add gpt2

scripts/convert_versatile_diffusion_to_diffusers.py

@@ -31,8 +31,12 @@ from diffusers import (
     UNet2DConditionModel,
     VersatileDiffusionPipeline,
 )
-from diffusers.pipelines.latent_diffusion.pipeline_latent_diffusion import LDMBertConfig, LDMBertModel
-from transformers import CLIPProcessor, CLIPTextModelWithProjection, CLIPTokenizer, CLIPVisionModelWithProjection
+from transformers import (
+    CLIPFeatureExtractor,
+    CLIPTextModelWithProjection,
+    CLIPTokenizer,
+    CLIPVisionModelWithProjection,
+)
 
 
 SCHEDULER_CONFIG = Namespace(
@@ -334,7 +338,7 @@ def convert_vd_unet_checkpoint(checkpoint, config, unet_key, extract_ema=False):
 
     # 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(f"Checkpoint has both EMA and non-EMA weights.")
+        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"
@@ -610,13 +614,6 @@ if __name__ == "__main__":
     parser.add_argument(
         "--optimus_checkpoint_path", default=None, type=str, required=False, help="Path to the checkpoint to convert."
     )
-    # !wget https://raw.githubusercontent.com/CompVis/stable-diffusion/main/configs/stable-diffusion/v1-inference.yaml
-    parser.add_argument(
-        "--original_config_file",
-        default=None,
-        type=str,
-        help="The YAML config file corresponding to the original architecture.",
-    )
     parser.add_argument(
         "--scheduler_type",
         default="pndm",
@@ -719,14 +716,14 @@ if __name__ == "__main__":
         vae.load_state_dict(converted_vae_checkpoint)
 
     tokenizer = CLIPTokenizer.from_pretrained("openai/clip-vit-large-patch14")
-    image_processor = CLIPProcessor.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_processor=image_processor,
+        image_feature_extractor=image_feature_extractor,
         text_encoder=text_encoder,
         image_encoder=image_encoder,
         image_unet=image_unet,

src/diffusers/__init__.py

@@ -73,7 +73,9 @@ if is_torch_available() and is_transformers_available():
         StableDiffusionInpaintPipeline,
         StableDiffusionInpaintPipelineLegacy,
         StableDiffusionPipeline,
+        VersatileDiffusionImageVariationPipeline,
         VersatileDiffusionPipeline,
+        VersatileDiffusionTextToImagePipeline,
         VQDiffusionPipeline,
     )
 else:

src/diffusers/pipelines/__init__.py

@@ -24,7 +24,11 @@ if is_torch_available() and is_transformers_available():
         StableDiffusionInpaintPipelineLegacy,
         StableDiffusionPipeline,
     )
-    from .versatile_diffusion import VersatileDiffusionPipeline
+    from .versatile_diffusion import (
+        VersatileDiffusionImageVariationPipeline,
+        VersatileDiffusionPipeline,
+        VersatileDiffusionTextToImagePipeline,
+    )
     from .vq_diffusion import VQDiffusionPipeline
 
 if is_transformers_available() and is_onnx_available():

src/diffusers/pipelines/versatile_diffusion/__init__.py

@@ -4,3 +4,5 @@ from ...utils import is_torch_available, is_transformers_available
 if is_transformers_available() and is_torch_available():
     from .modeling_gpt2_optimus import GPT2OptimusForLatentConnector
     from .pipeline_versatile_diffusion import VersatileDiffusionPipeline
+    from .pipeline_versatile_diffusion_image_variation import VersatileDiffusionImageVariationPipeline
+    from .pipeline_versatile_diffusion_text_to_image import VersatileDiffusionTextToImagePipeline

src/diffusers/pipelines/versatile_diffusion/pipeline_versatile_diffusion.py

@@ -1,108 +1,73 @@
-# 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.
+from typing import Any, Callable, Dict, List, Optional, Union
 
-import inspect
-from typing import List, Optional, Tuple, Union
-
-import numpy as np
 import torch
-import torch.utils.checkpoint
 
-import PIL
-from transformers import CLIPProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModel
+import PIL.Image
+from transformers import CLIPFeatureExtractor, CLIPTextModel, CLIPTokenizer, CLIPVisionModel
 
