smaller or equal to " + f"the number of heads used in cross_attention {self.config.attention_head_dim}" + ) + + for block in self.down_blocks: + if hasattr(block, "attentions") and block.attentions is not None: + block.set_attention_slice(slice_size) + + self.mid_block.set_attention_slice(slice_size) + + for block in self.up_blocks: + if hasattr(block, "attentions") and block.attentions is not None: + block.set_attention_slice(slice_size) + + def set_use_memory_efficient_attention_xformers(self, use_memory_efficient_attention_xformers: bool): + for block in self.down_blocks: + if hasattr(block, "attentions") and block.attentions is not None: + block.set_use_memory_efficient_attention_xformers(use_memory_efficient_attention_xformers) + + self.mid_block.set_use_memory_efficient_attention_xformers(use_memory_efficient_attention_xformers) + + for block in self.up_blocks: + if hasattr(block, "attentions") and block.attentions is not None: + block.set_use_memory_efficient_attention_xformers(use_memory_efficient_attention_xformers) + + def _set_gradient_checkpointing(self, module, value=False): + if isinstance(module, (CrossAttnDownBlockFlat, DownBlockFlat, CrossAttnUpBlockFlat, UpBlockFlat)): + module.gradient_checkpointing = value + + def forward( + self, + sample: torch.FloatTensor, + timestep: Union[torch.Tensor, float, int], + encoder_hidden_states: torch.Tensor, + return_dict: bool = True, + ) -> Union[UNet2DConditionOutput, Tuple]: + r""" + Args: + sample (`torch.FloatTensor`): (batch, channel, height, width) noisy inputs tensor + timestep (`torch.FloatTensor` or `float` or `int`): (batch) timesteps + encoder_hidden_states (`torch.FloatTensor`): (batch, channel, height, width) encoder hidden states + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`models.unet_2d_condition.UNet2DConditionOutput`] instead of a plain tuple. + + Returns: + [`~models.unet_2d_condition.UNet2DConditionOutput`] or `tuple`: + [`~models.unet_2d_condition.UNet2DConditionOutput`] if `return_dict` is True, otherwise a `tuple`. When + returning a tuple, the first element is the sample tensor. + """ + # By default samples have to be AT least a multiple of the overall upsampling factor. + # The overall upsampling factor is equal to 2 ** (# num of upsampling layears). + # However, the upsampling interpolation output size can be forced to fit any upsampling size + # on the fly if necessary. + default_overall_up_factor = 2**self.num_upsamplers + + # upsample size should be forwarded when sample is not a multiple of `default_overall_up_factor` + forward_upsample_size = False + upsample_size = None + + if any(s % default_overall_up_factor != 0 for s in sample.shape[-2:]): + logger.info("Forward upsample size to force interpolation output size.") + forward_upsample_size = True + + # 0. center input if necessary + if self.config.center_input_sample: + sample = 2 * sample - 1.0 + + # 1. time + timesteps = timestep + if not torch.is_tensor(timesteps): + # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can + timesteps = torch.tensor([timesteps], dtype=torch.long, device=sample.device) + elif torch.is_tensor(timesteps) and len(timesteps.shape) == 0: + timesteps = timesteps[None].to(sample.device) + + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timesteps = timesteps.expand(sample.shape[0]) + + t_emb = self.time_proj(timesteps) + + # timesteps does not contain any weights and will always return f32 tensors + # but time_embedding might actually be running in fp16. so we need to cast here. + # there might be better ways to encapsulate this. + t_emb = t_emb.to(dtype=self.dtype) + emb = self.time_embedding(t_emb) + + # 2. pre-process + sample = self.conv_in(sample) + + # 3. down + down_block_res_samples = (sample,) + for downsample_block in self.down_blocks: + if hasattr(downsample_block, "attentions") and downsample_block.attentions is not None: + sample, res_samples = downsample_block( + hidden_states=sample, + temb=emb, + encoder_hidden_states=encoder_hidden_states, + ) + else: + sample, res_samples = downsample_block(hidden_states=sample, temb=emb) + + down_block_res_samples += res_samples + + # 4. mid + sample = self.mid_block(sample, emb, encoder_hidden_states=encoder_hidden_states) + + # 5. up + for i, upsample_block in enumerate(self.up_blocks): + is_final_block = i == len(self.up_blocks) - 1 + + res_samples = down_block_res_samples[-len(upsample_block.resnets) :] + down_block_res_samples = down_block_res_samples[: -len(upsample_block.resnets)] + + # if we have not reached the final block and need to forward the + # upsample size, we do it here + if not is_final_block and forward_upsample_size: + upsample_size = down_block_res_samples[-1].shape[2:] + + if hasattr(upsample_block, "attentions") and upsample_block.attentions is not None: + sample = upsample_block( + hidden_states=sample, + temb=emb, + res_hidden_states_tuple=res_samples, + encoder_hidden_states=encoder_hidden_states, + upsample_size=upsample_size, + ) + else: + sample = upsample_block( + hidden_states=sample, temb=emb, res_hidden_states_tuple=res_samples, upsample_size=upsample_size + ) + # 6. post-process + sample = self.conv_norm_out(sample) + sample = self.conv_act(sample) + sample = self.conv_out(sample) + + if not return_dict: + return (sample,) + + return UNet2DConditionOutput(sample=sample) + + +class LinearMultiDim(nn.Linear): + def __init__(self, in_features, out_features=None, second_dim=4, *args, **kwargs): + in_features = [in_features, second_dim, 1] if isinstance(in_features, int) else list(in_features) + if out_features is None: + out_features = in_features + out_features = [out_features, second_dim, 1] if isinstance(out_features, int) else list(out_features) + self.in_features_multidim = in_features + self.out_features_multidim = out_features + super().__init__(np.array(in_features).prod(), np.array(out_features).prod()) + + def forward(self, input_tensor, *args, **kwargs): + shape = input_tensor.shape + n_dim = len(self.in_features_multidim) + input_tensor = input_tensor.reshape(*shape[0:-n_dim], self.in_features) + output_tensor = super().forward(input_tensor) + output_tensor = output_tensor.view(*shape[0:-n_dim], *self.out_features_multidim) + return output_tensor + + +class ResnetBlockFlat(nn.Module): + def __init__( + self, + *, + in_channels, + out_channels=None, + dropout=0.0, + temb_channels=512, + groups=32, + groups_out=None, + pre_norm=True, + eps=1e-6, + time_embedding_norm="default", + use_in_shortcut=None, + second_dim=4, + **kwargs, + ): + super().__init__() + self.pre_norm = pre_norm + self.pre_norm = True + + in_channels = [in_channels, second_dim, 1] if isinstance(in_channels, int) else list(in_channels) + self.in_channels_prod = np.array(in_channels).prod() + self.channels_multidim = in_channels + + if out_channels is not None: + out_channels = [out_channels, second_dim, 1] if isinstance(out_channels, int) else list(out_channels) + out_channels_prod = np.array(out_channels).prod() + self.out_channels_multidim = out_channels + else: + out_channels_prod = self.in_channels_prod + self.out_channels_multidim = self.channels_multidim + self.time_embedding_norm = time_embedding_norm + + if groups_out is None: + groups_out = groups + + self.norm1 = torch.nn.GroupNorm(num_groups=groups, num_channels=self.in_channels_prod, eps=eps, affine=True) + self.conv1 = torch.nn.Conv2d(self.in_channels_prod, out_channels_prod, kernel_size=1, padding=0) + + if temb_channels is not None: + self.time_emb_proj = torch.nn.Linear(temb_channels, out_channels_prod) + else: + self.time_emb_proj = None + + self.norm2 = torch.nn.GroupNorm(num_groups=groups_out, num_channels=out_channels_prod, eps=eps, affine=True) + self.dropout = torch.nn.Dropout(dropout) + self.conv2 = torch.nn.Conv2d(out_channels_prod, out_channels_prod, kernel_size=1, padding=0) + + self.nonlinearity = nn.SiLU() + + self.use_in_shortcut = ( + self.in_channels_prod != out_channels_prod if use_in_shortcut is None else use_in_shortcut + ) + + self.conv_shortcut = None + if self.use_in_shortcut: + self.conv_shortcut = torch.nn.Conv2d( + self.in_channels_prod, out_channels_prod, kernel_size=1, stride=1, padding=0 + ) + + def forward(self, input_tensor, temb): + shape = input_tensor.shape + n_dim = len(self.channels_multidim) + input_tensor = input_tensor.reshape(*shape[0:-n_dim], self.in_channels_prod, 1, 1) + input_tensor = input_tensor.view(-1, self.in_channels_prod, 1, 1) + + hidden_states = input_tensor + + hidden_states = self.norm1(hidden_states) + hidden_states = self.nonlinearity(hidden_states) + hidden_states = self.conv1(hidden_states) + + if temb is not None: + temb = self.time_emb_proj(self.nonlinearity(temb))[:, :, None, None] + hidden_states = hidden_states + temb + + hidden_states = self.norm2(hidden_states) + hidden_states = self.nonlinearity(hidden_states) + + hidden_states = self.dropout(hidden_states) + hidden_states = self.conv2(hidden_states) + + if self.conv_shortcut is not None: + input_tensor = self.conv_shortcut(input_tensor) + + output_tensor = input_tensor + hidden_states + + output_tensor = output_tensor.view(*shape[0:-n_dim], -1) + output_tensor = output_tensor.view(*shape[0:-n_dim], *self.out_channels_multidim) + + return output_tensor + + +# Copied from diffusers.models.unet_2d_blocks.DownBlock2D with DownBlock2D->DownBlockFlat, ResnetBlock2D->ResnetBlockFlat, Downsample2D->LinearMultiDim +class DownBlockFlat(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + temb_channels: int, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + output_scale_factor=1.0, + add_downsample=True, + downsample_padding=1, + ): + super().__init__() + resnets = [] + + for i in range(num_layers): + in_channels = in_channels if i == 0 else out_channels + resnets.append( + ResnetBlockFlat( + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + + self.resnets = nn.ModuleList(resnets) + + if add_downsample: + self.downsamplers = nn.ModuleList( + [ + LinearMultiDim( + out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op" + ) + ] + ) + else: + self.downsamplers = None + + self.gradient_checkpointing = False + + def forward(self, hidden_states, temb=None): + output_states = () + + for resnet in self.resnets: + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module): + def custom_forward(*inputs): + return module(*inputs) + + return custom_forward + + hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(resnet), hidden_states, temb) + else: + hidden_states = resnet(hidden_states, temb) + + output_states += (hidden_states,) + + if self.downsamplers is not None: + for downsampler in self.downsamplers: + hidden_states = downsampler(hidden_states) + + output_states += (hidden_states,) + + return hidden_states, output_states + + +# Copied from diffusers.models.unet_2d_blocks.CrossAttnDownBlock2D with CrossAttnDownBlock2D->CrossAttnDownBlockFlat, ResnetBlock2D->ResnetBlockFlat, Downsample2D->LinearMultiDim +class CrossAttnDownBlockFlat(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + temb_channels: int, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + attn_num_head_channels=1, + cross_attention_dim=1280, + attention_type="default", + output_scale_factor=1.0, + downsample_padding=1, + add_downsample=True, + dual_cross_attention=False, + use_linear_projection=False, + ): + super().__init__() + resnets = [] + attentions = [] + + self.attention_type = attention_type + self.attn_num_head_channels = attn_num_head_channels + + for i in range(num_layers): + in_channels = in_channels if i == 0 else out_channels + resnets.append( + ResnetBlockFlat( + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + if not dual_cross_attention: + attentions.append( + Transformer2DModel( + attn_num_head_channels, + out_channels // attn_num_head_channels, + in_channels=out_channels, + num_layers=1, + cross_attention_dim=cross_attention_dim, + norm_num_groups=resnet_groups, + use_linear_projection=use_linear_projection, + ) + ) + else: + attentions.append( + DualTransformer2DModel( + attn_num_head_channels, + out_channels // attn_num_head_channels, + in_channels=out_channels, + num_layers=1, + cross_attention_dim=cross_attention_dim, + norm_num_groups=resnet_groups, + ) + ) + self.attentions = nn.ModuleList(attentions) + self.resnets = nn.ModuleList(resnets) + + if add_downsample: + self.downsamplers = nn.ModuleList( + [ + LinearMultiDim( + out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op" + ) + ] + ) + else: + self.downsamplers = None + + self.gradient_checkpointing = False + + def set_attention_slice(self, slice_size): + if slice_size is not None and self.attn_num_head_channels % slice_size != 0: + raise ValueError( + f"Make sure slice_size {slice_size} is a divisor of " + f"the number of heads used in cross_attention {self.attn_num_head_channels}" + ) + if slice_size is not None and slice_size > self.attn_num_head_channels: + raise ValueError( + f"Chunk_size {slice_size} has to be smaller or equal to " + f"the number of heads used in cross_attention {self.attn_num_head_channels}" + ) + + for attn in self.attentions: + attn._set_attention_slice(slice_size) + + def set_use_memory_efficient_attention_xformers(self, use_memory_efficient_attention_xformers: bool): + for attn in self.attentions: + attn._set_use_memory_efficient_attention_xformers(use_memory_efficient_attention_xformers) + + def forward(self, hidden_states, temb=None, encoder_hidden_states=None): + output_states = () + + for resnet, attn in zip(self.resnets, self.attentions): + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module, return_dict=None): + def custom_forward(*inputs): + if return_dict is not None: + return module(*inputs, return_dict=return_dict) + else: + return module(*inputs) + + return custom_forward + + hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(resnet), hidden_states, temb) + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(attn, return_dict=False), hidden_states, encoder_hidden_states + )[0] + else: + hidden_states = resnet(hidden_states, temb) + hidden_states = attn(hidden_states, encoder_hidden_states=encoder_hidden_states).sample + + output_states += (hidden_states,) + + if self.downsamplers is not None: + for downsampler in self.downsamplers: + hidden_states = downsampler(hidden_states) + + output_states += (hidden_states,) + + return hidden_states, output_states + + +# Copied from diffusers.models.unet_2d_blocks.UpBlock2D with UpBlock2D->UpBlockFlat, ResnetBlock2D->ResnetBlockFlat, Upsample2D->LinearMultiDim +class UpBlockFlat(nn.Module): + def __init__( + self, + in_channels: int, + prev_output_channel: int, + out_channels: int, + temb_channels: int, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + output_scale_factor=1.0, + add_upsample=True, + ): + super().__init__() + resnets = [] + + for i in range(num_layers): + res_skip_channels = in_channels if (i == num_layers - 1) else out_channels + resnet_in_channels = prev_output_channel if i == 0 else out_channels + + resnets.append( + ResnetBlockFlat( + in_channels=resnet_in_channels + res_skip_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + + self.resnets = nn.ModuleList(resnets) + + if add_upsample: + self.upsamplers = nn.ModuleList([LinearMultiDim(out_channels, use_conv=True, out_channels=out_channels)]) + else: + self.upsamplers = None + + self.gradient_checkpointing = False + + def forward(self, hidden_states, res_hidden_states_tuple, temb=None, upsample_size=None): + for resnet in self.resnets: + # pop res hidden states + res_hidden_states = res_hidden_states_tuple[-1] + res_hidden_states_tuple = res_hidden_states_tuple[:-1] + hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1) + + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module): + def custom_forward(*inputs): + return module(*inputs) + + return custom_forward + + hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(resnet), hidden_states, temb) + else: + hidden_states = resnet(hidden_states, temb) + + if self.upsamplers is not None: + for upsampler in self.upsamplers: + hidden_states = upsampler(hidden_states, upsample_size) + + return hidden_states + + +# Copied from diffusers.models.unet_2d_blocks.CrossAttnUpBlock2D with CrossAttnUpBlock2D->CrossAttnUpBlockFlat, ResnetBlock2D->ResnetBlockFlat, Upsample2D->LinearMultiDim +class CrossAttnUpBlockFlat(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + prev_output_channel: int, + temb_channels: int, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + attn_num_head_channels=1, + cross_attention_dim=1280, + attention_type="default", + output_scale_factor=1.0, + add_upsample=True, + dual_cross_attention=False, + use_linear_projection=False, + ): + super().__init__() + resnets = [] + attentions = [] + + self.attention_type = attention_type + self.attn_num_head_channels = attn_num_head_channels + + for i in range(num_layers): + res_skip_channels = in_channels if (i == num_layers - 1) else out_channels + resnet_in_channels = prev_output_channel if i == 0 else out_channels + + resnets.append( + ResnetBlockFlat( + in_channels=resnet_in_channels + res_skip_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + if not dual_cross_attention: + attentions.append( + Transformer2DModel( + attn_num_head_channels, + out_channels // attn_num_head_channels, + in_channels=out_channels, + num_layers=1, + cross_attention_dim=cross_attention_dim, + norm_num_groups=resnet_groups, + use_linear_projection=use_linear_projection, + ) + ) + else: + attentions.append( + DualTransformer2DModel( + attn_num_head_channels, + out_channels // attn_num_head_channels, + in_channels=out_channels, + num_layers=1, + cross_attention_dim=cross_attention_dim, + norm_num_groups=resnet_groups, + ) + ) + self.attentions = nn.ModuleList(attentions) + self.resnets = nn.ModuleList(resnets) + + if add_upsample: + self.upsamplers = nn.ModuleList([LinearMultiDim(out_channels, use_conv=True, out_channels=out_channels)]) + else: + self.upsamplers = None + + self.gradient_checkpointing = False + + def set_attention_slice(self, slice_size): + if slice_size is not None and self.attn_num_head_channels % slice_size != 0: + raise ValueError( + f"Make sure slice_size {slice_size} is a divisor of " + f"the number of heads used in cross_attention {self.attn_num_head_channels}" + ) + if slice_size is not None and slice_size > self.attn_num_head_channels: + raise ValueError( + f"Chunk_size {slice_size} has to be smaller or equal to " + f"the number of heads used in cross_attention {self.attn_num_head_channels}" + ) + + for attn in self.attentions: + attn._set_attention_slice(slice_size) + + self.gradient_checkpointing = False + + def set_use_memory_efficient_attention_xformers(self, use_memory_efficient_attention_xformers: bool): + for attn in self.attentions: + attn._set_use_memory_efficient_attention_xformers(use_memory_efficient_attention_xformers) + + def forward( + self, + hidden_states, + res_hidden_states_tuple, + temb=None, + encoder_hidden_states=None, + upsample_size=None, + ): + for resnet, attn in zip(self.resnets, self.attentions): + # pop res hidden states + res_hidden_states = res_hidden_states_tuple[-1] + res_hidden_states_tuple = res_hidden_states_tuple[:-1] + hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1) + + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module, return_dict=None): + def custom_forward(*inputs): + if return_dict is not None: + return module(*inputs, return_dict=return_dict) + else: + return module(*inputs) + + return custom_forward + + hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(resnet), hidden_states, temb) + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(attn, return_dict=False), hidden_states, encoder_hidden_states + )[0] + else: + hidden_states = resnet(hidden_states, temb) + hidden_states = attn(hidden_states, encoder_hidden_states=encoder_hidden_states).sample + + if self.upsamplers is not None: + for upsampler in self.upsamplers: + hidden_states = upsampler(hidden_states, upsample_size) + + return hidden_states + + +# Copied from diffusers.models.unet_2d_blocks.UNetMidBlock2DCrossAttn with UNetMidBlock2DCrossAttn->UNetMidBlockFlatCrossAttn, ResnetBlock2D->ResnetBlockFlat +class UNetMidBlockFlatCrossAttn(nn.Module): + def __init__( + self, + in_channels: int, + temb_channels: int, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + attn_num_head_channels=1, + attention_type="default", + output_scale_factor=1.0, + cross_attention_dim=1280, + dual_cross_attention=False, + use_linear_projection=False, + **kwargs, + ): + super().__init__() + + self.attention_type = attention_type + self.attn_num_head_channels = attn_num_head_channels + resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32) + + # there is always at least one resnet + resnets = [ + ResnetBlockFlat( + in_channels=in_channels, + out_channels=in_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ] + attentions = [] + + for _ in range(num_layers): + if not dual_cross_attention: + attentions.append( + Transformer2DModel( + attn_num_head_channels, + in_channels // attn_num_head_channels, + in_channels=in_channels, + num_layers=1, + cross_attention_dim=cross_attention_dim, + norm_num_groups=resnet_groups, + use_linear_projection=use_linear_projection, + ) + ) + else: + attentions.append( + DualTransformer2DModel( + attn_num_head_channels, + in_channels // attn_num_head_channels, + in_channels=in_channels, + num_layers=1, + cross_attention_dim=cross_attention_dim, + norm_num_groups=resnet_groups, + ) + ) + resnets.append( + ResnetBlockFlat( + in_channels=in_channels, + out_channels=in_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + + self.attentions = nn.ModuleList(attentions) + self.resnets = nn.ModuleList(resnets) + + def set_attention_slice(self, slice_size): + if slice_size is not None and self.attn_num_head_channels % slice_size != 0: + raise ValueError( + f"Make sure slice_size {slice_size} is a divisor of " + f"the number of heads used in cross_attention {self.attn_num_head_channels}" + ) + if slice_size is not None and slice_size > self.attn_num_head_channels: + raise ValueError( + f"Chunk_size {slice_size} has to be smaller or equal to " + f"the number of heads used in cross_attention {self.attn_num_head_channels}" + ) + + for attn in self.attentions: + attn._set_attention_slice(slice_size) + + def set_use_memory_efficient_attention_xformers(self, use_memory_efficient_attention_xformers: bool): + for attn in self.attentions: + attn._set_use_memory_efficient_attention_xformers(use_memory_efficient_attention_xformers) + + def forward(self, hidden_states, temb=None, encoder_hidden_states=None): + hidden_states = self.resnets[0](hidden_states, temb) + for attn, resnet in zip(self.attentions, self.resnets[1:]): + hidden_states = attn(hidden_states, encoder_hidden_states).sample + hidden_states = resnet(hidden_states, temb) + + return hidden_states diff --git a/src/diffusers/pipelines/versatile_diffusion/pipeline_versatile_diffusion.py b/src/diffusers/pipelines/versatile_diffusion/pipeline_versatile_diffusion.py new file mode 100644 index 0000000000..1280419c34 --- /dev/null +++ b/src/diffusers/pipelines/versatile_diffusion/pipeline_versatile_diffusion.py @@ -0,0 +1,462 @@ +import inspect +from typing import Callable, List, Optional, Union + +import torch + +import PIL.Image +from transformers import CLIPFeatureExtractor, CLIPTextModel, CLIPTokenizer, CLIPVisionModel + +from ...models import AutoencoderKL, UNet2DConditionModel +from ...pipeline_utils import DiffusionPipeline +from ...schedulers import DDIMScheduler, LMSDiscreteScheduler, PNDMScheduler +from ...utils import logging +from .pipeline_versatile_diffusion_dual_guided import VersatileDiffusionDualGuidedPipeline +from .pipeline_versatile_diffusion_image_variation import VersatileDiffusionImageVariationPipeline +from .pipeline_versatile_diffusion_text_to_image import VersatileDiffusionTextToImagePipeline + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +class VersatileDiffusionPipeline(DiffusionPipeline): + r""" + Pipeline for text-to-image generation using Stable Diffusion. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModel`]): + Frozen text-encoder. Stable Diffusion uses the text portion of + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically + the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + safety_checker ([`StableDiffusionMegaSafetyChecker`]): + Classification module that estimates whether generated images could be considered offensive or harmful. + Please, refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for details. + feature_extractor ([`CLIPFeatureExtractor`]): + Model that extracts features from generated images to be used as inputs for the `safety_checker`. + """ + + tokenizer: CLIPTokenizer + image_feature_extractor: CLIPFeatureExtractor + text_encoder: CLIPTextModel + image_encoder: CLIPVisionModel + image_unet: UNet2DConditionModel + text_unet: UNet2DConditionModel + vae: AutoencoderKL + scheduler: Union[DDIMScheduler, PNDMScheduler, LMSDiscreteScheduler] + + def __init__( + self, + tokenizer: CLIPTokenizer, + image_feature_extractor: CLIPFeatureExtractor, + text_encoder: CLIPTextModel, + image_encoder: CLIPVisionModel, + image_unet: UNet2DConditionModel, + text_unet: UNet2DConditionModel, + vae: AutoencoderKL, + scheduler: Union[DDIMScheduler, PNDMScheduler, LMSDiscreteScheduler], + ): + super().__init__() + + self.register_modules( + tokenizer=tokenizer, + image_feature_extractor=image_feature_extractor, + text_encoder=text_encoder, + image_encoder=image_encoder, + image_unet=image_unet, + text_unet=text_unet, + vae=vae, + scheduler=scheduler, + ) + + def enable_attention_slicing(self, slice_size: Optional[Union[str, int]] = "auto"): + r""" + Enable sliced attention computation. + + When this option is enabled, the attention module will split the input tensor in slices, to compute attention + in several steps. This is useful to save some memory in exchange for a small speed decrease. + + Args: + slice_size (`str` or `int`, *optional*, defaults to `"auto"`): + When `"auto"`, halves the input to the attention heads, so attention will be computed in two steps. If + a number is provided, uses as many slices as `attention_head_dim // slice_size`. In this case, + `attention_head_dim` must be a multiple of `slice_size`. + """ + if slice_size == "auto": + # half the attention head size is usually a good trade-off between + # speed and memory + slice_size = self.image_unet.config.attention_head_dim // 2 + self.image_unet.set_attention_slice(slice_size) + self.text_unet.set_attention_slice(slice_size) + + def disable_attention_slicing(self): + r""" + Disable sliced attention computation. If `enable_attention_slicing` was previously invoked, this method will go + back to computing attention in one step. + """ + # set slice_size = `None` to disable `attention slicing` + self.enable_attention_slicing(None) + + @torch.no_grad() + def image_variation( + self, + image: Union[torch.FloatTensor, PIL.Image.Image], + height: int = 512, + width: int = 512, + num_inference_steps: int = 50, + guidance_scale: float = 7.5, + negative_prompt: Optional[Union[str, List[str]]] = None, + num_images_per_prompt: Optional[int] = 1, + eta: float = 0.0, + generator: Optional[torch.Generator] = None, + latents: Optional[torch.FloatTensor] = None, + output_type: Optional[str] = "pil", + return_dict: bool = True, + callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None, + callback_steps: Optional[int] = 1, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + image (`PIL.Image.Image`, `List[PIL.Image.Image]` or `torch.Tensor`): + The image prompt or prompts to guide the image generation. + height (`int`, *optional*, defaults to 512): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to 512): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 7.5): + Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598). + `guidance_scale` is defined as `w` of equation 2. of [Imagen + Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale > + 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`, + usually at the expense of lower image quality. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `guidance_scale` is less than `1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to + [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`torch.Generator`, *optional*): + A [torch generator](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make generation + deterministic. + latents (`torch.FloatTensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will ge generated by sampling using the supplied random `generator`. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + callback (`Callable`, *optional*): + A function that will be called every `callback_steps` steps during inference. The function will be + called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function will be called. If not specified, the callback will be + called at every step. + + Examples: + + ```py + >>> from diffusers import VersatileDiffusionPipeline + >>> import torch + >>> import requests + >>> from io import BytesIO + >>> from PIL import Image + + >>> # let's download an initial image + >>> url = "https://huggingface.co/datasets/diffusers/images/resolve/main/benz.jpg" + + >>> response = requests.get(url) + >>> image = Image.open(BytesIO(response.content)).convert("RGB") + + >>> pipe = VersatileDiffusionPipeline.from_pretrained( + ... "shi-labs/versatile-diffusion", torch_dtype=torch.float16 + ... ) + >>> pipe = pipe.to("cuda") + + >>> generator = torch.Generator(device="cuda").manual_seed(0) + >>> image = pipe(image, generator=generator).images[0] + >>> image.save("./car_variation.png") + ``` + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple. + When returning a tuple, the first element is a list with the generated images, and the second element is a + list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work" + (nsfw) content, according to the `safety_checker`. + """ + expected_components = inspect.signature(VersatileDiffusionImageVariationPipeline.__init__).parameters.keys() + components = {name: component for name, component in self.components.items() if name in expected_components} + return VersatileDiffusionImageVariationPipeline(**components)( + image=image, + height=height, + width=width, + num_inference_steps=num_inference_steps, + guidance_scale=guidance_scale, + negative_prompt=negative_prompt, + num_images_per_prompt=num_images_per_prompt, + eta=eta, + generator=generator, + latents=latents, + output_type=output_type, + return_dict=return_dict, + callback=callback, + callback_steps=callback_steps, + ) + + @torch.no_grad() + def text_to_image( + self, + prompt: Union[str, List[str]], + height: int = 512, + width: int = 512, + num_inference_steps: int = 50, + guidance_scale: float = 7.5, + negative_prompt: Optional[Union[str, List[str]]] = None, + num_images_per_prompt: Optional[int] = 1, + eta: float = 0.0, + generator: Optional[torch.Generator] = None, + latents: Optional[torch.FloatTensor] = None, + output_type: Optional[str] = "pil", + return_dict: bool = True, + callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None, + callback_steps: Optional[int] = 1, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `List[str]`): + The prompt or prompts to guide the image generation. + height (`int`, *optional*, defaults to 512): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to 512): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 7.5): + Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598). + `guidance_scale` is defined as `w` of equation 2. of [Imagen + Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale > + 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`, + usually at the expense of lower image quality. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `guidance_scale` is less than `1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to + [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`torch.Generator`, *optional*): + A [torch generator](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make generation + deterministic. + latents (`torch.FloatTensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will ge generated by sampling using the supplied random `generator`. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + callback (`Callable`, *optional*): + A function that will be called every `callback_steps` steps during inference. The function will be + called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function will be called. If not specified, the callback will be + called at every step. + + Examples: + + ```py + >>> from diffusers import VersatileDiffusionPipeline + >>> import torch + + >>> pipe = VersatileDiffusionPipeline.from_pretrained( + ... "shi-labs/versatile-diffusion", torch_dtype=torch.float16 + ... ) + >>> pipe = pipe.to("cuda") + + >>> generator = torch.Generator(device="cuda").manual_seed(0) + >>> image = pipe.text_to_image("an astronaut riding on a horse on mars", generator=generator).images[0] + >>> image.save("./astronaut.png") + ``` + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple. + When returning a tuple, the first element is a list with the generated images, and the second element is a + list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work" + (nsfw) content, according to the `safety_checker`. + """ + expected_components = inspect.signature(VersatileDiffusionTextToImagePipeline.__init__).parameters.keys() + components = {name: component for name, component in self.components.items() if name in expected_components} + temp_pipeline = VersatileDiffusionTextToImagePipeline(**components) + output = temp_pipeline( + prompt=prompt, + height=height, + width=width, + num_inference_steps=num_inference_steps, + guidance_scale=guidance_scale, + negative_prompt=negative_prompt, + num_images_per_prompt=num_images_per_prompt, + eta=eta, + generator=generator, + latents=latents, + output_type=output_type, + return_dict=return_dict, + callback=callback, + callback_steps=callback_steps, + ) + # swap the attention blocks back to the original state + temp_pipeline._swap_unet_attention_blocks() + + return output + + @torch.no_grad() + def dual_guided( + self, + prompt: Union[PIL.Image.Image, List[PIL.Image.Image]], + image: Union[str, List[str]], + text_to_image_strength: float = 0.5, + height: int = 512, + width: int = 512, + num_inference_steps: int = 50, + guidance_scale: float = 7.5, + num_images_per_prompt: Optional[int] = 1, + eta: float = 0.0, + generator: Optional[torch.Generator] = None, + latents: Optional[torch.FloatTensor] = None, + output_type: Optional[str] = "pil", + return_dict: bool = True, + callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None, + callback_steps: Optional[int] = 1, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `List[str]`): + The prompt or prompts to guide the image generation. + height (`int`, *optional*, defaults to 512): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to 512): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 7.5): + Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598). + `guidance_scale` is defined as `w` of equation 2. of [Imagen + Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale > + 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`, + usually at the expense of lower image quality. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `guidance_scale` is less than `1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to + [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`torch.Generator`, *optional*): + A [torch generator](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make generation + deterministic. + latents (`torch.FloatTensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will ge generated by sampling using the supplied random `generator`. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + callback (`Callable`, *optional*): + A function that will be called every `callback_steps` steps during inference. The function will be + called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function will be called. If not specified, the callback will be + called at every step. + + Examples: + + ```py + >>> from diffusers import VersatileDiffusionPipeline + >>> import torch + >>> import requests + >>> from io import BytesIO + >>> from PIL import Image + + >>> # let's download an initial image + >>> url = "https://huggingface.co/datasets/diffusers/images/resolve/main/benz.jpg" + + >>> response = requests.get(url) + >>> image = Image.open(BytesIO(response.content)).convert("RGB") + >>> text = "a red car in the sun" + + >>> pipe = VersatileDiffusionPipeline.from_pretrained( + ... "shi-labs/versatile-diffusion", torch_dtype=torch.float16 + ... ) + >>> pipe = pipe.to("cuda") + + >>> generator = torch.Generator(device="cuda").manual_seed(0) + >>> text_to_image_strength = 0.75 + + >>> image = pipe.dual_guided( + ... prompt=text, image=image, text_to_image_strength=text_to_image_strength, generator=generator + ... ).images[0] + >>> image.save("./car_variation.png") + ``` + + Returns: + [`~pipelines.stable_diffusion.ImagePipelineOutput`] or `tuple`: + [`~pipelines.stable_diffusion.ImagePipelineOutput`] if `return_dict` is True, otherwise a `tuple. When + returning a tuple, the first element is a list with the generated images. + """ + + expected_components = inspect.signature(VersatileDiffusionDualGuidedPipeline.