-from ...models import AutoencoderKL, UNet2DConditionModel, VQModel
-from ...models.attention import Transformer2DModel
-from ...pipeline_utils import DiffusionPipeline, ImagePipelineOutput
+from ...models import AutoencoderKL, UNet2DConditionModel
+from ...pipeline_utils import DiffusionPipeline
 from ...schedulers import DDIMScheduler, LMSDiscreteScheduler, PNDMScheduler
+from ...utils import logging
+from . import VersatileDiffusionImageVariationPipeline, VersatileDiffusionTextToImagePipeline
 
 
-class VersatileMixedModel:
-    """
-    A context managet that swaps the transformer modules between the image and text unet during inference, depending on
-    the latent type and condition type.
-    """
-
-    def __init__(self, image_unet, text_unet, latent_type, condition_type):
-        self.image_unet = image_unet
-        self.text_unet = text_unet
-        self.latent_type = latent_type
-        self.condition_type = condition_type
-
-    def swap_transformer_modules(self):
-        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 __enter__(self):
-        if self.latent_type != self.condition_type:
-            self.swap_transformer_modules()
-        return self.image_unet if self.latent_type == "image" else self.text_unet
-
-    def __exit__(self, *exc):
-        # swap the modules back
-        if self.latent_type != self.condition_type:
-            self.swap_transformer_modules()
+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.)
 
-    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`):
+    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
-            [BertTokenizer](https://huggingface.co/docs/transformers/model_doc/bert#transformers.BertTokenizer).
+            [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_processor: CLIPProcessor
+    image_feature_extractor: CLIPFeatureExtractor
     text_encoder: CLIPTextModel
     image_encoder: CLIPVisionModel
     image_unet: UNet2DConditionModel
     text_unet: UNet2DConditionModel
-    vae: Union[VQModel, AutoencoderKL]
+    vae: AutoencoderKL
     scheduler: Union[DDIMScheduler, PNDMScheduler, LMSDiscreteScheduler]
 
     def __init__(
         self,
         tokenizer: CLIPTokenizer,
-        image_processor: CLIPProcessor,
+        image_feature_extractor: CLIPFeatureExtractor,
         text_encoder: CLIPTextModel,
         image_encoder: CLIPVisionModel,
         image_unet: UNet2DConditionModel,
         text_unet: UNet2DConditionModel,
-        vae: Union[VQModel, AutoencoderKL],
+        vae: AutoencoderKL,
         scheduler: Union[DDIMScheduler, PNDMScheduler, LMSDiscreteScheduler],
     ):
         super().__init__()
+
         self.register_modules(
             tokenizer=tokenizer,
-            image_processor=image_processor,
+            image_feature_extractor=image_feature_extractor,
             text_encoder=text_encoder,
             image_encoder=image_encoder,
             image_unet=image_unet,
@@ -110,203 +75,105 @@ class VersatileDiffusionPipeline(DiffusionPipeline):
             vae=vae,
             scheduler=scheduler,
         )
-        self.image_transformer_blocks = {
-            name: module for name, module in image_unet.named_modules() if isinstance(module, Transformer2DModel)
-        }
-        self.text_transformer_blocks = {
-            name: module for name, module in text_unet.named_modules() if isinstance(module, Transformer2DModel)
-        }
 
-    def _encode_prompt(self, prompt, do_classifier_free_guidance):
+    @property
+    def components(self) -> Dict[str, Any]:
+        return {k: getattr(self, k) for k in self.config.keys() if not k.startswith("_")}
+
+    def enable_attention_slicing(self, slice_size: Optional[Union[str, int]] = "auto"):
         r"""
-        Encodes the prompt into text encoder hidden states.
+        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:
-            prompt (`str` or `list(int)`):
-                prompt to be encoded
-            do_classifier_free_guidance (`bool`):
-                whether to use classifier free guidance or not
+            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 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
-
-        if do_classifier_free_guidance:
-            uncond_input = self.tokenizer([""] * batch_size, padding="max_length", max_length=77, return_tensors="pt")
-            uncond_embeddings = self.text_encoder(uncond_input.input_ids.to(self.device))
-            uncond_embeddings = normalize_embeddings(uncond_embeddings)
-
-        # get prompt text embeddings
-        text_input = self.tokenizer(prompt, padding="max_length", max_length=77, return_tensors="pt")
-        text_embeddings = self.text_encoder(text_input.input_ids.to(self.device))
-        text_embeddings = normalize_embeddings(text_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
-        text_embeddings = torch.cat([uncond_embeddings, text_embeddings])
-
-        return text_embeddings
-
-    def _encode_image_prompt(self, prompt, do_classifier_free_guidance):
+    def disable_attention_slicing(self):
         r"""
-        Encodes the image prompt into image encoder hidden states.
-
-        Args:
-            prompt (`str` or `list(int)`):
-                prompt to be encoded
-            do_classifier_free_guidance (`bool`):
-                whether to use classifier free guidance or not
+        Disable sliced attention computation. If `enable_attention_slicing` was previously invoked, this method will go
+        back to computing attention in one step.
         """
-
-        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
-
-        if do_classifier_free_guidance:
-            dummy_images = [np.zeros((512, 512, 3))] * batch_size
-            dummy_images = self.image_processor(images=dummy_images, return_tensors="pt")
-            uncond_embeddings = self.image_encoder(dummy_images.pixel_values.to(self.device))
-            uncond_embeddings = normalize_embeddings(uncond_embeddings)
-
-        # get prompt text embeddings
-        image_input = self.image_processor(images=prompt, return_tensors="pt")
-        image_embeddings = self.image_encoder(image_input.pixel_values.to(self.device))
-        image_embeddings = normalize_embeddings(image_embeddings)
-
-        # For classifier free guidance, we need to do two forward passes.
-        # Here we concatenate the unconditional and image embeddings into a single batch
-        # to avoid doing two forward passes
-        image_embeddings = torch.cat([uncond_embeddings, image_embeddings])
-
-        return image_embeddings
+        # set slice_size = `None` to disable `attention slicing`
+        self.enable_attention_slicing(None)
 
     @torch.no_grad()
-    def __call__(
+    def image_variation(
         self,
-        prompt: Optional[Union[str, List[str]]] = None,
-        image_prompt: Optional[Union[torch.Tensor, PIL.Image.Image]] = None,
-        height: Optional[int] = 512,
-        width: Optional[int] = 512,
-        num_inference_steps: Optional[int] = 50,
-        guidance_scale: Optional[float] = 1.0,
-        eta: Optional[float] = 0.0,
+        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,
-        **kwargs,
-    ) -> Union[Tuple, ImagePipelineOutput]:
-        r"""
-        Args:
-            prompt (`str` or `List[str]`):
-                The prompt or prompts to guide the image generation.
-            height (`int`, *optional*, defaults to 256):
-                The height in pixels of the generated image.
-            width (`int`, *optional*, defaults to 256):
-                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 1.0):
-                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` at
-                the, usually at the expense of lower image quality.
-            generator (`torch.Generator`, *optional*):
-                A [torch generator](https://pytorch.org/docs/stable/generated/torch.Generator.html) 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*):
-                Whether or not to return a [`~pipeline_utils.ImagePipelineOutput`] instead of a plain tuple.
-
-        Returns:
-            [`~pipeline_utils.ImagePipelineOutput`] or `tuple`: [`~pipelines.utils.ImagePipelineOutput`] if
-            `return_dict` is True, otherwise a `tuple. When returning a tuple, the first element is a list with the
-            generated images.
-        """
-        do_classifier_free_guidance = guidance_scale > 1.0
-
-        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 prompt is not None:
-            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)}")
-
-            condition_embeddings = self._encode_prompt(prompt, do_classifier_free_guidance)
-            prompt_type = "text"
-        elif image_prompt is not None:
-            if isinstance(image_prompt, PIL.Image.Image):
-                batch_size = 1
-            elif isinstance(image_prompt, torch.Tensor):
-                batch_size = image_prompt.shape[0]
-            else:
-                raise ValueError(
-                    f"`image_prompt` has to be of type `PIL.Image.Image` or `torch.Tensor` but is {type(image_prompt)}"
-                )
-
-            condition_embeddings = self._encode_image_prompt(image_prompt, do_classifier_free_guidance)
-            prompt_type = "image"
-        else:
-            raise ValueError("Either `prompt` or `image_prompt` has to be provided.")
-
-        latents = torch.randn(
-            (batch_size, self.image_unet.in_channels, height // 8, width // 8), generator=generator, device=self.device
+        callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+        callback_steps: Optional[int] = 1,
+    ):
+        return VersatileDiffusionImageVariationPipeline(**self.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,
         )
 