__init__).parameters.keys() + components = {name: component for name, component in self.components.items() if name in expected_components} + temp_pipeline = VersatileDiffusionDualGuidedPipeline(**components) + output = temp_pipeline( + prompt=prompt, + image=image, + text_to_image_strength=text_to_image_strength, + height=height, + width=width, + num_inference_steps=num_inference_steps, + guidance_scale=guidance_scale, + num_images_per_prompt=num_images_per_prompt, + eta=eta, + generator=generator, + latents=latents, + output_type=output_type, + return_dict=return_dict, + callback=callback, + callback_steps=callback_steps, + ) + temp_pipeline._revert_dual_attention() + + return output diff --git a/src/diffusers/pipelines/versatile_diffusion/pipeline_versatile_diffusion_dual_guided.py b/src/diffusers/pipelines/versatile_diffusion/pipeline_versatile_diffusion_dual_guided.py new file mode 100644 index 0000000000..ad4e8b0d0a --- /dev/null +++ b/src/diffusers/pipelines/versatile_diffusion/pipeline_versatile_diffusion_dual_guided.py @@ -0,0 +1,628 @@ +# Copyright 2022 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Callable, List, Optional, Tuple, Union + +import numpy as np +import torch +import torch.utils.checkpoint + +import PIL +from transformers import ( + CLIPFeatureExtractor, + CLIPTextModelWithProjection, + CLIPTokenizer, + CLIPVisionModelWithProjection, +) + +from ...models import AutoencoderKL, UNet2DConditionModel +from ...models.attention import DualTransformer2DModel, Transformer2DModel +from ...pipeline_utils import DiffusionPipeline, ImagePipelineOutput +from ...schedulers import DDIMScheduler, LMSDiscreteScheduler, PNDMScheduler +from ...utils import is_accelerate_available, logging +from .modeling_text_unet import UNetFlatConditionModel + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +class VersatileDiffusionDualGuidedPipeline(DiffusionPipeline): + r""" + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + Parameters: + vqvae ([`VQModel`]): + Vector-quantized (VQ) Model to encode and decode images to and from latent representations. + bert ([`LDMBertModel`]): + Text-encoder model based on [BERT](https://huggingface.co/docs/transformers/model_doc/bert) architecture. + tokenizer (`transformers.BertTokenizer`): + Tokenizer of class + [BertTokenizer](https://huggingface.co/docs/transformers/model_doc/bert#transformers.BertTokenizer). + unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + """ + tokenizer: CLIPTokenizer + image_feature_extractor: CLIPFeatureExtractor + text_encoder: CLIPTextModelWithProjection + image_encoder: CLIPVisionModelWithProjection + image_unet: UNet2DConditionModel + text_unet: UNetFlatConditionModel + vae: AutoencoderKL + scheduler: Union[DDIMScheduler, PNDMScheduler, LMSDiscreteScheduler] + + def __init__( + self, + tokenizer: CLIPTokenizer, + image_feature_extractor: CLIPFeatureExtractor, + text_encoder: CLIPTextModelWithProjection, + image_encoder: CLIPVisionModelWithProjection, + image_unet: UNet2DConditionModel, + text_unet: UNetFlatConditionModel, + vae: AutoencoderKL, + scheduler: Union[DDIMScheduler, PNDMScheduler, LMSDiscreteScheduler], + ): + super().__init__() + self.register_modules( + tokenizer=tokenizer, + image_feature_extractor=image_feature_extractor, + text_encoder=text_encoder, + image_encoder=image_encoder, + image_unet=image_unet, + text_unet=text_unet, + vae=vae, + scheduler=scheduler, + ) + + if self.text_unet is not None and ( + "dual_cross_attention" not in self.image_unet.config or not self.image_unet.config.dual_cross_attention + ): + # if loading from a universal checkpoint rather than a saved dual-guided pipeline + self._convert_to_dual_attention() + + def remove_unused_weights(self): + self.register_modules(text_unet=None) + + def _convert_to_dual_attention(self): + """ + Replace image_unet's `Transformer2DModel` blocks with `DualTransformer2DModel` that contains transformer blocks + from both `image_unet` and `text_unet` + """ + for name, module in self.image_unet.named_modules(): + if isinstance(module, Transformer2DModel): + parent_name, index = name.rsplit(".", 1) + index = int(index) + + image_transformer = self.image_unet.get_submodule(parent_name)[index] + text_transformer = self.text_unet.get_submodule(parent_name)[index] + + config = image_transformer.config + dual_transformer = DualTransformer2DModel( + num_attention_heads=config.num_attention_heads, + attention_head_dim=config.attention_head_dim, + in_channels=config.in_channels, + num_layers=config.num_layers, + dropout=config.dropout, + norm_num_groups=config.norm_num_groups, + cross_attention_dim=config.cross_attention_dim, + attention_bias=config.attention_bias, + sample_size=config.sample_size, + num_vector_embeds=config.num_vector_embeds, + activation_fn=config.activation_fn, + num_embeds_ada_norm=config.num_embeds_ada_norm, + ) + dual_transformer.transformers[0] = image_transformer + dual_transformer.transformers[1] = text_transformer + + self.image_unet.get_submodule(parent_name)[index] = dual_transformer + self.image_unet.register_to_config(dual_cross_attention=True) + + def _revert_dual_attention(self): + """ + Revert the image_unet `DualTransformer2DModel` blocks back to `Transformer2DModel` with image_unet weights Call + this function if you reuse `image_unet` in another pipeline, e.g. `VersatileDiffusionPipeline` + """ + for name, module in self.image_unet.named_modules(): + if isinstance(module, DualTransformer2DModel): + parent_name, index = name.rsplit(".", 1) + index = int(index) + self.image_unet.get_submodule(parent_name)[index] = module.transformers[0] + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_xformers_memory_efficient_attention with unet->image_unet + def enable_xformers_memory_efficient_attention(self): + r""" + Enable memory efficient attention as implemented in xformers. + + When this option is enabled, you should observe lower GPU memory usage and a potential speed up at inference + time. Speed up at training time is not guaranteed. + + Warning: When Memory Efficient Attention and Sliced attention are both enabled, the Memory Efficient Attention + is used. + """ + self.image_unet.set_use_memory_efficient_attention_xformers(True) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_xformers_memory_efficient_attention with unet->image_unet + def disable_xformers_memory_efficient_attention(self): + r""" + Disable memory efficient attention as implemented in xformers. + """ + self.image_unet.set_use_memory_efficient_attention_xformers(False) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_attention_slicing with unet->image_unet + def enable_attention_slicing(self, slice_size: Optional[Union[str, int]] = "auto"): + r""" + Enable sliced attention computation. + + When this option is enabled, the attention module will split the input tensor in slices, to compute attention + in several steps. This is useful to save some memory in exchange for a small speed decrease. + + Args: + slice_size (`str` or `int`, *optional*, defaults to `"auto"`): + When `"auto"`, halves the input to the attention heads, so attention will be computed in two steps. If + a number is provided, uses as many slices as `attention_head_dim // slice_size`. In this case, + `attention_head_dim` must be a multiple of `slice_size`. + """ + if slice_size == "auto": + # half the attention head size is usually a good trade-off between + # speed and memory + slice_size = self.image_unet.config.attention_head_dim // 2 + self.image_unet.set_attention_slice(slice_size) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_attention_slicing + def disable_attention_slicing(self): + r""" + Disable sliced attention computation. If `enable_attention_slicing` was previously invoked, this method will go + back to computing attention in one step. + """ + # set slice_size = `None` to disable `attention slicing` + self.enable_attention_slicing(None) + + def enable_sequential_cpu_offload(self, gpu_id=0): + r""" + Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet, + text_encoder, vae and safety checker have their state dicts saved to CPU and then are moved to a + `torch.device('meta') and loaded to GPU only when their specific submodule has its `forward` method called. + """ + if is_accelerate_available(): + from accelerate import cpu_offload + else: + raise ImportError("Please install accelerate via `pip install accelerate`") + + device = torch.device(f"cuda:{gpu_id}") + + for cpu_offloaded_model in [self.image_unet, self.text_unet, self.text_encoder, self.vae]: + if cpu_offloaded_model is not None: + cpu_offload(cpu_offloaded_model, device) + + @property + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._execution_device with unet->image_unet + def _execution_device(self): + r""" + Returns the device on which the pipeline's models will be executed. After calling + `pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module + hooks. + """ + if self.device != torch.device("meta") or not hasattr(self.image_unet, "_hf_hook"): + return self.device + for module in self.image_unet.modules(): + if ( + hasattr(module, "_hf_hook") + and hasattr(module._hf_hook, "execution_device") + and module._hf_hook.execution_device is not None + ): + return torch.device(module._hf_hook.execution_device) + return self.device + + def _encode_text_prompt(self, prompt, device, num_images_per_prompt, do_classifier_free_guidance): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list(int)`): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + """ + + def normalize_embeddings(encoder_output): + embeds = self.text_encoder.text_projection(encoder_output.last_hidden_state) + embeds_pooled = encoder_output.text_embeds + embeds = embeds / torch.norm(embeds_pooled.unsqueeze(1), dim=-1, keepdim=True) + return embeds + + batch_size = len(prompt) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="max_length", return_tensors="pt").input_ids + + if not torch.equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer.batch_decode(untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = text_inputs.attention_mask.to(device) + else: + attention_mask = None + + text_embeddings = self.text_encoder( + text_input_ids.to(device), + attention_mask=attention_mask, + ) + text_embeddings = normalize_embeddings(text_embeddings) + + # duplicate text embeddings for each generation per prompt, using mps friendly method + bs_embed, seq_len, _ = text_embeddings.shape + text_embeddings = text_embeddings.repeat(1, num_images_per_prompt, 1) + text_embeddings = text_embeddings.view(bs_embed * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance: + uncond_tokens = [""] * batch_size + max_length = text_input_ids.shape[-1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = uncond_input.attention_mask.to(device) + else: + attention_mask = None + + uncond_embeddings = self.text_encoder( + uncond_input.input_ids.to(device), + attention_mask=attention_mask, + ) + uncond_embeddings = normalize_embeddings(uncond_embeddings) + + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = uncond_embeddings.shape[1] + uncond_embeddings = uncond_embeddings.repeat(1, num_images_per_prompt, 1) + uncond_embeddings = uncond_embeddings.view(batch_size * num_images_per_prompt, seq_len, -1) + + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + text_embeddings = torch.cat([uncond_embeddings, text_embeddings]) + + return text_embeddings + + def _encode_image_prompt(self, prompt, device, num_images_per_prompt, do_classifier_free_guidance): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list(int)`): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + """ + + def normalize_embeddings(encoder_output): + embeds = self.image_encoder.vision_model.post_layernorm(encoder_output.last_hidden_state) + embeds = self.image_encoder.visual_projection(embeds) + embeds_pooled = embeds[:, 0:1] + embeds = embeds / torch.norm(embeds_pooled, dim=-1, keepdim=True) + return embeds + + batch_size = len(prompt) if isinstance(prompt, list) else 1 + + # get prompt text embeddings + image_input = self.image_feature_extractor(images=prompt, return_tensors="pt") + pixel_values = image_input.pixel_values.to(device).to(self.image_encoder.dtype) + image_embeddings = self.image_encoder(pixel_values) + image_embeddings = normalize_embeddings(image_embeddings) + + # duplicate image embeddings for each generation per prompt, using mps friendly method + bs_embed, seq_len, _ = image_embeddings.shape + image_embeddings = image_embeddings.repeat(1, num_images_per_prompt, 1) + image_embeddings = image_embeddings.view(bs_embed * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance: + uncond_images = [np.zeros((512, 512, 3)) + 0.5] * batch_size + uncond_images = self.image_feature_extractor(images=uncond_images, return_tensors="pt") + pixel_values = uncond_images.pixel_values.to(device).to(self.image_encoder.dtype) + uncond_embeddings = self.image_encoder(pixel_values) + uncond_embeddings = normalize_embeddings(uncond_embeddings) + + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = uncond_embeddings.shape[1] + uncond_embeddings = uncond_embeddings.repeat(1, num_images_per_prompt, 1) + uncond_embeddings = uncond_embeddings.view(batch_size * num_images_per_prompt, seq_len, -1) + + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and conditional embeddings into a single batch + # to avoid doing two forward passes + image_embeddings = torch.cat([uncond_embeddings, image_embeddings]) + + return image_embeddings + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents + def decode_latents(self, latents): + latents = 1 / 0.18215 * latents + image = self.vae.decode(latents).sample + image = (image / 2 + 0.5).clamp(0, 1) + # we always cast to float32 as this does not cause significant overhead and is compatible with bfloa16 + image = image.cpu().permute(0, 2, 3, 1).float().numpy() + return image + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def check_inputs(self, prompt, image, height, width, callback_steps): + if not isinstance(prompt, str) and not isinstance(prompt, PIL.Image.Image) and not isinstance(prompt, list): + raise ValueError(f"`prompt` has to be of type `str` `PIL.Image` or `list` but is {type(prompt)}") + if not isinstance(image, str) and not isinstance(image, PIL.Image.Image) and not isinstance(image, list): + raise ValueError(f"`image` has to be of type `str` `PIL.Image` or `list` but is {type(image)}") + + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if (callback_steps is None) or ( + callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0) + ): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = (batch_size, num_channels_latents, height // 8, width // 8) + if latents is None: + if device.type == "mps": + # randn does not work reproducibly on mps + latents = torch.randn(shape, generator=generator, device="cpu", dtype=dtype).to(device) + else: + latents = torch.randn(shape, generator=generator, device=device, dtype=dtype) + else: + if latents.shape != shape: + raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}") + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + def set_transformer_params(self, mix_ratio: float = 0.5, condition_types: Tuple = ("text", "image")): + for name, module in self.image_unet.named_modules(): + if isinstance(module, DualTransformer2DModel): + module.mix_ratio = mix_ratio + + for i, type in enumerate(condition_types): + if type == "text": + module.condition_lengths[i] = self.text_encoder.config.max_position_embeddings + module.transformer_index_for_condition[i] = 1 # use the second (text) transformer + else: + module.condition_lengths[i] = 257 + module.transformer_index_for_condition[i] = 0 # use the first (image) transformer + + @torch.no_grad() + def __call__( + self, + prompt: Union[PIL.Image.Image, List[PIL.Image.Image]], + image: Union[str, List[str]], + text_to_image_strength: float = 0.5, + height: int = 512, + width: int = 512, + num_inference_steps: int = 50, + guidance_scale: float = 7.5, + num_images_per_prompt: Optional[int] = 1, + eta: float = 0.0, + generator: Optional[torch.Generator] = None, + latents: Optional[torch.FloatTensor] = None, + output_type: Optional[str] = "pil", + return_dict: bool = True, + callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None, + callback_steps: Optional[int] = 1, + **kwargs, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `List[str]`): + The prompt or prompts to guide the image generation. + height (`int`, *optional*, defaults to 512): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to 512): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 7.5): + Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598). + `guidance_scale` is defined as `w` of equation 2. of [Imagen + Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale > + 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`, + usually at the expense of lower image quality. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `guidance_scale` is less than `1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to + [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`torch.Generator`, *optional*): + A [torch generator](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make generation + deterministic. + latents (`torch.FloatTensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will ge generated by sampling using the supplied random `generator`. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + callback (`Callable`, *optional*): + A function that will be called every `callback_steps` steps during inference. The function will be + called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function will be called. If not specified, the callback will be + called at every step. + + Examples: + + ```py + >>> from diffusers import VersatileDiffusionDualGuidedPipeline + >>> import torch + >>> import requests + >>> from io import BytesIO + >>> from PIL import Image + + >>> # let's download an initial image + >>> url = "https://huggingface.co/datasets/diffusers/images/resolve/main/benz.jpg" + + >>> response = requests.get(url) + >>> image = Image.open(BytesIO(response.content)).convert("RGB") + >>> text = "a red car in the sun" + + >>> pipe = VersatileDiffusionDualGuidedPipeline.from_pretrained( + ... "shi-labs/versatile-diffusion", torch_dtype=torch.float16 + ... ) + >>> pipe.remove_unused_weights() + >>> pipe = pipe.to("cuda") + + >>> generator = torch.Generator(device="cuda").manual_seed(0) + >>> text_to_image_strength = 0.75 + + >>> image = pipe( + ... prompt=text, image=image, text_to_image_strength=text_to_image_strength, generator=generator + ... ).images[0] + >>> image.save("./car_variation.png") + ``` + + Returns: + [`~pipelines.stable_diffusion.ImagePipelineOutput`] or `tuple`: + [`~pipelines.stable_diffusion.ImagePipelineOutput`] if `return_dict` is True, otherwise a `tuple. When + returning a tuple, the first element is a list with the generated images. + """ + + # 1. Check inputs. Raise error if not correct + self.check_inputs(prompt, image, height, width, callback_steps) + + # 2. Define call parameters + prompt = [prompt] if not isinstance(prompt, list) else prompt + image = [image] if not isinstance(image, list) else image + batch_size = len(prompt) + device = self._execution_device + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = guidance_scale > 1.0 + + # 3. Encode input prompts + text_embeddings = self._encode_text_prompt(prompt, device, num_images_per_prompt, do_classifier_free_guidance) + image_embeddings = self._encode_image_prompt(image, device, num_images_per_prompt, do_classifier_free_guidance) + dual_prompt_embeddings = torch.cat([text_embeddings, image_embeddings], dim=1) + prompt_types = ("text", "image") + + # 4. Prepare timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + # 5. Prepare latent variables + num_channels_latents = self.image_unet.