-        self.scheduler.set_timesteps(num_inference_steps)
-
-        # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
-        accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
-
-        extra_kwargs = {}
-        if accepts_eta:
-            extra_kwargs["eta"] = eta
-
-        with VersatileMixedModel(self.image_unet, self.text_unet, "image", prompt_type) as unet:
-            for t in self.progress_bar(self.scheduler.timesteps):
-                if not do_classifier_free_guidance:
-                    latents_input = latents
-                else:
-                    latents_input = torch.cat([latents] * 2)
-
-                # predict the noise residual
-                noise_pred = unet(latents_input, t, encoder_hidden_states=condition_embeddings).sample
-                # perform guidance
-                if guidance_scale != 1.0:
-                    noise_pred_uncond, noise_prediction_cond = noise_pred.chunk(2)
-                    noise_pred = noise_pred_uncond + guidance_scale * (noise_prediction_cond - noise_pred_uncond)
-
-                # compute the previous noisy sample x_t -> x_t-1
-                latents = self.scheduler.step(noise_pred, t, latents, **extra_kwargs).prev_sample
-
-        # scale and decode the image latents with vae
-        latents = 1 / 0.18215 * latents
-        image = self.vae.decode(latents).sample
-
-        image = (image / 2 + 0.5).clamp(0, 1)
-        image = image.cpu().permute(0, 2, 3, 1).numpy()
-        if output_type == "pil":
-            image = self.numpy_to_pil(image)
-
-        if not return_dict:
-            return (image,)
-
-        return ImagePipelineOutput(images=image)
+    @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,
+    ):
+        return VersatileDiffusionTextToImagePipeline(**self.components)(
+            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,
+        )

src/diffusers/pipelines/versatile_diffusion/pipeline_versatile_diffusion_image_to_text.py