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + dual_prompt_embeddings.dtype, + device, + generator, + latents, + ) + + # 6. Prepare extra step kwargs. + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 7. Combine the attention blocks of the image and text UNets + self.set_transformer_params(text_to_image_strength, prompt_types) + + # 8. Denoising loop + for i, t in enumerate(self.progress_bar(timesteps)): + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # predict the noise residual + noise_pred = self.image_unet(latent_model_input, t, encoder_hidden_states=dual_prompt_embeddings).sample + + # perform guidance + if do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample + + # call the callback, if provided + if callback is not None and i % callback_steps == 0: + callback(i, t, latents) + + # 9. Post-processing + image = self.decode_latents(latents) + + # 10. Convert to PIL + if output_type == "pil": + image = self.numpy_to_pil(image) + + if not return_dict: + return (image,) + + return ImagePipelineOutput(images=image) diff --git a/src/diffusers/pipelines/versatile_diffusion/pipeline_versatile_diffusion_image_variation.py b/src/diffusers/pipelines/versatile_diffusion/pipeline_versatile_diffusion_image_variation.py new file mode 100644 index 0000000000..652b7b735a --- /dev/null +++ b/src/diffusers/pipelines/versatile_diffusion/pipeline_versatile_diffusion_image_variation.py @@ -0,0 +1,462 @@ +# Copyright 2022 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Callable, List, Optional, Union + +import numpy as np +import torch +import torch.utils.checkpoint + +import PIL +from transformers import CLIPFeatureExtractor, CLIPVisionModelWithProjection + +from ...models import AutoencoderKL, UNet2DConditionModel +from ...pipeline_utils import DiffusionPipeline, ImagePipelineOutput +from ...schedulers import DDIMScheduler, LMSDiscreteScheduler, PNDMScheduler +from ...utils import is_accelerate_available, logging + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +class VersatileDiffusionImageVariationPipeline(DiffusionPipeline): + r""" + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + Parameters: + vqvae ([`VQModel`]): + Vector-quantized (VQ) Model to encode and decode images to and from latent representations. + bert ([`LDMBertModel`]): + Text-encoder model based on [BERT](https://huggingface.co/docs/transformers/model_doc/bert) architecture. + tokenizer (`transformers.BertTokenizer`): + Tokenizer of class + [BertTokenizer](https://huggingface.co/docs/transformers/model_doc/bert#transformers.BertTokenizer). + unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + """ + image_feature_extractor: CLIPFeatureExtractor + image_encoder: CLIPVisionModelWithProjection + image_unet: UNet2DConditionModel + vae: AutoencoderKL + scheduler: Union[DDIMScheduler, PNDMScheduler, LMSDiscreteScheduler] + + def __init__( + self, + image_feature_extractor: CLIPFeatureExtractor, + image_encoder: CLIPVisionModelWithProjection, + image_unet: UNet2DConditionModel, + vae: AutoencoderKL, + scheduler: Union[DDIMScheduler, PNDMScheduler, LMSDiscreteScheduler], + ): + super().__init__() + self.register_modules( + image_feature_extractor=image_feature_extractor, + image_encoder=image_encoder, + image_unet=image_unet, + vae=vae, + scheduler=scheduler, + ) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_xformers_memory_efficient_attention with unet->image_unet + def enable_xformers_memory_efficient_attention(self): + r""" + Enable memory efficient attention as implemented in xformers. + + When this option is enabled, you should observe lower GPU memory usage and a potential speed up at inference + time. Speed up at training time is not guaranteed. + + Warning: When Memory Efficient Attention and Sliced attention are both enabled, the Memory Efficient Attention + is used. + """ + self.image_unet.set_use_memory_efficient_attention_xformers(True) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_xformers_memory_efficient_attention with unet->image_unet + def disable_xformers_memory_efficient_attention(self): + r""" + Disable memory efficient attention as implemented in xformers. + """ + self.image_unet.set_use_memory_efficient_attention_xformers(False) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_attention_slicing with unet->image_unet + def enable_attention_slicing(self, slice_size: Optional[Union[str, int]] = "auto"): + r""" + Enable sliced attention computation. + + When this option is enabled, the attention module will split the input tensor in slices, to compute attention + in several steps. This is useful to save some memory in exchange for a small speed decrease. + + Args: + slice_size (`str` or `int`, *optional*, defaults to `"auto"`): + When `"auto"`, halves the input to the attention heads, so attention will be computed in two steps. If + a number is provided, uses as many slices as `attention_head_dim // slice_size`. In this case, + `attention_head_dim` must be a multiple of `slice_size`. + """ + if slice_size == "auto": + # half the attention head size is usually a good trade-off between + # speed and memory + slice_size = self.image_unet.config.attention_head_dim // 2 + self.image_unet.set_attention_slice(slice_size) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_attention_slicing + def disable_attention_slicing(self): + r""" + Disable sliced attention computation. If `enable_attention_slicing` was previously invoked, this method will go + back to computing attention in one step. + """ + # set slice_size = `None` to disable `attention slicing` + self.enable_attention_slicing(None) + + def enable_sequential_cpu_offload(self, gpu_id=0): + r""" + Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet, + text_encoder, vae and safety checker have their state dicts saved to CPU and then are moved to a + `torch.device('meta') and loaded to GPU only when their specific submodule has its `forward` method called. + """ + if is_accelerate_available(): + from accelerate import cpu_offload + else: + raise ImportError("Please install accelerate via `pip install accelerate`") + + device = torch.device(f"cuda:{gpu_id}") + + for cpu_offloaded_model in [self.image_unet, self.text_unet, self.text_encoder, self.vae]: + if cpu_offloaded_model is not None: + cpu_offload(cpu_offloaded_model, device) + + @property + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._execution_device with unet->image_unet + def _execution_device(self): + r""" + Returns the device on which the pipeline's models will be executed. After calling + `pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module + hooks. + """ + if self.device != torch.device("meta") or not hasattr(self.image_unet, "_hf_hook"): + return self.device + for module in self.image_unet.modules(): + if ( + hasattr(module, "_hf_hook") + and hasattr(module._hf_hook, "execution_device") + and module._hf_hook.execution_device is not None + ): + return torch.device(module._hf_hook.execution_device) + return self.device + + def _encode_prompt(self, prompt, device, num_images_per_prompt, do_classifier_free_guidance, negative_prompt): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list(int)`): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `List[str]`): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `guidance_scale` is less than `1`). + """ + + def normalize_embeddings(encoder_output): + embeds = self.image_encoder.vision_model.post_layernorm(encoder_output.last_hidden_state) + embeds = self.image_encoder.visual_projection(embeds) + embeds_pooled = embeds[:, 0:1] + embeds = embeds / torch.norm(embeds_pooled, dim=-1, keepdim=True) + return embeds + + batch_size = len(prompt) if isinstance(prompt, list) else 1 + + # get prompt text embeddings + image_input = self.image_feature_extractor(images=prompt, return_tensors="pt") + pixel_values = image_input.pixel_values.to(device).to(self.image_encoder.dtype) + image_embeddings = self.image_encoder(pixel_values) + image_embeddings = normalize_embeddings(image_embeddings) + + # duplicate image embeddings for each generation per prompt, using mps friendly method + bs_embed, seq_len, _ = image_embeddings.shape + image_embeddings = image_embeddings.repeat(1, num_images_per_prompt, 1) + image_embeddings = image_embeddings.view(bs_embed * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance: + uncond_images: List[str] + if negative_prompt is None: + uncond_images = [np.zeros((512, 512, 3)) + 0.5] * batch_size + elif type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, PIL.Image.Image): + uncond_images = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_images = negative_prompt + + uncond_images = self.image_feature_extractor(images=uncond_images, return_tensors="pt") + pixel_values = uncond_images.pixel_values.to(device).to(self.image_encoder.dtype) + uncond_embeddings = self.image_encoder(pixel_values) + uncond_embeddings = normalize_embeddings(uncond_embeddings) + + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = uncond_embeddings.shape[1] + uncond_embeddings = uncond_embeddings.repeat(1, num_images_per_prompt, 1) + uncond_embeddings = uncond_embeddings.view(batch_size * num_images_per_prompt, seq_len, -1) + + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and conditional embeddings into a single batch + # to avoid doing two forward passes + image_embeddings = torch.cat([uncond_embeddings, image_embeddings]) + + return image_embeddings + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents + def decode_latents(self, latents): + latents = 1 / 0.18215 * latents + image = self.vae.decode(latents).sample + image = (image / 2 + 0.5).clamp(0, 1) + # we always cast to float32 as this does not cause significant overhead and is compatible with bfloa16 + image = image.cpu().permute(0, 2, 3, 1).float().numpy() + return image + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def check_inputs(self, image, height, width, callback_steps): + if not isinstance(image, PIL.Image.Image) and not isinstance(image, torch.Tensor): + raise ValueError(f"`image` has to be of type `PIL.Image.Image` or `torch.Tensor` but is {type(image)}") + + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if (callback_steps is None) or ( + callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0) + ): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = (batch_size, num_channels_latents, height // 8, width // 8) + if latents is None: + if device.type == "mps": + # randn does not work reproducibly on mps + latents = torch.randn(shape, generator=generator, device="cpu", dtype=dtype).to(device) + else: + latents = torch.randn(shape, generator=generator, device=device, dtype=dtype) + else: + if latents.shape != shape: + raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}") + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + @torch.no_grad() + def __call__( + self, + image: Union[PIL.Image.Image, List[PIL.Image.Image], torch.Tensor], + height: int = 512, + width: int = 512, + num_inference_steps: int = 50, + guidance_scale: float = 7.5, + negative_prompt: Optional[Union[str, List[str]]] = None, + num_images_per_prompt: Optional[int] = 1, + eta: float = 0.0, + generator: Optional[torch.Generator] = None, + latents: Optional[torch.FloatTensor] = None, + output_type: Optional[str] = "pil", + return_dict: bool = True, + callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None, + callback_steps: Optional[int] = 1, + **kwargs, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + image (`PIL.Image.Image`, `List[PIL.Image.Image]` or `torch.Tensor`): + The image prompt or prompts to guide the image generation. + height (`int`, *optional*, defaults to 512): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to 512): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 7.5): + Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598). + `guidance_scale` is defined as `w` of equation 2. of [Imagen + Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale > + 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`, + usually at the expense of lower image quality. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `guidance_scale` is less than `1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to + [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`torch.Generator`, *optional*): + A [torch generator](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make generation + deterministic. + latents (`torch.FloatTensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will ge generated by sampling using the supplied random `generator`. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + callback (`Callable`, *optional*): + A function that will be called every `callback_steps` steps during inference. The function will be + called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function will be called. If not specified, the callback will be + called at every step. + + Examples: + + ```py + >>> from diffusers import VersatileDiffusionImageVariationPipeline + >>> import torch + >>> import requests + >>> from io import BytesIO + >>> from PIL import Image + + >>> # let's download an initial image + >>> url = "https://huggingface.co/datasets/diffusers/images/resolve/main/benz.jpg" + + >>> response = requests.get(url) + >>> image = Image.open(BytesIO(response.content)).convert("RGB") + + >>> pipe = VersatileDiffusionImageVariationPipeline.from_pretrained( + ... "shi-labs/versatile-diffusion", torch_dtype=torch.float16 + ... ) + >>> pipe = pipe.to("cuda") + + >>> generator = torch.Generator(device="cuda").manual_seed(0) + >>> image = pipe(image, generator=generator).images[0] + >>> image.save("./car_variation.png") + ``` + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple. + When returning a tuple, the first element is a list with the generated images, and the second element is a + list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work" + (nsfw) content, according to the `safety_checker`. + """ + + # 1. Check inputs. Raise error if not correct + self.check_inputs(image, height, width, callback_steps) + + # 2. Define call parameters + batch_size = 1 if isinstance(image, PIL.Image.Image) else len(image) + device = self._execution_device + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = guidance_scale > 1.0 + + # 3. Encode input prompt + image_embeddings = self._encode_prompt( + image, device, num_images_per_prompt, do_classifier_free_guidance, negative_prompt + ) + + # 4. Prepare timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + # 5. Prepare latent variables + num_channels_latents = self.image_unet.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + image_embeddings.dtype, + device, + generator, + latents, + ) + + # 6. Prepare extra step kwargs. + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 7. Denoising loop + for i, t in enumerate(self.progress_bar(timesteps)): + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # predict the noise residual + noise_pred = self.image_unet(latent_model_input, t, encoder_hidden_states=image_embeddings).sample + + # perform guidance + if do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample + + # call the callback, if provided + if callback is not None and i % callback_steps == 0: + callback(i, t, latents) + + # 8. Post-processing + image = self.decode_latents(latents) + + # 9. Convert to PIL + if output_type == "pil": + image = self.numpy_to_pil(image) + + if not return_dict: + return (image,) + + return ImagePipelineOutput(images=image) diff --git a/src/diffusers/pipelines/versatile_diffusion/pipeline_versatile_diffusion_text_to_image.py b/src/diffusers/pipelines/versatile_diffusion/pipeline_versatile_diffusion_text_to_image.py new file mode 100644 index 0000000000..d07d734a64 --- /dev/null +++ b/src/diffusers/pipelines/versatile_diffusion/pipeline_versatile_diffusion_text_to_image.py @@ -0,0 +1,514 @@ +# Copyright 2022 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Callable, List, Optional, Union + +import torch +import torch.utils.checkpoint + +from transformers import CLIPFeatureExtractor, CLIPTextModelWithProjection, CLIPTokenizer + +from ...models import AutoencoderKL, UNet2DConditionModel +from ...models.attention import Transformer2DModel +from ...pipeline_utils import DiffusionPipeline, ImagePipelineOutput +from ...schedulers import DDIMScheduler, LMSDiscreteScheduler, PNDMScheduler +from ...utils import is_accelerate_available, logging +from .modeling_text_unet import UNetFlatConditionModel + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +class VersatileDiffusionTextToImagePipeline(DiffusionPipeline): + r""" + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + Parameters: + vqvae ([`VQModel`]): + Vector-quantized (VQ) Model to encode and decode images to and from latent representations. + bert ([`LDMBertModel`]): + Text-encoder model based on [BERT](https://huggingface.co/docs/transformers/model_doc/bert) architecture. + tokenizer (`transformers.BertTokenizer`): + Tokenizer of class + [BertTokenizer](https://huggingface.co/docs/transformers/model_doc/bert#transformers.BertTokenizer). + unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + """ + tokenizer: CLIPTokenizer + image_feature_extractor: CLIPFeatureExtractor + text_encoder: CLIPTextModelWithProjection + image_unet: UNet2DConditionModel + text_unet: UNetFlatConditionModel + vae: AutoencoderKL + scheduler: Union[DDIMScheduler, PNDMScheduler, LMSDiscreteScheduler] + + def __init__( + self, + tokenizer: CLIPTokenizer, + text_encoder: CLIPTextModelWithProjection, + image_unet: UNet2DConditionModel, + text_unet: UNetFlatConditionModel, + vae: AutoencoderKL, + scheduler: Union[DDIMScheduler, PNDMScheduler, LMSDiscreteScheduler], + ): + super().