@@ -0,0 +1,408 @@
+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.embeddings import TimestepEmbedding, Timesteps
+from ...models.unet_2d_condition import UNet2DConditionOutput
+from ...utils import logging
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+class UNetMultiDimConditionModel(ModelMixin, ConfigMixin):
+    r"""
+    UNet2DConditionModel 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`, *optional*): The size of the input 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`):
+            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")`):
+            The tuple of downsample blocks to use.
+        up_block_types (`Tuple[str]`, *optional*, defaults to `("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D",)`):
+            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] = (
+            "CrossAttnDownBlockMultiDim",
+            "CrossAttnDownBlockMultiDim",
+            "CrossAttnDownBlockMultiDim",
+            "DownBlockMultiDim",
+        ),
+        up_block_types: Tuple[str] = (
+            "UpBlockMultiDim",
+            "CrossAttnUpBlockMultiDim",
+            "CrossAttnUpBlockMultiDim",
+            "CrossAttnUpBlockMultiDim",
+        ),
+        block_out_channels: Tuple[int] = (320, 640, 1280, 1280),
+        block_second_dim: Tuple[int] = (4, 4, 4, 4),
+        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: int = 8,
+    ):
+        super().__init__()
+
+        self.sample_size = sample_size
+        time_embed_dim = block_out_channels[0] * 4
+
+        # input
+        self.conv_in = LinearMultiDim([in_channels, 1, 1], 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([])
+
+        # 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 = self.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,
+                downsample_padding=downsample_padding,
+            )
+            self.down_blocks.append(down_block)
+
+        # mid
+        self.mid_block = UNetMidBlockMultiDimCrossAttn(
+            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,
+            resnet_groups=norm_num_groups,
+        )
+
+        # count how many layers upsample the images
+        self.num_upsamplers = 0
+
+        # up
+        reversed_block_out_channels = list(reversed(block_out_channels))
+        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 = self.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=attention_head_dim,
+            )
+            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 = nn.Conv2d(block_out_channels[0], out_channels, 3, padding=1)
+
+    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,
+    ):
+        down_block_type = down_block_type[7:] if down_block_type.startswith("UNetRes") else down_block_type
+        if down_block_type == "DownBlockMultiDim":
+            return DownBlockMultiDim(
+                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 == "CrossAttnDownBlockMultiDim":
+            if cross_attention_dim is None:
+                raise ValueError("cross_attention_dim must be specified for CrossAttnDownBlock2D")
+            return CrossAttnDownBlockMultiDim(
+                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,
+            )
+
+    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, (CrossAttnDownBlockMultiDim, DownBlockMultiDim, CrossAttnUpBlockMultiDim, UpBlockMultiDim)
+        ):
+            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, *args, **kwargs):
+        in_features = [in_features] if isinstance(in_features, int) else list(in_features)
+        out_features = [out_features] 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(), *args, **kwargs)
+
+    def forward(self, x):
+        shape = x.shape
+        n = len(self.in_features_multidim)
+        x = x.view(*shape[0:-n], self.in_features)
+        y = super().forward(x)
+        y = y.view(*shape[0:-n], *self.out_features_multidim)
+        return y
+
+
+class UNetMidBlockMultiDimCrossAttn(nn.Module):
+    pass
+
+
+class DownBlockMultiDim(nn.Module):
+    pass
+
+
+class CrossAttnDownBlockMultiDim(nn.Module):
+    pass
+
+
+class UpBlockMultiDim(nn.Module):
+    pass
+
+
+class CrossAttnUpBlockMultiDim(nn.Module):
+    pass

src/diffusers/pipelines/versatile_diffusion/pipeline_versatile_diffusion_image_variation.py

@@ -0,0 +1,216 @@
+# 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 List, Optional, Tuple, Union
+
+import numpy as np
+import torch
+import torch.utils.checkpoint
+
+import PIL
+from transformers import CLIPProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModel
+
+from ...models import AutoencoderKL, UNet2DConditionModel, VQModel
+from ...pipeline_utils import DiffusionPipeline, ImagePipelineOutput
+from ...schedulers import DDIMScheduler, LMSDiscreteScheduler, PNDMScheduler
+
+
+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`].
+    """
+    tokenizer: CLIPTokenizer
+    image_processor: CLIPProcessor
+    text_encoder: CLIPTextModel
+    image_encoder: CLIPVisionModel
+    image_unet: UNet2DConditionModel
+    text_unet: UNet2DConditionModel
+    vae: Union[VQModel, AutoencoderKL]
+    scheduler: Union[DDIMScheduler, PNDMScheduler, LMSDiscreteScheduler]
+
+    def __init__(
+        self,
+        tokenizer: CLIPTokenizer,
+        image_processor: CLIPProcessor,
+        image_encoder: CLIPVisionModel,
+        image_unet: UNet2DConditionModel,
+        vae: Union[VQModel, AutoencoderKL],
+        scheduler: Union[DDIMScheduler, PNDMScheduler, LMSDiscreteScheduler],
+    ):
+        super().__init__()
+        self.register_modules(
+            tokenizer=tokenizer,
+            image_processor=image_processor,
+            image_encoder=image_encoder,
+            image_unet=image_unet,
+            vae=vae,
+            scheduler=scheduler,
+        )
+
+    def _encode_prompt(self, prompt, do_classifier_free_guidance):
+        r"""
+        Encodes the image prompt into image encoder hidden states.
+
+        Args:
+            prompt (`str` or `list(int)`):
+                prompt to be encoded
+            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
+
+        if do_classifier_free_guidance:
+            dummy_images = [np.zeros((512, 512, 3))] * batch_size
+            dummy_images = self.image_processor(images=dummy_images, return_tensors="pt")
+            uncond_embeddings = self.image_encoder(dummy_images.pixel_values.to(self.device))
+            uncond_embeddings = normalize_embeddings(uncond_embeddings)
+
+        # get prompt text embeddings
+        image_input = self.image_processor(images=prompt, return_tensors="pt")
+        image_embeddings = self.image_encoder(image_input.pixel_values.to(self.device))
+        image_embeddings = normalize_embeddings(image_embeddings)
+
+        # For classifier free guidance, we need to do two forward passes.
+        # Here we concatenate the unconditional and image embeddings into a single batch
+        # to avoid doing two forward passes
+        image_embeddings = torch.cat([uncond_embeddings, image_embeddings])
+
+        return image_embeddings
+
+    @torch.no_grad()
+    def __call__(
+        self,
+        image: Optional[Union[torch.Tensor, PIL.Image.Image]] = None,
+        height: Optional[int] = 512,
+        width: Optional[int] = 512,
+        num_inference_steps: Optional[int] = 50,
+        guidance_scale: Optional[float] = 1.0,
+        eta: Optional[float] = 0.0,
+        generator: Optional[torch.Generator] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        **kwargs,
+    ) -> Union[Tuple, ImagePipelineOutput]:
+        r"""
+        Args:
+            prompt (`str` or `List[str]`):
+                The prompt or prompts to guide the image generation.
+            height (`int`, *optional*, defaults to 256):
+                The height in pixels of the generated image.
+            width (`int`, *optional*, defaults to 256):
+                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 1.0):
+                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.
+            generator (`torch.Generator`, *optional*):
+                A [torch generator](https://pytorch.org/docs/stable/generated/torch.Generator.html) 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*):
+                Whether or not to return a [`~pipeline_utils.ImagePipelineOutput`] instead of a plain tuple.
+
+        Returns:
+            [`~pipeline_utils.ImagePipelineOutput`] or `tuple`: [`~pipelines.utils.ImagePipelineOutput`] if
+            `return_dict` is True, otherwise a `tuple. When returning a tuple, the first element is a list with the
+            generated images.
+        """
+        do_classifier_free_guidance = guidance_scale > 1.0
+
+        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 isinstance(image, PIL.Image.Image):
+            batch_size = 1
+        elif isinstance(image, torch.Tensor):
+            batch_size = image.shape[0]
+        else:
+            raise ValueError(
+                f"`image_prompt` has to be of type `PIL.Image.Image` or `torch.Tensor` but is {type(image)}"
+            )
+
+        condition_embeddings = self._encode_prompt(image, do_classifier_free_guidance)
+
+        latents = torch.randn(
+            (batch_size, self.image_unet.in_channels, height // 8, width // 8), generator=generator, device=self.device
+        )
+
+        self.scheduler.set_timesteps(num_inference_steps)
+
+        # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+        accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+
+        extra_kwargs = {}
+        if accepts_eta:
+            extra_kwargs["eta"] = eta
+
+        for t in self.progress_bar(self.scheduler.timesteps):
+            if not do_classifier_free_guidance:
+                latents_input = latents
+            else:
+                latents_input = torch.cat([latents] * 2)
+
+            # predict the noise residual
+            noise_pred = self.image_unet(latents_input, t, encoder_hidden_states=condition_embeddings).sample
+            # perform guidance
+            if guidance_scale != 1.0:
+                noise_pred_uncond, noise_prediction_cond = noise_pred.chunk(2)
+                noise_pred = noise_pred_uncond + guidance_scale * (noise_prediction_cond - noise_pred_uncond)
+
+            # compute the previous noisy sample x_t -> x_t-1
+            latents = self.scheduler.step(noise_pred, t, latents, **extra_kwargs).prev_sample
+
+        # scale and decode the image latents with vae
+        latents = 1 / 0.18215 * latents
+        image = self.vae.decode(latents).sample
+
+        image = (image / 2 + 0.5).clamp(0, 1)
+        image = image.cpu().permute(0, 2, 3, 1).numpy()
+        if output_type == "pil":
+            image = self.numpy_to_pil(image)
+
+        if not return_dict:
+            return (image,)
+
+        return ImagePipelineOutput(images=image)

src/diffusers/pipelines/versatile_diffusion/pipeline_versatile_diffusion_text_to_image.py