__init__() + self.register_modules( + tokenizer=tokenizer, + text_encoder=text_encoder, + image_unet=image_unet, + text_unet=text_unet, + vae=vae, + scheduler=scheduler, + ) + + if self.text_unet is not None: + self._swap_unet_attention_blocks() + + def _swap_unet_attention_blocks(self): + """ + Swap the `Transformer2DModel` blocks between the image and text UNets + """ + for name, module in self.image_unet.named_modules(): + if isinstance(module, Transformer2DModel): + parent_name, index = name.rsplit(".", 1) + index = int(index) + self.image_unet.get_submodule(parent_name)[index], self.text_unet.get_submodule(parent_name)[index] = ( + self.text_unet.get_submodule(parent_name)[index], + self.image_unet.get_submodule(parent_name)[index], + ) + + def remove_unused_weights(self): + self.register_modules(text_unet=None) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_xformers_memory_efficient_attention with unet->image_unet + def enable_xformers_memory_efficient_attention(self): + r""" + Enable memory efficient attention as implemented in xformers. + + When this option is enabled, you should observe lower GPU memory usage and a potential speed up at inference + time. Speed up at training time is not guaranteed. + + Warning: When Memory Efficient Attention and Sliced attention are both enabled, the Memory Efficient Attention + is used. + """ + self.image_unet.set_use_memory_efficient_attention_xformers(True) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_xformers_memory_efficient_attention with unet->image_unet + def disable_xformers_memory_efficient_attention(self): + r""" + Disable memory efficient attention as implemented in xformers. + """ + self.image_unet.set_use_memory_efficient_attention_xformers(False) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_attention_slicing with unet->image_unet + def enable_attention_slicing(self, slice_size: Optional[Union[str, int]] = "auto"): + r""" + Enable sliced attention computation. + + When this option is enabled, the attention module will split the input tensor in slices, to compute attention + in several steps. This is useful to save some memory in exchange for a small speed decrease. + + Args: + slice_size (`str` or `int`, *optional*, defaults to `"auto"`): + When `"auto"`, halves the input to the attention heads, so attention will be computed in two steps. If + a number is provided, uses as many slices as `attention_head_dim // slice_size`. In this case, + `attention_head_dim` must be a multiple of `slice_size`. + """ + if slice_size == "auto": + # half the attention head size is usually a good trade-off between + # speed and memory + slice_size = self.image_unet.config.attention_head_dim // 2 + self.image_unet.set_attention_slice(slice_size) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_attention_slicing + def disable_attention_slicing(self): + r""" + Disable sliced attention computation. If `enable_attention_slicing` was previously invoked, this method will go + back to computing attention in one step. + """ + # set slice_size = `None` to disable `attention slicing` + self.enable_attention_slicing(None) + + def enable_sequential_cpu_offload(self, gpu_id=0): + r""" + Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet, + text_encoder, vae and safety checker have their state dicts saved to CPU and then are moved to a + `torch.device('meta') and loaded to GPU only when their specific submodule has its `forward` method called. + """ + if is_accelerate_available(): + from accelerate import cpu_offload + else: + raise ImportError("Please install accelerate via `pip install accelerate`") + + device = torch.device(f"cuda:{gpu_id}") + + for cpu_offloaded_model in [self.image_unet, self.text_unet, self.text_encoder, self.vae]: + if cpu_offloaded_model is not None: + cpu_offload(cpu_offloaded_model, device) + + @property + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._execution_device with unet->image_unet + def _execution_device(self): + r""" + Returns the device on which the pipeline's models will be executed. After calling + `pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module + hooks. + """ + if self.device != torch.device("meta") or not hasattr(self.image_unet, "_hf_hook"): + return self.device + for module in self.image_unet.modules(): + if ( + hasattr(module, "_hf_hook") + and hasattr(module._hf_hook, "execution_device") + and module._hf_hook.execution_device is not None + ): + return torch.device(module._hf_hook.execution_device) + return self.device + + def _encode_prompt(self, prompt, device, num_images_per_prompt, do_classifier_free_guidance, negative_prompt): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list(int)`): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `List[str]`): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `guidance_scale` is less than `1`). + """ + + def normalize_embeddings(encoder_output): + embeds = self.text_encoder.text_projection(encoder_output.last_hidden_state) + embeds_pooled = encoder_output.text_embeds + embeds = embeds / torch.norm(embeds_pooled.unsqueeze(1), dim=-1, keepdim=True) + return embeds + + batch_size = len(prompt) if isinstance(prompt, list) else 1 + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="max_length", return_tensors="pt").input_ids + + if not torch.equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer.batch_decode(untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = text_inputs.attention_mask.to(device) + else: + attention_mask = None + + text_embeddings = self.text_encoder( + text_input_ids.to(device), + attention_mask=attention_mask, + ) + text_embeddings = normalize_embeddings(text_embeddings) + + # duplicate text embeddings for each generation per prompt, using mps friendly method + bs_embed, seq_len, _ = text_embeddings.shape + text_embeddings = text_embeddings.repeat(1, num_images_per_prompt, 1) + text_embeddings = text_embeddings.view(bs_embed * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance: + uncond_tokens: List[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + max_length = text_input_ids.shape[-1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = uncond_input.attention_mask.to(device) + else: + attention_mask = None + + uncond_embeddings = self.text_encoder( + uncond_input.input_ids.to(device), + attention_mask=attention_mask, + ) + uncond_embeddings = normalize_embeddings(uncond_embeddings) + + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = uncond_embeddings.shape[1] + uncond_embeddings = uncond_embeddings.repeat(1, num_images_per_prompt, 1) + uncond_embeddings = uncond_embeddings.view(batch_size * num_images_per_prompt, seq_len, -1) + + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + text_embeddings = torch.cat([uncond_embeddings, text_embeddings]) + + return text_embeddings + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents + def decode_latents(self, latents): + latents = 1 / 0.18215 * latents + image = self.vae.decode(latents).sample + image = (image / 2 + 0.5).clamp(0, 1) + # we always cast to float32 as this does not cause significant overhead and is compatible with bfloa16 + image = image.cpu().permute(0, 2, 3, 1).float().numpy() + return image + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.check_inputs + def check_inputs(self, prompt, height, width, callback_steps): + if not isinstance(prompt, str) and not isinstance(prompt, list): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if (callback_steps is None) or ( + callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0) + ): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = (batch_size, num_channels_latents, height // 8, width // 8) + if latents is None: + if device.type == "mps": + # randn does not work reproducibly on mps + latents = torch.randn(shape, generator=generator, device="cpu", dtype=dtype).to(device) + else: + latents = torch.randn(shape, generator=generator, device=device, dtype=dtype) + else: + if latents.shape != shape: + raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}") + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + @torch.no_grad() + def __call__( + self, + prompt: Union[str, List[str]], + height: int = 512, + width: int = 512, + num_inference_steps: int = 50, + guidance_scale: float = 7.5, + negative_prompt: Optional[Union[str, List[str]]] = None, + num_images_per_prompt: Optional[int] = 1, + eta: float = 0.0, + generator: Optional[torch.Generator] = None, + latents: Optional[torch.FloatTensor] = None, + output_type: Optional[str] = "pil", + return_dict: bool = True, + callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None, + callback_steps: Optional[int] = 1, + **kwargs, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `List[str]`): + The prompt or prompts to guide the image generation. + height (`int`, *optional*, defaults to 512): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to 512): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 7.5): + Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598). + `guidance_scale` is defined as `w` of equation 2. of [Imagen + Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale > + 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`, + usually at the expense of lower image quality. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `guidance_scale` is less than `1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to + [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`torch.Generator`, *optional*): + A [torch generator](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make generation + deterministic. + latents (`torch.FloatTensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will ge generated by sampling using the supplied random `generator`. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + callback (`Callable`, *optional*): + A function that will be called every `callback_steps` steps during inference. The function will be + called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function will be called. If not specified, the callback will be + called at every step. + + Examples: + + ```py + >>> from diffusers import VersatileDiffusionTextToImagePipeline + >>> import torch + + >>> pipe = VersatileDiffusionTextToImagePipeline.from_pretrained( + ... "shi-labs/versatile-diffusion", torch_dtype=torch.float16 + ... ) + >>> pipe.remove_unused_weights() + >>> pipe = pipe.to("cuda") + + >>> generator = torch.Generator(device="cuda").manual_seed(0) + >>> image = pipe("an astronaut riding on a horse on mars", generator=generator).images[0] + >>> image.save("./astronaut.png") + ``` + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple. + When returning a tuple, the first element is a list with the generated images, and the second element is a + list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work" + (nsfw) content, according to the `safety_checker`. + """ + + # 1. Check inputs. Raise error if not correct + self.check_inputs(prompt, height, width, callback_steps) + + # 2. Define call parameters + batch_size = 1 if isinstance(prompt, str) else len(prompt) + device = self._execution_device + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = guidance_scale > 1.0 + + # 3. Encode input prompt + text_embeddings = self._encode_prompt( + prompt, device, num_images_per_prompt, do_classifier_free_guidance, negative_prompt + ) + + # 4. Prepare timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + # 5. Prepare latent variables + num_channels_latents = self.image_unet.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + text_embeddings.dtype, + device, + generator, + latents, + ) + + # 6. Prepare extra step kwargs. + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 7. Denoising loop + for i, t in enumerate(self.progress_bar(timesteps)): + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # predict the noise residual + noise_pred = self.image_unet(latent_model_input, t, encoder_hidden_states=text_embeddings).sample + + # perform guidance + if do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample + + # call the callback, if provided + if callback is not None and i % callback_steps == 0: + callback(i, t, latents) + + # 9. Post-processing + image = self.decode_latents(latents) + + # 10. Convert to PIL + if output_type == "pil": + image = self.numpy_to_pil(image) + + if not return_dict: + return (image,) + + return ImagePipelineOutput(images=image) diff --git a/src/diffusers/schedulers/scheduling_euler_ancestral_discrete.py b/src/diffusers/schedulers/scheduling_euler_ancestral_discrete.py index f3abf017d9..8117f30560 100644 --- a/src/diffusers/schedulers/scheduling_euler_ancestral_discrete.py +++ b/src/diffusers/schedulers/scheduling_euler_ancestral_discrete.py @@ -189,7 +189,7 @@ class EulerAncestralDiscreteScheduler(SchedulerMixin, ConfigMixin): ) if not self.is_scale_input_called: - logger.warn( + logger.warning( "The `scale_model_input` function should be called before `step` to ensure correct denoising. " "See `StableDiffusionPipeline` for a usage example." ) diff --git a/src/diffusers/schedulers/scheduling_euler_discrete.py b/src/diffusers/schedulers/scheduling_euler_discrete.py index d9991bc3a0..3b2262fcc6 100644 --- a/src/diffusers/schedulers/scheduling_euler_discrete.py +++ b/src/diffusers/schedulers/scheduling_euler_discrete.py @@ -198,7 +198,7 @@ class EulerDiscreteScheduler(SchedulerMixin, ConfigMixin): ) if not self.is_scale_input_called: - logger.warn( + logger.warning( "The `scale_model_input` function should be called before `step` to ensure correct denoising. " "See `StableDiffusionPipeline` for a usage example." ) diff --git a/src/diffusers/schedulers/scheduling_lms_discrete.py b/src/diffusers/schedulers/scheduling_lms_discrete.py index 8a9aedb41b..cc9e8d7256 100644 --- a/src/diffusers/schedulers/scheduling_lms_discrete.py +++ b/src/diffusers/schedulers/scheduling_lms_discrete.py @@ -243,19 +243,18 @@ class LMSDiscreteScheduler(SchedulerMixin, ConfigMixin): timesteps: torch.FloatTensor, ) -> torch.FloatTensor: # Make sure sigmas and timesteps have the same device and dtype as original_samples - self.sigmas = self.sigmas.to(device=original_samples.device, dtype=original_samples.dtype) + sigmas = self.sigmas.to(device=original_samples.device, dtype=original_samples.dtype) if original_samples.device.type == "mps" and torch.is_floating_point(timesteps): # mps does not support float64 - self.timesteps = self.timesteps.to(original_samples.device, dtype=torch.float32) + schedule_timesteps = self.timesteps.to(original_samples.device, dtype=torch.float32) timesteps = timesteps.to(original_samples.device, dtype=torch.float32) else: - self.timesteps = self.timesteps.to(original_samples.device) + schedule_timesteps = self.timesteps.to(original_samples.device) timesteps = timesteps.to(original_samples.device) - schedule_timesteps = self.timesteps step_indices = [(schedule_timesteps == t).nonzero().item() for t in timesteps] - sigma = self.sigmas[step_indices].flatten() + sigma = sigmas[step_indices].flatten() while len(sigma.shape) < len(original_samples.shape): sigma = sigma.unsqueeze(-1) diff --git a/src/diffusers/schedulers/scheduling_utils_flax.py b/src/diffusers/schedulers/scheduling_utils_flax.py index b3024ca450..5dc28c25d9 100644 --- a/src/diffusers/schedulers/scheduling_utils_flax.py +++ b/src/diffusers/schedulers/scheduling_utils_flax.py @@ -118,7 +118,10 @@ class FlaxSchedulerMixin: """ config, kwargs = cls.load_config( - pretrained_model_name_or_path=pretrained_model_name_or_path, return_unused_kwargs=True, **kwargs + pretrained_model_name_or_path=pretrained_model_name_or_path, + subfolder=subfolder, + return_unused_kwargs=True, + **kwargs, ) scheduler, unused_kwargs = cls.from_config(config, return_unused_kwargs=True, **kwargs) diff --git a/src/diffusers/utils/__init__.py b/src/diffusers/utils/__init__.py index 909d878ed6..e86f3b801a 100644 --- a/src/diffusers/utils/__init__.py +++ b/src/diffusers/utils/__init__.py @@ -33,6 +33,7 @@ from .import_utils import ( is_torch_available, is_torch_version, is_transformers_available, + is_transformers_version, is_unidecode_available, requires_backends, ) diff --git a/src/diffusers/utils/dummy_torch_and_transformers_and_onnx_objects.py b/src/diffusers/utils/dummy_torch_and_transformers_and_onnx_objects.py index 221020030e..ae9412a956 100644 --- a/src/diffusers/utils/dummy_torch_and_transformers_and_onnx_objects.py +++ b/src/diffusers/utils/dummy_torch_and_transformers_and_onnx_objects.py @@ -34,6 +34,21 @@ class OnnxStableDiffusionInpaintPipeline(metaclass=DummyObject): requires_backends(cls, ["torch", "transformers", "onnx"]) +class OnnxStableDiffusionInpaintPipelineLegacy(metaclass=DummyObject): + _backends = ["torch", "transformers", "onnx"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers", "onnx"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers", "onnx"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers", "onnx"]) + + class OnnxStableDiffusionPipeline(metaclass=DummyObject): _backends = ["torch", "transformers", "onnx"] diff --git a/src/diffusers/utils/dummy_torch_and_transformers_objects.py b/src/diffusers/utils/dummy_torch_and_transformers_objects.py index 92c163ba74..d255c174c7 100644 --- a/src/diffusers/utils/dummy_torch_and_transformers_objects.py +++ b/src/diffusers/utils/dummy_torch_and_transformers_objects.py @@ -64,6 +64,21 @@ class LDMTextToImagePipeline(metaclass=DummyObject): requires_backends(cls, ["torch", "transformers"]) +class StableDiffusionImageVariationPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + class StableDiffusionImg2ImgPipeline(metaclass=DummyObject): _backends = ["torch", "transformers"] @@ -124,6 +139,81 @@ class StableDiffusionPipeline(metaclass=DummyObject): requires_backends(cls, ["torch", "transformers"]) +class StableDiffusionPipelineSafe(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class VersatileDiffusionDualGuidedPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class VersatileDiffusionImageVariationPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class VersatileDiffusionPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class VersatileDiffusionTextToImagePipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + class VQDiffusionPipeline(metaclass=DummyObject): _backends = ["torch", "transformers"] diff --git a/src/diffusers/utils/import_utils.py b/src/diffusers/utils/import_utils.py index 005cbb6170..c0294b4a3d 100644 --- a/src/diffusers/utils/import_utils.py +++ b/src/diffusers/utils/import_utils.py @@ -303,6 +303,17 @@ def requires_backends(obj, backends): if failed: raise ImportError("".join(failed)) + if name in [ + "VersatileDiffusionTextToImagePipeline", + "VersatileDiffusionPipeline", + "VersatileDiffusionDualGuidedPipeline", + "StableDiffusionImageVariationPipeline", + ] and is_transformers_version("<", "4.25.0.dev0"): + raise ImportError( + f"You need to install `transformers` from 'main' in order to use {name}: \n```\n pip install" + " git+https://github.com/huggingface/transformers \n```" + ) + class DummyObject(type): """ @@ -347,3 +358,17 @@ def is_torch_version(operation: str, version: str): A string version of PyTorch """ return compare_versions(parse(_torch_version), operation, version) + + +def is_transformers_version(operation: str, version: str): + """ + Args: + Compares the current Transformers version to a given reference with an operation. + operation (`str`): + A string representation of an operator, such as `">"` or `"<="` + version (`str`): + A string version of PyTorch + """ + if not _transformers_available: + return False + return compare_versions(parse(_transformers_version), operation, version) diff --git a/tests/models/test_models_unet_2d.py b/tests/models/test_models_unet_2d.py index 81437311c6..02c6d314bf 100644 --- a/tests/models/test_models_unet_2d.py +++ b/tests/models/test_models_unet_2d.py @@ -296,6 +296,44 @@ class UNet2DConditionModelTests(ModelTesterMixin, unittest.TestCase): for name, param in named_params.items(): self.assertTrue(torch_all_close(param.grad.data, named_params_2[name].grad.data, atol=5e-5)) + def test_model_with_attention_head_dim_tuple(self): + init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common() + + init_dict["attention_head_dim"] = (8, 16) + + model = self.model_class(**init_dict) + model.to(torch_device) + model.eval() + + with torch.no_grad(): + output = model(**inputs_dict) + + if isinstance(output, dict): + output = output.sample + + self.assertIsNotNone(output) + expected_shape = inputs_dict["sample"].shape + self.assertEqual(output.shape, expected_shape, "Input and output shapes do not match") + + def test_model_with_use_linear_projection(self): + init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common() + + init_dict["use_linear_projection"] = True + + model = self.model_class(**init_dict) + model.to(torch_device) + model.eval() + + with torch.no_grad(): + output = model(**inputs_dict) + + if isinstance(output, dict): + output = output.sample + + self.assertIsNotNone(output) + expected_shape = inputs_dict["sample"].shape + self.assertEqual(output.shape, expected_shape, "Input and output shapes do not match") + class NCSNppModelTests(ModelTesterMixin, unittest.TestCase): model_class = UNet2DModel diff --git a/tests/pipelines/latent_diffusion/test_latent_diffusion_superresolution.py b/tests/pipelines/latent_diffusion/test_latent_diffusion_superresolution.py index c04210dede..6f1f51c7ba 100644 --- a/tests/pipelines/latent_diffusion/test_latent_diffusion_superresolution.py +++ b/tests/pipelines/latent_diffusion/test_latent_diffusion_superresolution.py @@ -87,6 +87,27 @@ class LDMSuperResolutionPipelineFastTests(PipelineTesterMixin, unittest.TestCase expected_slice = np.array([0.8678, 0.8245, 0.6381, 0.6830, 0.4385, 0.5599, 0.4641, 0.6201, 0.5150]) assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2 + @unittest.skipIf(torch_device != "cuda", "This test requires a GPU") + def test_inference_superresolution_fp16(self): + unet = self.dummy_uncond_unet + scheduler = DDIMScheduler() + vqvae = self.dummy_vq_model + + # put models in fp16 + unet = unet.half() + vqvae = vqvae.half() + + ldm = LDMSuperResolutionPipeline(unet=unet, vqvae=vqvae, scheduler=scheduler) + ldm.to(torch_device) + ldm.set_progress_bar_config(disable=None) + + init_image = self.dummy_image.to(torch_device) + + generator = torch.Generator(device=torch_device).manual_seed(0) + image = ldm(init_image, generator=generator, num_inference_steps=2, output_type="numpy").images + + assert image.shape == (1, 64, 64, 3) + @slow @require_torch diff --git a/tests/pipelines/stable_diffusion/test_onnx_stable_diffusion_inpaint_legacy.py b/tests/pipelines/stable_diffusion/test_onnx_stable_diffusion_inpaint_legacy.py new file mode 100644 index 0000000000..577023f705 --- /dev/null +++ b/tests/pipelines/stable_diffusion/test_onnx_stable_diffusion_inpaint_legacy.py @@ -0,0 +1,95 @@ +# coding=utf-8 +# Copyright 2022 HuggingFace Inc. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import unittest + +import numpy as np + +from diffusers import OnnxStableDiffusionInpaintPipelineLegacy +from diffusers.utils.testing_utils import ( + is_onnx_available, + load_image, + load_numpy, + require_onnxruntime, + require_torch_gpu, + slow, +) + + +if is_onnx_available(): + import onnxruntime as ort + + +@slow +@require_onnxruntime +@require_torch_gpu +class StableDiffusionOnnxInpaintLegacyPipelineIntegrationTests(unittest.TestCase): + @property + def gpu_provider(self): + return ( + "CUDAExecutionProvider", + { + "gpu_mem_limit": "15000000000", # 15GB + "arena_extend_strategy": "kSameAsRequested", + }, + ) + + @property + def gpu_options(self): + options = ort.SessionOptions() + options.enable_mem_pattern = False + return options + + def test_inference(self): + init_image = load_image( + "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main" + "/in_paint/overture-creations-5sI6fQgYIuo.png" + ) + mask_image = load_image( + "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main" + "/in_paint/overture-creations-5sI6fQgYIuo_mask.png" + ) + expected_image = load_numpy( + "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main" + "/in_paint/red_cat_sitting_on_a_park_bench_onnx.npy" + ) + + # using the PNDM scheduler by default + pipe = OnnxStableDiffusionInpaintPipelineLegacy.from_pretrained( + "CompVis/stable-diffusion-v1-4", + revision="onnx", + provider=self.gpu_provider, + sess_options=self.gpu_options, + ) + pipe.set_progress_bar_config(disable=None) + + prompt = "A red cat sitting on a park bench" + + generator = np.random.RandomState(0) + output = pipe( + prompt=prompt, + init_image=init_image, + mask_image=mask_image, + strength=0.75, + guidance_scale=7.5, + num_inference_steps=15, + generator=generator, + output_type="np", + ) + + image = output.images[0] + + assert image.shape == (512, 512, 3) + assert np.abs(expected_image - image).max() < 1e-2 diff --git a/tests/pipelines/stable_diffusion/test_stable_diffusion_image_variation.py b/tests/pipelines/stable_diffusion/test_stable_diffusion_image_variation.py new file mode 100644 index 0000000000..2935275d0f --- /dev/null +++ b/tests/pipelines/stable_diffusion/test_stable_diffusion_image_variation.py @@ -0,0 +1,424 @@ +# coding=utf-8 +# Copyright 2022 HuggingFace Inc. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import gc +import random +import unittest + +import numpy as np +import torch + +from diffusers import ( + AutoencoderKL, + LMSDiscreteScheduler, + PNDMScheduler, + StableDiffusionImageVariationPipeline, + UNet2DConditionModel, +) +from diffusers.utils import floats_tensor, load_image, load_numpy, slow, torch_device +from diffusers.utils.testing_utils import require_torch_gpu +from transformers import CLIPVisionConfig, CLIPVisionModelWithProjection + +from ...test_pipelines_common import PipelineTesterMixin + + +torch.backends.cuda.matmul.allow_tf32 = False + + +class StableDiffusionImageVariationPipelineFastTests(PipelineTesterMixin, unittest.TestCase): + def tearDown(self): + # clean up the VRAM after each test + super().tearDown() + gc.collect() + torch.cuda.empty_cache() + + @property + def dummy_image(self): + batch_size = 1 + num_channels = 3 + sizes = (32, 32) + + image = floats_tensor((batch_size, num_channels) + sizes, rng=random.Random(0)).to(torch_device) + return image + + @property + def dummy_cond_unet(self): + torch.manual_seed(0) + model = UNet2DConditionModel( + block_out_channels=(32, 64), + layers_per_block=2, + sample_size=32, + in_channels=4, + out_channels=4, + down_block_types=("DownBlock2D", "CrossAttnDownBlock2D"), + up_block_types=("CrossAttnUpBlock2D", "UpBlock2D"), + cross_attention_dim=32, + ) + return model + + @property + def dummy_vae(self): + torch.manual_seed(0) + model = AutoencoderKL( + block_out_channels=[32, 64], + in_channels=3, + out_channels=3, + down_block_types=["DownEncoderBlock2D", "DownEncoderBlock2D"], + up_block_types=["UpDecoderBlock2D", "UpDecoderBlock2D"], + latent_channels=4, + ) + return model + + @property + def dummy_image_encoder(self): + torch.manual_seed(0) + config = CLIPVisionConfig( + hidden_size=32, + projection_dim=32, + intermediate_size=37, + layer_norm_eps=1e-05, + num_attention_heads=4, + num_hidden_layers=5, + image_size=32, + patch_size=4, + ) + return CLIPVisionModelWithProjection(config) + + @property + def dummy_extractor(self): + def extract(*args, **kwargs): + class Out: + def __init__(self): + self.pixel_values = torch.ones([0]) + + def to(self, device): + self.pixel_values.to(device) + return self + + return Out() + + return extract + + def test_stable_diffusion_img_variation_default_case(self): + device = "cpu" # ensure determinism for the device-dependent torch.Generator + unet = self.dummy_cond_unet + scheduler = PNDMScheduler(skip_prk_steps=True) + vae = self.dummy_vae + image_encoder = self.dummy_image_encoder + + init_image = self.dummy_image.to(device) + + # make sure here that pndm scheduler skips prk + sd_pipe = StableDiffusionImageVariationPipeline( + unet=unet, + scheduler=scheduler, + vae=vae, + image_encoder=image_encoder, + safety_checker=None, + feature_extractor=self.dummy_extractor, + ) + sd_pipe = sd_pipe.to(device) + sd_pipe.set_progress_bar_config(disable=None) + + generator = torch.Generator(device=device).manual_seed(0) + output = sd_pipe( + init_image, + generator=generator, + guidance_scale=6.0, + num_inference_steps=2, + output_type="np", + ) + + image = output.images + + generator = torch.Generator(device=device).manual_seed(0) + image_from_tuple = sd_pipe( + init_image, + generator=generator, + guidance_scale=6.0, + num_inference_steps=2, + output_type="np", + return_dict=False, + )[0] + + image_slice = image[0, -3:, -3:, -1] + print(image_slice.flatten()) + image_from_tuple_slice = image_from_tuple[0, -3:, -3:, -1] + + assert image.shape == (1, 128, 128, 3) + expected_slice = np.array([0.4935, 0.4784, 0.4802, 0.5027, 0.4805, 0.5149, 0.5143, 0.4879, 0.4731]) + assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-3 + assert np.abs(image_from_tuple_slice.flatten() - expected_slice).max() < 1e-3 + + def test_stable_diffusion_img_variation_multiple_images(self): + device = "cpu" # ensure determinism for the device-dependent torch.Generator + unet = self.dummy_cond_unet + scheduler = PNDMScheduler(skip_prk_steps=True) + vae = self.dummy_vae + image_encoder = self.dummy_image_encoder + + init_image = self.dummy_image.to(device).repeat(2, 1, 1, 1) + + # make sure here that pndm scheduler skips prk + sd_pipe = StableDiffusionImageVariationPipeline( + unet=unet, + scheduler=scheduler, + vae=vae, + image_encoder=image_encoder, + safety_checker=None, + feature_extractor=self.dummy_extractor, + ) + sd_pipe = sd_pipe.to(device) + sd_pipe.set_progress_bar_config(disable=None) + + generator = torch.Generator(device=device).manual_seed(0) + output = sd_pipe( + init_image, + generator=generator, + guidance_scale=6.0, + num_inference_steps=2, + output_type="np", + ) + + image = output.images + + image_slice = image[-1, -3:, -3:, -1] + + assert image.shape == (2, 128, 128, 3) + expected_slice = np.array([0.4939, 0.4627, 0.4831, 0.5710, 0.5387, 0.4428, 0.5230, 0.5545, 0.4586]) + assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-3 + + def test_stable_diffusion_img_variation_num_images_per_prompt(self): + device = "cpu" + unet = self.dummy_cond_unet + scheduler = PNDMScheduler(skip_prk_steps=True) + vae = self.dummy_vae + image_encoder = self.dummy_image_encoder + + init_image = self.dummy_image.to(device) + + # make sure here that pndm scheduler skips prk + sd_pipe = StableDiffusionImageVariationPipeline( + unet=unet, + scheduler=scheduler, + vae=vae, + image_encoder=image_encoder, + safety_checker=None, + feature_extractor=self.dummy_extractor, + ) + sd_pipe = sd_pipe.to(device) + sd_pipe.set_progress_bar_config(disable=None) + + # test num_images_per_prompt=1 (default) + images = sd_pipe( + init_image, + num_inference_steps=2, + output_type="np", + ).images + + assert images.shape == (1, 128, 128, 3) + + # test num_images_per_prompt=1 (default) for batch of images + batch_size = 2 + images = sd_pipe( + init_image.repeat(batch_size, 1, 1, 1), + num_inference_steps=2, + output_type="np", + ).images + + assert images.shape == (batch_size, 128, 128, 3) + + # test num_images_per_prompt for single prompt + num_images_per_prompt = 2 + images = sd_pipe( + init_image, + num_inference_steps=2, + output_type="np", + num_images_per_prompt=num_images_per_prompt, + ).images + + assert images.shape == (num_images_per_prompt, 128, 128, 3) + + # test num_images_per_prompt for batch of prompts + batch_size = 2 + images = sd_pipe( + init_image.repeat(batch_size, 1, 1, 1), + num_inference_steps=2, + output_type="np", + num_images_per_prompt=num_images_per_prompt, + ).images + + assert images.shape == (batch_size * num_images_per_prompt, 128, 128, 3) + + @unittest.skipIf(torch_device != "cuda", "This test requires a GPU") + def test_stable_diffusion_img_variation_fp16(self): + """Test that stable diffusion img2img works with fp16""" + unet = self.dummy_cond_unet + scheduler = PNDMScheduler(skip_prk_steps=True) + vae = self.dummy_vae + image_encoder = self.dummy_image_encoder + + init_image = self.dummy_image.to(torch_device).float() + + # put models in fp16 + unet = unet.half() + vae = vae.half() + image_encoder = image_encoder.half() + + # make sure here that pndm scheduler skips prk + sd_pipe = StableDiffusionImageVariationPipeline( + unet=unet, + scheduler=scheduler, + vae=vae, + image_encoder=image_encoder, + safety_checker=None, + feature_extractor=self.dummy_extractor, + ) + sd_pipe = sd_pipe.to(torch_device) + sd_pipe.set_progress_bar_config(disable=None) + + generator = torch.Generator(device=torch_device).manual_seed(0) + image = sd_pipe( + init_image, + generator=generator, + num_inference_steps=2, + output_type="np", + ).images + + assert image.shape == (1, 128, 128, 3) + + +@slow +@require_torch_gpu +class StableDiffusionImageVariationPipelineIntegrationTests(unittest.TestCase): + def tearDown(self): + # clean up the VRAM after each test + super().tearDown() + gc.collect() + torch.cuda.empty_cache() + + def test_stable_diffusion_img_variation_pipeline_default(self): + init_image = load_image( + "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/img2img/vermeer.jpg" + ) + init_image = init_image.resize((512, 512)) + expected_image = load_numpy( + "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/img2img/vermeer.npy" + ) + + model_id = "fusing/sd-image-variations-diffusers" + pipe = StableDiffusionImageVariationPipeline.from_pretrained( + model_id, + safety_checker=None, + ) + pipe.to(torch_device) + pipe.set_progress_bar_config(disable=None) + pipe.enable_attention_slicing() + + generator = torch.Generator(device=torch_device).manual_seed(0) + output = pipe( + init_image, + guidance_scale=7.5, + generator=generator, + output_type="np", + ) + image = output.images[0] + + assert image.shape == (512, 512, 3) + # img2img is flaky across GPUs even in fp32, so using MAE here + assert np.abs(expected_image - image).max() < 1e-3 + + def test_stable_diffusion_img_variation_intermediate_state(self): + number_of_steps = 0 + + def test_callback_fn(step: int, timestep: int, latents: torch.FloatTensor) -> None: + test_callback_fn.has_been_called = True + nonlocal number_of_steps + number_of_steps += 1 + if step == 0: + latents = latents.detach().cpu().numpy() + assert latents.shape == (1, 4, 64, 64) + latents_slice = latents[0, -3:, -3:, -1] + expected_slice = np.array([1.83, 1.293, -0.09705, 1.256, -2.293, 1.091, -0.0809, -0.65, -2.953]) + assert np.abs(latents_slice.flatten() - expected_slice).max() < 1e-3 + elif step == 37: + latents = latents.detach().cpu().numpy() + assert latents.shape == (1, 4, 64, 64) + latents_slice = latents[0, -3:, -3:, -1] + expected_slice = np.array([2.285, 2.703, 1.969, 0.696, -1.323, 0.9253, -0.5464, -1.521, -2.537]) + assert np.abs(latents_slice.flatten() - expected_slice).max() < 1e-2 + + test_callback_fn.has_been_called = False + + init_image = load_image( + "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main" + "/img2img/sketch-mountains-input.jpg" + ) + init_image = init_image.resize((512, 512)) + + pipe = StableDiffusionImageVariationPipeline.from_pretrained( + "fusing/sd-image-variations-diffusers", + torch_dtype=torch.float16, + ) + pipe.to(torch_device) + pipe.set_progress_bar_config(disable=None) + pipe.enable_attention_slicing() + + generator = torch.Generator(device=torch_device).manual_seed(0) + with torch.autocast(torch_device): + pipe( + init_image, + num_inference_steps=50, + guidance_scale=7.5, + generator=generator, + callback=test_callback_fn, + callback_steps=1, + ) + assert test_callback_fn.has_been_called + assert number_of_steps == 51 + + def test_stable_diffusion_pipeline_with_sequential_cpu_offloading(self): + torch.cuda.empty_cache() + torch.cuda.reset_max_memory_allocated() + torch.cuda.reset_peak_memory_stats() + + init_image = load_image( + "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main" + "/img2img/sketch-mountains-input.jpg" + ) + init_image = init_image.resize((512, 512)) + + model_id = "fusing/sd-image-variations-diffusers" + lms = LMSDiscreteScheduler.from_pretrained(model_id, subfolder="scheduler") + pipe = StableDiffusionImageVariationPipeline.from_pretrained( + model_id, scheduler=lms, safety_checker=None, torch_dtype=torch.float16 + ) + pipe.to(torch_device) + pipe.set_progress_bar_config(disable=None) + pipe.enable_attention_slicing(1) + pipe.enable_sequential_cpu_offload() + + generator = torch.Generator(device=torch_device).manual_seed(0) + _ = pipe( + init_image, + guidance_scale=7.5, + generator=generator, + output_type="np", + num_inference_steps=5, + ) + + mem_bytes = torch.cuda.max_memory_allocated() + # make sure that less than 2.6 GB is allocated + assert mem_bytes < 2.6 * 10**9 diff --git a/tests/pipelines/stable_diffusion/test_stable_diffusion_inpaint.py b/tests/pipelines/stable_diffusion/test_stable_diffusion_inpaint.py index ce231a1a46..2f9348c5b5 100644 --- a/tests/pipelines/stable_diffusion/test_stable_diffusion_inpaint.py +++ b/tests/pipelines/stable_diffusion/test_stable_diffusion_inpaint.py @@ -22,12 +22,14 @@ import torch from diffusers import ( AutoencoderKL, + LMSDiscreteScheduler, PNDMScheduler, StableDiffusionInpaintPipeline, UNet2DConditionModel, UNet2DModel, VQModel, ) +from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_inpaint import prepare_mask_and_masked_image from diffusers.utils import floats_tensor, load_image, load_numpy, slow, torch_device from diffusers.utils.testing_utils import require_torch_gpu from PIL import Image @@ -421,6 +423,45 @@ class StableDiffusionInpaintPipelineIntegrationTests(unittest.TestCase): assert image.shape == (512, 512, 3) assert np.abs(expected_image - image).max() < 1e-2 + def test_stable_diffusion_inpaint_pipeline_k_lms(self): + init_image = load_image( + "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main" + "/in_paint/overture-creations-5sI6fQgYIuo.png" + ) + mask_image = load_image( + "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main" + "/in_paint/overture-creations-5sI6fQgYIuo_mask.png" + ) + expected_image = load_numpy( + "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/in_paint" + "/yellow_cat_sitting_on_a_park_bench_k_lms.npy" + ) + + model_id = "runwayml/stable-diffusion-inpainting" + pipe = StableDiffusionInpaintPipeline.from_pretrained(model_id, safety_checker=None) + pipe.to(torch_device) + + # switch to LMS + pipe.scheduler = LMSDiscreteScheduler.from_config(pipe.scheduler.config) + + pipe.set_progress_bar_config(disable=None) + pipe.enable_attention_slicing() + + prompt = "Face of a yellow cat, high resolution, sitting on a park bench" + + generator = torch.Generator(device=torch_device).manual_seed(0) + output = pipe( + prompt=prompt, + image=init_image, + mask_image=mask_image, + generator=generator, + output_type="np", + ) + image = output.images[0] + + assert image.shape == (512, 512, 3) + assert np.abs(expected_image - image).max() < 1e-2 + @unittest.skipIf(torch_device == "cpu", "This test is supposed to run on GPU") def test_stable_diffusion_pipeline_with_sequential_cpu_offloading(self): torch.cuda.empty_cache() @@ -466,3 +507,172 @@ class StableDiffusionInpaintPipelineIntegrationTests(unittest.TestCase): mem_bytes = torch.cuda.max_memory_allocated() # make sure that less than 2.2 GB is allocated assert mem_bytes < 2.2 * 10**9 + + +class StableDiffusionInpaintingPrepareMaskAndMaskedImageTests(unittest.TestCase): + def test_pil_inputs(self): + im = np.random.randint(0, 255, (32, 32, 3), dtype=np.uint8) + im = Image.fromarray(im) + mask = np.random.randint(0, 255, (32, 32), dtype=np.uint8) > 127.5 + mask = Image.fromarray((mask * 255).astype(np.uint8)) + + t_mask, t_masked = prepare_mask_and_masked_image(im, mask) + + self.assertTrue(isinstance(t_mask, torch.Tensor)) + self.assertTrue(isinstance(t_masked, torch.Tensor)) + + self.assertEqual(t_mask.ndim, 4) + self.assertEqual(t_masked.ndim, 4) + + self.assertEqual(t_mask.shape, (1, 1, 32, 32)) + self.assertEqual(t_masked.shape, (1, 3, 32, 32)) + + self.assertTrue(t_mask.dtype == torch.float32) + self.assertTrue(t_masked.dtype == torch.float32) + + self.assertTrue(t_mask.min() >= 0.0) + self.assertTrue(t_mask.max() <= 1.0) + self.assertTrue(t_masked.min() >= -1.0) + self.assertTrue(t_masked.min() <= 1.0) + + self.assertTrue(t_mask.sum() > 0.0) + + def test_np_inputs(self): + im_np = np.random.randint(0, 255, (32, 32, 3), dtype=np.uint8) + im_pil = Image.fromarray(im_np) + mask_np = np.random.randint(0, 255, (32, 32), dtype=np.uint8) > 127.5 + mask_pil = Image.fromarray((mask_np * 255).astype(np.uint8)) + + t_mask_np, t_masked_np = prepare_mask_and_masked_image(im_np, mask_np) + t_mask_pil, t_masked_pil = prepare_mask_and_masked_image(im_pil, mask_pil) + + self.assertTrue((t_mask_np == t_mask_pil).all()) + self.assertTrue((t_masked_np == t_masked_pil).all()) + + def test_torch_3D_2D_inputs(self): + im_tensor = torch.randint(0, 255, (3, 32, 32), dtype=torch.uint8) + mask_tensor = torch.randint(0, 255, (32, 32), dtype=torch.uint8) > 127.5 + im_np = im_tensor.numpy().transpose(1, 2, 0) + mask_np = mask_tensor.numpy() + + t_mask_tensor, t_masked_tensor = prepare_mask_and_masked_image(im_tensor / 127.5 - 1, mask_tensor) + t_mask_np, t_masked_np = prepare_mask_and_masked_image(im_np, mask_np) + + self.assertTrue((t_mask_tensor == t_mask_np).all()) + self.assertTrue((t_masked_tensor == t_masked_np).all()) + + def test_torch_3D_3D_inputs(self): + im_tensor = torch.randint(0, 255, (3, 32, 32), dtype=torch.uint8) + mask_tensor = torch.randint(0, 255, (1, 32, 32), dtype=torch.uint8) > 127.5 + im_np = im_tensor.numpy().transpose(1, 2, 0) + mask_np = mask_tensor.numpy()[0] + + t_mask_tensor, t_masked_tensor = prepare_mask_and_masked_image(im_tensor / 127.5 - 1, mask_tensor) + t_mask_np, t_masked_np = prepare_mask_and_masked_image(im_np, mask_np) + + self.assertTrue((t_mask_tensor == t_mask_np).all()) + self.assertTrue((t_masked_tensor == t_masked_np).all()) + + def test_torch_4D_2D_inputs(self): + im_tensor = torch.randint(0, 255, (1, 3, 32, 32), dtype=torch.uint8) + mask_tensor = torch.randint(0, 255, (32, 32), dtype=torch.uint8) > 127.5 + im_np = im_tensor.numpy()[0].transpose(1, 2, 0) + mask_np = mask_tensor.numpy() + + t_mask_tensor, t_masked_tensor = prepare_mask_and_masked_image(im_tensor / 127.5 - 1, mask_tensor) + t_mask_np, t_masked_np = prepare_mask_and_masked_image(im_np, mask_np) + + self.assertTrue((t_mask_tensor == t_mask_np).all()) + self.assertTrue((t_masked_tensor == t_masked_np).all()) + + def test_torch_4D_3D_inputs(self): + im_tensor = torch.randint(0, 255, (1, 3, 32, 32), dtype=torch.uint8) + mask_tensor = torch.randint(0, 255, (1, 32, 32), dtype=torch.uint8) > 127.5 + im_np = im_tensor.numpy()[0].transpose(1, 2, 0) + mask_np = mask_tensor.numpy()[0] + + t_mask_tensor, t_masked_tensor = prepare_mask_and_masked_image(im_tensor / 127.5 - 1, mask_tensor) + t_mask_np, t_masked_np = prepare_mask_and_masked_image(im_np, mask_np) + + self.assertTrue((t_mask_tensor == t_mask_np).all()) + self.assertTrue((t_masked_tensor == t_masked_np).all()) + + def test_torch_4D_4D_inputs(self): + im_tensor = torch.randint(0, 255, (1, 3, 32, 32), dtype=torch.uint8) + mask_tensor = torch.randint(0, 255, (1, 1, 32, 32), dtype=torch.uint8) > 127.5 + im_np = im_tensor.numpy()[0].transpose(1, 2, 0) + mask_np = mask_tensor.numpy()[0][0] + + t_mask_tensor, t_masked_tensor = prepare_mask_and_masked_image(im_tensor / 127.5 - 1, mask_tensor) + t_mask_np, t_masked_np = prepare_mask_and_masked_image(im_np, mask_np) + + self.assertTrue((t_mask_tensor == t_mask_np).all()) + self.assertTrue((t_masked_tensor == t_masked_np).all()) + + def test_torch_batch_4D_3D(self): + im_tensor = torch.randint(0, 255, (2, 3, 32, 32), dtype=torch.uint8) + mask_tensor = torch.randint(0, 255, (2, 32, 32), dtype=torch.uint8) > 127.5 + + im_nps = [im.numpy().transpose(1, 2, 0) for im in im_tensor] + mask_nps = [mask.numpy() for mask in mask_tensor] + + t_mask_tensor, t_masked_tensor = prepare_mask_and_masked_image(im_tensor / 127.5 - 1, mask_tensor) + nps = [prepare_mask_and_masked_image(i, m) for i, m in zip(im_nps, mask_nps)] + t_mask_np = torch.cat([n[0] for n in nps]) + t_masked_np = torch.cat([n[1] for n in nps]) + + self.assertTrue((t_mask_tensor == t_mask_np).all()) + self.assertTrue((t_masked_tensor == t_masked_np).all()) + + def test_torch_batch_4D_4D(self): + im_tensor = torch.randint(0, 255, (2, 3, 32, 32), dtype=torch.uint8) + mask_tensor = torch.randint(0, 255, (2, 1, 32, 32), dtype=torch.uint8) > 127.5 + + im_nps = [im.numpy().transpose(1, 2, 0) for im in im_tensor] + mask_nps = [mask.numpy()[0] for mask in mask_tensor] + + t_mask_tensor, t_masked_tensor = prepare_mask_and_masked_image(im_tensor / 127.5 - 1, mask_tensor) + nps = [prepare_mask_and_masked_image(i, m) for i, m in zip(im_nps, mask_nps)] + t_mask_np = torch.cat([n[0] for n in nps]) + t_masked_np = torch.cat([n[1] for n in nps]) + + self.assertTrue((t_mask_tensor == t_mask_np).all()) + self.assertTrue((t_masked_tensor == t_masked_np).all()) + + def test_shape_mismatch(self): + # test height and width + with self.assertRaises(AssertionError): + prepare_mask_and_masked_image(torch.randn(3, 32, 32), torch.randn(64, 64)) + # test batch dim + with self.assertRaises(AssertionError): + prepare_mask_and_masked_image(torch.randn(2, 3, 32, 32), torch.randn(4, 64, 64)) + # test batch dim + with self.assertRaises(AssertionError): + prepare_mask_and_masked_image(torch.randn(2, 3, 32, 32), torch.randn(4, 1, 64, 64)) + + def test_type_mismatch(self): + # test tensors-only + with self.assertRaises(TypeError): + prepare_mask_and_masked_image(torch.rand(3, 32, 32), torch.rand(3, 32, 32).numpy()) + # test tensors-only + with self.assertRaises(TypeError): + prepare_mask_and_masked_image(torch.rand(3, 32, 32).numpy(), torch.rand(3, 32, 32)) + + def test_channels_first(self): + # test channels first for 3D tensors + with self.assertRaises(AssertionError): + prepare_mask_and_masked_image(torch.rand(32, 32, 3), torch.rand(3, 32, 32)) + + def test_tensor_range(self): + # test im <= 1 + with self.assertRaises(ValueError): + prepare_mask_and_masked_image(torch.ones(3, 32, 32) * 2, torch.rand(32, 32)) + # test im >= -1 + with self.assertRaises(ValueError): + prepare_mask_and_masked_image(torch.ones(3, 32, 32) * (-2), torch.rand(32, 32)) + # test mask <= 1 + with self.assertRaises(ValueError): + prepare_mask_and_masked_image(torch.rand(3, 32, 32), torch.ones(32, 32) * 2) + # test mask >= 0 + with self.assertRaises(ValueError): + prepare_mask_and_masked_image(torch.rand(3, 32, 32), torch.ones(32, 32) * -1) diff --git a/tests/pipelines/stable_diffusion_safe/__init__.py b/tests/pipelines/stable_diffusion_safe/__init__.py new file mode 100644 index 0000000000..e69de29bb2 diff --git a/tests/pipelines/stable_diffusion_safe/test_safe_diffusion.py b/tests/pipelines/stable_diffusion_safe/test_safe_diffusion.py new file mode 100644 index 0000000000..dcb3f27303 --- /dev/null +++ b/tests/pipelines/stable_diffusion_safe/test_safe_diffusion.py @@ -0,0 +1,435 @@ +# coding=utf-8 +# Copyright 2022 HuggingFace Inc. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import gc +import random +import tempfile +import unittest + +import numpy as np +import torch + +from diffusers import AutoencoderKL, DDIMScheduler, LMSDiscreteScheduler, PNDMScheduler, UNet2DConditionModel +from diffusers.pipelines.stable_diffusion_safe import StableDiffusionPipelineSafe as StableDiffusionPipeline +from diffusers.utils import floats_tensor, slow, torch_device +from diffusers.utils.testing_utils import require_torch_gpu +from transformers import CLIPTextConfig, CLIPTextModel, CLIPTokenizer + +from ...test_pipelines_common import PipelineTesterMixin + + +torch.backends.cuda.matmul.allow_tf32 = False + + +class SafeDiffusionPipelineFastTests(PipelineTesterMixin, unittest.TestCase): + def tearDown(self): + # clean up the VRAM after each test + super().tearDown() + gc.collect() + torch.cuda.empty_cache() + + @property + def dummy_image(self): + batch_size = 1 + num_channels = 3 + sizes = (32, 32) + + image = floats_tensor((batch_size, num_channels) + sizes, rng=random.Random(0)).to(torch_device) + return image + + @property + def dummy_cond_unet(self): + torch.manual_seed(0) + model = UNet2DConditionModel( + block_out_channels=(32, 64), + layers_per_block=2, + sample_size=32, + in_channels=4, + out_channels=4, + down_block_types=("DownBlock2D", "CrossAttnDownBlock2D"), + up_block_types=("CrossAttnUpBlock2D", "UpBlock2D"), + cross_attention_dim=32, + ) + return model + + @property + def dummy_vae(self): + torch.manual_seed(0) + model = AutoencoderKL( + block_out_channels=[32, 64], + in_channels=3, + out_channels=3, + down_block_types=["DownEncoderBlock2D", "DownEncoderBlock2D"], + up_block_types=["UpDecoderBlock2D", "UpDecoderBlock2D"], + latent_channels=4, + ) + return model + + @property + def dummy_text_encoder(self): + torch.manual_seed(0) + config = CLIPTextConfig( + bos_token_id=0, + eos_token_id=2, + hidden_size=32, + intermediate_size=37, + layer_norm_eps=1e-05, + num_attention_heads=4, + num_hidden_layers=5, + pad_token_id=1, + vocab_size=1000, + ) + return CLIPTextModel(config) + + @property + def dummy_extractor(self): + def extract(*args, **kwargs): + class Out: + def __init__(self): + self.pixel_values = torch.ones([0]) + + def to(self, device): + self.pixel_values.to(device) + return self + + return Out() + + return extract + + def test_safe_diffusion_ddim(self): + device = "cpu" # ensure determinism for the device-dependent torch.Generator + unet = self.dummy_cond_unet + scheduler = DDIMScheduler( + beta_start=0.00085, + beta_end=0.012, + beta_schedule="scaled_linear", + clip_sample=False, + set_alpha_to_one=False, + ) + + vae = self.dummy_vae + bert = self.dummy_text_encoder + tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip") + + # make sure here that pndm scheduler skips prk + sd_pipe = StableDiffusionPipeline( + unet=unet, + scheduler=scheduler, + vae=vae, + text_encoder=bert, + tokenizer=tokenizer, + safety_checker=None, + feature_extractor=self.dummy_extractor, + ) + sd_pipe = sd_pipe.to(device) + sd_pipe.set_progress_bar_config(disable=None) + + prompt = "A painting of a squirrel eating a burger" + + generator = torch.Generator(device=device).manual_seed(0) + output = sd_pipe([prompt], generator=generator, guidance_scale=6.0, num_inference_steps=2, output_type="np") + image = output.images + + generator = torch.Generator(device=device).manual_seed(0) + image_from_tuple = sd_pipe( + [prompt], + generator=generator, + guidance_scale=6.0, + num_inference_steps=2, + output_type="np", + return_dict=False, + )[0] + + image_slice = image[0, -3:, -3:, -1] + image_from_tuple_slice = image_from_tuple[0, -3:, -3:, -1] + + assert image.shape == (1, 128, 128, 3) + expected_slice = np.array([0.5112, 0.4692, 0.4715, 0.5206, 0.4894, 0.5114, 0.5096, 0.4932, 0.4755]) + + assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2 + assert np.abs(image_from_tuple_slice.flatten() - expected_slice).max() < 1e-2 + + def test_stable_diffusion_pndm(self): + device = "cpu" # ensure determinism for the device-dependent torch.Generator + unet = self.dummy_cond_unet + scheduler = PNDMScheduler(skip_prk_steps=True) + vae = self.dummy_vae + bert = self.dummy_text_encoder + tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip") + + # make sure here that pndm scheduler skips prk + sd_pipe = StableDiffusionPipeline( + unet=unet, + scheduler=scheduler, + vae=vae, + text_encoder=bert, + tokenizer=tokenizer, + safety_checker=None, + feature_extractor=self.dummy_extractor, + ) + sd_pipe = sd_pipe.to(device) + sd_pipe.set_progress_bar_config(disable=None) + + prompt = "A painting of a squirrel eating a burger" + generator = torch.Generator(device=device).manual_seed(0) + output = sd_pipe([prompt], generator=generator, guidance_scale=6.0, num_inference_steps=2, output_type="np") + + image = output.images + + generator = torch.Generator(device=device).manual_seed(0) + image_from_tuple = sd_pipe( + [prompt], + generator=generator, + guidance_scale=6.0, + num_inference_steps=2, + output_type="np", + return_dict=False, + )[0] + + image_slice = image[0, -3:, -3:, -1] + image_from_tuple_slice = image_from_tuple[0, -3:, -3:, -1] + + assert image.shape == (1, 128, 128, 3) + expected_slice = np.array([0.4937, 0.4649, 0.4716, 0.5145, 0.4889, 0.513, 0.513, 0.4905, 0.4738]) + assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2 + assert np.abs(image_from_tuple_slice.flatten() - expected_slice).max() < 1e-2 + + def test_stable_diffusion_no_safety_checker(self): + pipe = StableDiffusionPipeline.from_pretrained( + "hf-internal-testing/tiny-stable-diffusion-lms-pipe", safety_checker=None + ) + assert isinstance(pipe, StableDiffusionPipeline) + assert isinstance(pipe.scheduler, LMSDiscreteScheduler) + assert pipe.safety_checker is None + + image = pipe("example prompt", num_inference_steps=2).images[0] + assert image is not None + + # check that there's no error when saving a pipeline with one of the models being None + with tempfile.TemporaryDirectory() as tmpdirname: + pipe.save_pretrained(tmpdirname) + pipe = StableDiffusionPipeline.from_pretrained(tmpdirname) + + # sanity check that the pipeline still works + assert pipe.safety_checker is None + image = pipe("example prompt", num_inference_steps=2).images[0] + assert image is not None + + @unittest.skipIf(torch_device != "cuda", "This test requires a GPU") + def test_stable_diffusion_fp16(self): + """Test that stable diffusion works with fp16""" + unet = self.dummy_cond_unet + scheduler = PNDMScheduler(skip_prk_steps=True) + vae = self.dummy_vae + bert = self.dummy_text_encoder + tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip") + + # put models in fp16 + unet = unet.half() + vae = vae.half() + bert = bert.half() + + # make sure here that pndm scheduler skips prk + sd_pipe = StableDiffusionPipeline( + unet=unet, + scheduler=scheduler, + vae=vae, + text_encoder=bert, + tokenizer=tokenizer, + safety_checker=None, + feature_extractor=self.dummy_extractor, + ) + sd_pipe = sd_pipe.to(torch_device) + sd_pipe.set_progress_bar_config(disable=None) + + prompt = "A painting of a squirrel eating a burger" + generator = torch.Generator(device=torch_device).manual_seed(0) + image = sd_pipe([prompt], generator=generator, num_inference_steps=2, output_type="np").images + + assert image.shape == (1, 128, 128, 3) + + +@slow +@require_torch_gpu +class SafeDiffusionPipelineIntegrationTests(unittest.TestCase): + def tearDown(self): + # clean up the VRAM after each test + super().tearDown() + gc.collect() + torch.cuda.empty_cache() + + def test_harm_safe_stable_diffusion(self): + sd_pipe = StableDiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5", safety_checker=None) + sd_pipe.scheduler = LMSDiscreteScheduler.from_config(sd_pipe.scheduler.config) + sd_pipe = sd_pipe.to(torch_device) + sd_pipe.set_progress_bar_config(disable=None) + + prompt = ( + "portrait of girl with smokey eyes makeup in abandoned hotel, grange clothes, redshift, wide high angle" + " coloured polaroid photograph with flash, kodak film, hyper real, stunning moody cinematography, with" + " anamorphic lenses, by maripol, fallen angels by wong kar - wai, style of suspiria and neon demon and" + " children from bahnhof zoo, detailed " + ) + seed = 4003660346 + guidance_scale = 7 + + # without safety guidance (sld_guidance_scale = 0) + generator = torch.Generator(device=torch_device).manual_seed(seed) + output = sd_pipe( + [prompt], + generator=generator, + guidance_scale=guidance_scale, + num_inference_steps=50, + output_type="np", + width=512, + height=512, + sld_guidance_scale=0, + ) + + image = output.images + image_slice = image[0, -3:, -3:, -1] + expected_slice = [0.2278, 0.2231, 0.2249, 0.2333, 0.2303, 0.1885, 0.2273, 0.2144, 0.2176] + + assert image.shape == (1, 512, 512, 3) + assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2 + + # without safety guidance (strong configuration) + generator = torch.Generator(device=torch_device).manual_seed(seed) + output = sd_pipe( + [prompt], + generator=generator, + guidance_scale=guidance_scale, + num_inference_steps=50, + output_type="np", + width=512, + height=512, + sld_guidance_scale=2000, + sld_warmup_steps=7, + sld_threshold=0.025, + sld_momentum_scale=0.5, + sld_mom_beta=0.7, + ) + + image = output.images + image_slice = image[0, -3:, -3:, -1] + expected_slice = [0.2383, 0.2276, 0.236, 0.2192, 0.2186, 0.2053, 0.1971, 0.1901, 0.1719] + + assert image.shape == (1, 512, 512, 3) + assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2 + + def test_nudity_safe_stable_diffusion(self): + sd_pipe = StableDiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5", safety_checker=None) + sd_pipe.scheduler = LMSDiscreteScheduler.from_config(sd_pipe.scheduler.config) + sd_pipe = sd_pipe.to(torch_device) + sd_pipe.set_progress_bar_config(disable=None) + + prompt = "padme amidala taking a bath artwork, safe for work, no nudity" + seed = 2734971755 + guidance_scale = 7 + + generator = torch.Generator(device=torch_device).manual_seed(seed) + output = sd_pipe( + [prompt], + generator=generator, + guidance_scale=guidance_scale, + num_inference_steps=50, + output_type="np", + width=512, + height=512, + sld_guidance_scale=0, + ) + + image = output.images + image_slice = image[0, -3:, -3:, -1] + expected_slice = [0.3502, 0.3622, 0.3396, 0.3642, 0.3478, 0.3318, 0.35, 0.3348, 0.3297] + + assert image.shape == (1, 512, 512, 3) + assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2 + + generator = torch.Generator(device=torch_device).manual_seed(seed) + output = sd_pipe( + [prompt], + generator=generator, + guidance_scale=guidance_scale, + num_inference_steps=50, + output_type="np", + width=512, + height=512, + sld_guidance_scale=2000, + sld_warmup_steps=7, + sld_threshold=0.025, + sld_momentum_scale=0.5, + sld_mom_beta=0.7, + ) + + image = output.images + image_slice = image[0, -3:, -3:, -1] + expected_slice = [0.5531, 0.5206, 0.4895, 0.5156, 0.5182, 0.4751, 0.4802, 0.4803, 0.4443] + + assert image.shape == (1, 512, 512, 3) + assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2 + + def test_nudity_safetychecker_safe_stable_diffusion(self): + sd_pipe = StableDiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5") + sd_pipe = sd_pipe.to(torch_device) + sd_pipe.set_progress_bar_config(disable=None) + + prompt = ( + "the four horsewomen of the apocalypse, painting by tom of finland, gaston bussiere, craig mullins, j. c." + " leyendecker" + ) + seed = 1044355234 + guidance_scale = 12 + + generator = torch.Generator(device=torch_device).manual_seed(seed) + output = sd_pipe( + [prompt], + generator=generator, + guidance_scale=guidance_scale, + num_inference_steps=50, + output_type="np", + width=512, + height=512, + sld_guidance_scale=0, + ) + + image = output.images + image_slice = image[0, -3:, -3:, -1] + expected_slice = np.array([0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0]) + + assert image.shape == (1, 512, 512, 3) + assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-7 + + generator = torch.Generator(device=torch_device).manual_seed(seed) + output = sd_pipe( + [prompt], + generator=generator, + guidance_scale=guidance_scale, + num_inference_steps=50, + output_type="np", + width=512, + height=512, + sld_guidance_scale=2000, + sld_warmup_steps=7, + sld_threshold=0.025, + sld_momentum_scale=0.5, + sld_mom_beta=0.7, + ) + + image = output.images + image_slice = image[0, -3:, -3:, -1] + expected_slice = np.array([0.5818, 0.6285, 0.6835, 0.6019, 0.625, 0.6754, 0.6096, 0.6334, 0.6561]) + assert image.shape == (1, 512, 512, 3) + assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2 diff --git a/tests/pipelines/versatile_diffusion/__init__.py b/tests/pipelines/versatile_diffusion/__init__.py new file mode 100644 index 0000000000..e69de29bb2 diff --git a/tests/pipelines/versatile_diffusion/test_versatile_diffusion_dual_guided.py b/tests/pipelines/versatile_diffusion/test_versatile_diffusion_dual_guided.py new file mode 100644 index 0000000000..9fb6ca522f --- /dev/null +++ b/tests/pipelines/versatile_diffusion/test_versatile_diffusion_dual_guided.py @@ -0,0 +1,112 @@ +# coding=utf-8 +# Copyright 2022 HuggingFace Inc. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import gc +import tempfile +import unittest + +import numpy as np +import torch + +from diffusers import VersatileDiffusionDualGuidedPipeline +from diffusers.utils.testing_utils import load_image, require_torch_gpu, slow, torch_device + +from ...test_pipelines_common import PipelineTesterMixin + + +torch.backends.cuda.matmul.allow_tf32 = False + + +class VersatileDiffusionDualGuidedPipelineFastTests(PipelineTesterMixin, unittest.TestCase): + pass + + +@slow +@require_torch_gpu +class VersatileDiffusionDualGuidedPipelineIntegrationTests(unittest.TestCase): + def tearDown(self): + # clean up the VRAM after each test + super().tearDown() + gc.collect() + torch.cuda.empty_cache() + + def test_remove_unused_weights_save_load(self): + pipe = VersatileDiffusionDualGuidedPipeline.from_pretrained("shi-labs/versatile-diffusion") + # remove text_unet + pipe.remove_unused_weights() + pipe.to(torch_device) + pipe.set_progress_bar_config(disable=None) + + second_prompt = load_image( + "https://raw.githubusercontent.com/SHI-Labs/Versatile-Diffusion/master/assets/benz.jpg" + ) + + generator = torch.Generator(device=torch_device).manual_seed(0) + image = pipe( + prompt="first prompt", + image=second_prompt, + text_to_image_strength=0.75, + generator=generator, + guidance_scale=7.5, + num_inference_steps=2, + output_type="numpy", + ).images + + with tempfile.TemporaryDirectory() as tmpdirname: + pipe.save_pretrained(tmpdirname) + pipe = VersatileDiffusionDualGuidedPipeline.from_pretrained(tmpdirname) + + pipe.to(torch_device) + pipe.set_progress_bar_config(disable=None) + + generator = generator.manual_seed(0) + new_image = pipe( + prompt="first prompt", + image=second_prompt, + text_to_image_strength=0.75, + generator=generator, + guidance_scale=7.5, + num_inference_steps=2, + output_type="numpy", + ).images + + assert np.abs(image - new_image).sum() < 1e-5, "Models don't have the same forward pass" + + def test_inference_dual_guided(self): + pipe = VersatileDiffusionDualGuidedPipeline.from_pretrained("shi-labs/versatile-diffusion") + pipe.remove_unused_weights() + pipe.to(torch_device) + pipe.set_progress_bar_config(disable=None) + + first_prompt = "cyberpunk 2077" + second_prompt = load_image( + "https://raw.githubusercontent.com/SHI-Labs/Versatile-Diffusion/master/assets/benz.jpg" + ) + generator = torch.Generator(device=torch_device).manual_seed(0) + image = pipe( + prompt=first_prompt, + image=second_prompt, + text_to_image_strength=0.75, + generator=generator, + guidance_scale=7.5, + num_inference_steps=50, + output_type="numpy", + ).images + + image_slice = image[0, 253:256, 253:256, -1] + + assert image.shape == (1, 512, 512, 3) + expected_slice = np.array([0.014, 0.0112, 0.0136, 0.0145, 0.0107, 0.0113, 0.0272, 0.0215, 0.0216]) + assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2 diff --git a/tests/pipelines/versatile_diffusion/test_versatile_diffusion_image_variation.py b/tests/pipelines/versatile_diffusion/test_versatile_diffusion_image_variation.py new file mode 100644 index 0000000000..4eddc271db --- /dev/null +++ b/tests/pipelines/versatile_diffusion/test_versatile_diffusion_image_variation.py @@ -0,0 +1,58 @@ +# coding=utf-8 +# Copyright 2022 HuggingFace Inc. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import unittest + +import numpy as np +import torch + +from diffusers import VersatileDiffusionImageVariationPipeline +from diffusers.utils.testing_utils import load_image, require_torch_gpu, slow, torch_device + +from ...test_pipelines_common import PipelineTesterMixin + + +torch.backends.cuda.matmul.allow_tf32 = False + + +class VersatileDiffusionImageVariationPipelineFastTests(PipelineTesterMixin, unittest.TestCase): + pass + + +@slow +@require_torch_gpu +class VersatileDiffusionImageVariationPipelineIntegrationTests(unittest.TestCase): + def test_inference_image_variations(self): + pipe = VersatileDiffusionImageVariationPipeline.from_pretrained("shi-labs/versatile-diffusion") + pipe.to(torch_device) + pipe.set_progress_bar_config(disable=None) + + image_prompt = load_image( + "https://raw.githubusercontent.com/SHI-Labs/Versatile-Diffusion/master/assets/benz.jpg" + ) + generator = torch.Generator(device=torch_device).manual_seed(0) + image = pipe( + image=image_prompt, + generator=generator, + guidance_scale=7.5, + num_inference_steps=50, + output_type="numpy", + ).images + + image_slice = image[0, 253:256, 253:256, -1] + + assert image.shape == (1, 512, 512, 3) + expected_slice = np.array([0.0113, 0.2241, 0.4024, 0.0839, 0.0871, 0.2725, 0.2581, 0.0, 0.1096]) + assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2 diff --git a/tests/pipelines/versatile_diffusion/test_versatile_diffusion_mega.py b/tests/pipelines/versatile_diffusion/test_versatile_diffusion_mega.py new file mode 100644 index 0000000000..1209abf6a8 --- /dev/null +++ b/tests/pipelines/versatile_diffusion/test_versatile_diffusion_mega.py @@ -0,0 +1,129 @@ +# coding=utf-8 +# Copyright 2022 HuggingFace Inc. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import gc +import tempfile +import unittest + +import numpy as np +import torch + +from diffusers import VersatileDiffusionPipeline +from diffusers.utils.testing_utils import load_image, require_torch_gpu, slow, torch_device + +from ...test_pipelines_common import PipelineTesterMixin + + +torch.backends.cuda.matmul.allow_tf32 = False + + +class VersatileDiffusionMegaPipelineFastTests(PipelineTesterMixin, unittest.TestCase): + pass + + +@slow +@require_torch_gpu +class VersatileDiffusionMegaPipelineIntegrationTests(unittest.TestCase): + def tearDown(self): + # clean up the VRAM after each test + super().tearDown() + gc.collect() + torch.cuda.empty_cache() + + def test_from_pretrained_save_pretrained(self): + pipe = VersatileDiffusionPipeline.from_pretrained("shi-labs/versatile-diffusion", torch_dtype=torch.float16) + pipe.to(torch_device) + pipe.set_progress_bar_config(disable=None) + + prompt_image = load_image( + "https://raw.githubusercontent.com/SHI-Labs/Versatile-Diffusion/master/assets/benz.jpg" + ) + + generator = torch.Generator(device=torch_device).manual_seed(0) + image = pipe.dual_guided( + prompt="first prompt", + image=prompt_image, + text_to_image_strength=0.75, + generator=generator, + guidance_scale=7.5, + num_inference_steps=2, + output_type="numpy", + ).images + + with tempfile.TemporaryDirectory() as tmpdirname: + pipe.save_pretrained(tmpdirname) + pipe = VersatileDiffusionPipeline.from_pretrained(tmpdirname, torch_dtype=torch.float16) + pipe.to(torch_device) + pipe.set_progress_bar_config(disable=None) + + generator = generator.manual_seed(0) + new_image = pipe.dual_guided( + prompt="first prompt", + image=prompt_image, + text_to_image_strength=0.75, + generator=generator, + guidance_scale=7.5, + num_inference_steps=2, + output_type="numpy", + ).images + + assert np.abs(image - new_image).sum() < 1e-5, "Models don't have the same forward pass" + + def test_inference_dual_guided_then_text_to_image(self): + pipe = VersatileDiffusionPipeline.from_pretrained("shi-labs/versatile-diffusion", torch_dtype=torch.float16) + pipe.to(torch_device) + pipe.set_progress_bar_config(disable=None) + + prompt = "cyberpunk 2077" + init_image = load_image( + "https://raw.githubusercontent.com/SHI-Labs/Versatile-Diffusion/master/assets/benz.jpg" + ) + generator = torch.Generator(device=torch_device).manual_seed(0) + image = pipe.dual_guided( + prompt=prompt, + image=init_image, + text_to_image_strength=0.75, + generator=generator, + guidance_scale=7.5, + num_inference_steps=50, + output_type="numpy", + ).images + + image_slice = image[0, 253:256, 253:256, -1] + + assert image.shape == (1, 512, 512, 3) + expected_slice = np.array([0.014, 0.0112, 0.0136, 0.0145, 0.0107, 0.0113, 0.0272, 0.0215, 0.0216]) + assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2 + + prompt = "A painting of a squirrel eating a burger " + generator = torch.Generator(device=torch_device).manual_seed(0) + image = pipe.text_to_image( + prompt=prompt, generator=generator, guidance_scale=7.5, num_inference_steps=50, output_type="numpy" + ).images + + image_slice = image[0, 253:256, 253:256, -1] + + assert image.shape == (1, 512, 512, 3) + expected_slice = np.array([0.0408, 0.0181, 0.0, 0.0388, 0.0046, 0.0461, 0.0411, 0.0, 0.0222]) + assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2 + + pipe = VersatileDiffusionPipeline.from_pretrained("shi-labs/versatile-diffusion", torch_dtype=torch.float16) + image = pipe.image_variation(init_image, generator=generator, output_type="numpy").images[0] + + image_slice = image[0, 253:256, 253:256, -1] + + assert image.shape == (1, 512, 512, 3) + expected_slice = np.array([0.0657, 0.0529, 0.0455, 0.0802, 0.0570, 0.0179, 0.0267, 0.0483, 0.0769]) + assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2 diff --git a/tests/pipelines/versatile_diffusion/test_versatile_diffusion_text_to_image.py b/tests/pipelines/versatile_diffusion/test_versatile_diffusion_text_to_image.py new file mode 100644 index 0000000000..027819efee --- /dev/null +++ b/tests/pipelines/versatile_diffusion/test_versatile_diffusion_text_to_image.py @@ -0,0 +1,86 @@ +# coding=utf-8 +# Copyright 2022 HuggingFace Inc. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import gc +import tempfile +import unittest + +import numpy as np +import torch + +from diffusers import VersatileDiffusionTextToImagePipeline +from diffusers.utils.testing_utils import require_torch_gpu, slow, torch_device + +from ...test_pipelines_common import PipelineTesterMixin + + +torch.backends.cuda.matmul.allow_tf32 = False + + +class VersatileDiffusionTextToImagePipelineFastTests(PipelineTesterMixin, unittest.TestCase): + pass + + +@slow +@require_torch_gpu +class VersatileDiffusionTextToImagePipelineIntegrationTests(unittest.TestCase): + def tearDown(self): + # clean up the VRAM after each test + super().tearDown() + gc.collect() + torch.cuda.empty_cache() + + def test_remove_unused_weights_save_load(self): + pipe = VersatileDiffusionTextToImagePipeline.from_pretrained("shi-labs/versatile-diffusion") + # remove text_unet + pipe.remove_unused_weights() + pipe.to(torch_device) + pipe.set_progress_bar_config(disable=None) + + prompt = "A painting of a squirrel eating a burger " + generator = torch.Generator(device=torch_device).manual_seed(0) + image = pipe( + prompt=prompt, generator=generator, guidance_scale=7.5, num_inference_steps=2, output_type="numpy" + ).images + + with tempfile.TemporaryDirectory() as tmpdirname: + pipe.save_pretrained(tmpdirname) + pipe = VersatileDiffusionTextToImagePipeline.from_pretrained(tmpdirname) + pipe.to(torch_device) + pipe.set_progress_bar_config(disable=None) + + generator = generator.manual_seed(0) + new_image = pipe( + prompt=prompt, generator=generator, guidance_scale=7.5, num_inference_steps=2, output_type="numpy" + ).images + + assert np.abs(image - new_image).sum() < 1e-5, "Models don't have the same forward pass" + + def test_inference_text2img(self): + pipe = VersatileDiffusionTextToImagePipeline.from_pretrained("shi-labs/versatile-diffusion") + pipe.to(torch_device) + pipe.set_progress_bar_config(disable=None) + + prompt = "A painting of a squirrel eating a burger " + generator = torch.Generator(device=torch_device).manual_seed(0) + image = pipe( + prompt=prompt, generator=generator, guidance_scale=7.5, num_inference_steps=50, output_type="numpy" + ).images + + image_slice = image[0, 253:256, 253:256, -1] + + assert image.shape == (1, 512, 512, 3) + expected_slice = np.array([0.0408, 0.0181, 0.0, 0.0388, 0.0046, 0.0461, 0.0411, 0.0, 0.0222]) + assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2 diff --git a/tests/test_pipelines.py b/tests/test_pipelines.py index b593791c94..19493e3231 100644 --- a/tests/test_pipelines.py +++ b/tests/test_pipelines.py @@ -18,6 +18,7 @@ import os import random import tempfile import unittest +from functools import partial import numpy as np import torch @@ -46,6 +47,7 @@ from diffusers.pipeline_utils import DiffusionPipeline from diffusers.schedulers.scheduling_utils import SCHEDULER_CONFIG_NAME from diffusers.utils import CONFIG_NAME, WEIGHTS_NAME, floats_tensor, slow, torch_device from diffusers.utils.testing_utils import CaptureLogger, get_tests_dir, require_torch_gpu +from parameterized import parameterized from PIL import Image from transformers import CLIPFeatureExtractor, CLIPModel, CLIPTextConfig, CLIPTextModel, CLIPTokenizer @@ -247,7 +249,6 @@ class CustomPipelineTests(unittest.TestCase): class PipelineFastTests(unittest.TestCase): - @property def dummy_image(self): batch_size = 1 num_channels = 3 @@ -256,13 +257,12 @@ class PipelineFastTests(unittest.TestCase): image = floats_tensor((batch_size, num_channels) + sizes, rng=random.Random(0)).to(torch_device) return image - @property - def dummy_uncond_unet(self): + def dummy_uncond_unet(self, sample_size=32): torch.manual_seed(0) model = UNet2DModel( block_out_channels=(32, 64), layers_per_block=2, - sample_size=32, + sample_size=sample_size, in_channels=3, out_channels=3, down_block_types=("DownBlock2D", "AttnDownBlock2D"), @@ -270,13 +270,12 @@ class PipelineFastTests(unittest.TestCase): ) return model - @property - def dummy_cond_unet(self): + def dummy_cond_unet(self, sample_size=32): torch.manual_seed(0) model = UNet2DConditionModel( block_out_channels=(32, 64), layers_per_block=2, - sample_size=32, + sample_size=sample_size, in_channels=4, out_channels=4, down_block_types=("DownBlock2D", "CrossAttnDownBlock2D"), @@ -285,13 +284,12 @@ class PipelineFastTests(unittest.TestCase): ) return model - @property - def dummy_cond_unet_inpaint(self): + def dummy_cond_unet_inpaint(self, sample_size=32): torch.manual_seed(0) model = UNet2DConditionModel( block_out_channels=(32, 64), layers_per_block=2, - sample_size=32, + sample_size=sample_size, in_channels=9, out_channels=4, down_block_types=("DownBlock2D", "CrossAttnDownBlock2D"), @@ -300,7 +298,6 @@ class PipelineFastTests(unittest.TestCase): ) return model - @property def dummy_vq_model(self): torch.manual_seed(0) model = VQModel( @@ -313,7 +310,6 @@ class PipelineFastTests(unittest.TestCase): ) return model - @property def dummy_vae(self): torch.manual_seed(0) model = AutoencoderKL( @@ -326,7 +322,6 @@ class PipelineFastTests(unittest.TestCase): ) return model - @property def dummy_text_encoder(self): torch.manual_seed(0) config = CLIPTextConfig( @@ -342,7 +337,6 @@ class PipelineFastTests(unittest.TestCase): ) return CLIPTextModel(config) - @property def dummy_extractor(self): def extract(*args, **kwargs): class Out: @@ -357,15 +351,43 @@ class PipelineFastTests(unittest.TestCase): return extract - def test_components(self): + @parameterized.expand( + [ + [DDIMScheduler, DDIMPipeline, 32], + [partial(DDPMScheduler, predict_epsilon=True), DDPMPipeline, 32], + [DDIMScheduler, DDIMPipeline, (32, 64)], + [partial(DDPMScheduler, predict_epsilon=True), DDPMPipeline, (64, 32)], + ] + ) + def test_uncond_unet_components(self, scheduler_fn=DDPMScheduler, pipeline_fn=DDPMPipeline, sample_size=32): + unet = self.dummy_uncond_unet(sample_size) + # DDIM doesn't take `predict_epsilon`, and DDPM requires it -- so using partial in parameterized decorator + scheduler = scheduler_fn() + pipeline = pipeline_fn(unet, scheduler).to(torch_device) + + # Device type MPS is not supported for torch.Generator() api. + if torch_device == "mps": + generator = torch.manual_seed(0) + else: + generator = torch.Generator(device=torch_device).manual_seed(0) + + out_image = pipeline( + generator=generator, + num_inference_steps=2, + output_type="np", + ).images + sample_size = (sample_size, sample_size) if isinstance(sample_size, int) else sample_size + assert out_image.shape == (1, *sample_size, 3) + + def test_stable_diffusion_components(self): """Test that components property works correctly""" - unet = self.dummy_cond_unet + unet = self.dummy_cond_unet() scheduler = PNDMScheduler(skip_prk_steps=True) - vae = self.dummy_vae - bert = self.dummy_text_encoder + vae = self.dummy_vae() + bert = self.dummy_text_encoder() tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip") - image = self.dummy_image.cpu().permute(0, 2, 3, 1)[0] + image = self.dummy_image().cpu().permute(0, 2, 3, 1)[0] init_image = Image.fromarray(np.uint8(image)).convert("RGB") mask_image = Image.fromarray(np.uint8(image + 4)).convert("RGB").resize((128, 128)) @@ -377,7 +399,7 @@ class PipelineFastTests(unittest.TestCase): text_encoder=bert, tokenizer=tokenizer, safety_checker=None, - feature_extractor=self.dummy_extractor, + feature_extractor=self.dummy_extractor(), ).to(torch_device) img2img = StableDiffusionImg2ImgPipeline(**inpaint.components).to(torch_device) text2img = StableDiffusionPipeline(**inpaint.components).to(torch_device) diff --git a/utils/check_copies.py b/utils/check_copies.py index 395cefb9c4..16782397da 100644 --- a/utils/check_copies.py +++ b/utils/check_copies.py @@ -153,6 +153,10 @@ def is_copy_consistent(filename, overwrite=False): observed_code_lines = lines[start_index:line_index] observed_code = "".join(observed_code_lines) + # Remove any nested `Copied from` comments to avoid circular copies + theoretical_code = [line for line in theoretical_code.split("\n") if _re_copy_warning.search(line) is None] + theoretical_code = "\n".join(theoretical_code) + # Before comparing, use the `replace_pattern` on the original code. if len(replace_pattern) > 0: patterns = replace_pattern.replace("with", "").split(",")