@@ -0,0 +1,493 @@
+# 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, CLIPTextModel, CLIPTokenizer
+
+from ...models import AutoencoderKL, UNet2DConditionModel, VQModel
+from ...models.attention import Transformer2DModel
+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 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: CLIPTextModel
+    image_unet: UNet2DConditionModel
+    text_unet: UNet2DConditionModel
+    vae: Union[VQModel, AutoencoderKL]
+    scheduler: Union[DDIMScheduler, PNDMScheduler, LMSDiscreteScheduler]
+
+    def __init__(
+        self,
+        tokenizer: CLIPTokenizer,
+        text_encoder: CLIPTextModel,
+        image_unet: UNet2DConditionModel,
+        text_unet: UNet2DConditionModel,
+        vae: Union[VQModel, 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,
+        )
+
+    def swap_unet_attention_blocks(self):
+        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],
+                )
+
+    # 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.
+
+        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. Swap the attention blocks between the image and text UNets
+        self.swap_unet_attention_blocks()
+
+        # 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=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. Swap the attention blocks backs in case the UNets are reused in another pipeline
+        self.swap_unet_attention_blocks()
+
+        # 10. Post-processing
+        image = self.decode_latents(latents)
+
+        # 11. Convert to PIL
+        if output_type == "pil":
+            image = self.numpy_to_pil(image)
+
+        if not return_dict:
+            return (image,)
+
+        return ImagePipelineOutput(images=image)

src/diffusers/utils/dummy_torch_and_transformers_objects.py

@@ -124,6 +124,36 @@ class StableDiffusionPipeline(metaclass=DummyObject):
         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 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"]
 

tests/pipelines/versatile_diffusion/test_versatile_diffusion_image_variation.py

@@ -18,7 +18,7 @@ import unittest
 import numpy as np
 import torch
 
-from diffusers import VersatileDiffusionPipeline
+from diffusers import VersatileDiffusionImageVariationPipeline
 from diffusers.utils.testing_utils import load_image, require_torch, slow, torch_device
 
 from ...test_pipelines_common import PipelineTesterMixin
@@ -27,32 +27,15 @@ from ...test_pipelines_common import PipelineTesterMixin
 torch.backends.cuda.matmul.allow_tf32 = False
 
 
-class VersatileDiffusionPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
+class VersatileDiffusionImageVariationPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
     pass
 
 
 @slow
 @require_torch
-class VersatileDiffusionPipelineIntegrationTests(unittest.TestCase):
-    def test_inference_text2img(self):
-        pipe = VersatileDiffusionPipeline.from_pretrained("diffusers/vd-official-test")
-        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, -3:, -3:, -1]
-
-        assert image.shape == (1, 512, 512, 3)
-        expected_slice = np.array([0.9256, 0.9340, 0.8933, 0.9361, 0.9113, 0.8727, 0.9122, 0.8745, 0.8099])
-        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2
-
+class VersatileDiffusionImageVariationPipelineIntegrationTests(unittest.TestCase):
     def test_inference_image_variations(self):
-        pipe = VersatileDiffusionPipeline.from_pretrained("diffusers/vd-official-test")
+        pipe = VersatileDiffusionImageVariationPipeline.from_pretrained("diffusers/vd-official-test")
         pipe.to(torch_device)
         pipe.set_progress_bar_config(disable=None)
 
@@ -61,7 +44,7 @@ class VersatileDiffusionPipelineIntegrationTests(unittest.TestCase):
         )
         generator = torch.Generator(device=torch_device).manual_seed(0)
         image = pipe(
-            image_prompt=image_prompt,
+            image=image_prompt,
             generator=generator,
             guidance_scale=7.5,
             num_inference_steps=50,

tests/pipelines/versatile_diffusion/test_versatile_diffusion_text_to_image.py

@@ -0,0 +1,52 @@
+# 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 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 test_inference_text2img(self):
+        pipe = VersatileDiffusionTextToImagePipeline.from_pretrained("diffusers/vd-official-test")
+        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([0.0657, 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