AI/diffusers
1
0
Fork 0
mirror of https://github.com/huggingface/diffusers.git synced 2026-01-27 17:22:53 +03:00
Code Activity

[Community Pipeline] Unclip Image Interpolation (#2400)

Browse Source 
* unclip img interpolation poc

* Added code sample and refactoring.
This commit is contained in:
Naga Sai Abhinay
2023-03-06 06:25:30 +05:30
committed by GitHub
parent 2b4f849db9
commit 2e3541d7f4
2 changed files with 538 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

45
examples/community/README.md
View File

@@ -28,6 +28,7 @@ Stable Diffusion v1.1-1.4 Comparison | Run all 4 model checkpoints for Stable Di
MagicMix | Diffusion Pipeline for semantic mixing of an image and a text prompt | [MagicMix](#magic-mix) | - | [Partho Das](https://github.com/daspartho) |
| Stable UnCLIP | Diffusion Pipeline for combining prior model (generate clip image embedding from text, UnCLIPPipeline `"kakaobrain/karlo-v1-alpha"`) and decoder pipeline (decode clip image embedding to image, StableDiffusionImageVariationPipeline `"lambdalabs/sd-image-variations-diffusers"` ). | [Stable UnCLIP](#stable-unclip) | - |[Ray Wang](https://wrong.wang) |
| UnCLIP Text Interpolation Pipeline | Diffusion Pipeline that allows passing two prompts and produces images while interpolating between the text-embeddings of the two prompts | [UnCLIP Text Interpolation Pipeline](#unclip-text-interpolation-pipeline) | - | [Naga Sai Abhinay Devarinti](https://github.com/Abhinay1997/) |
| UnCLIP Image Interpolation Pipeline | Diffusion Pipeline that allows passing two images/image_embeddings and produces images while interpolating between their image-embeddings | [UnCLIP Image Interpolation Pipeline](#unclip-image-interpolation-pipeline) | - | [Naga Sai Abhinay Devarinti](https://github.com/Abhinay1997/) |
@@ -989,3 +990,47 @@ The resulting images in order:-
![result_3](https://huggingface.co/datasets/NagaSaiAbhinay/UnCLIPTextInterpolationSamples/resolve/main/lion_to_cub_3.png)
![result_4](https://huggingface.co/datasets/NagaSaiAbhinay/UnCLIPTextInterpolationSamples/resolve/main/lion_to_cub_4.png)
![result_5](https://huggingface.co/datasets/NagaSaiAbhinay/UnCLIPTextInterpolationSamples/resolve/main/lion_to_cub_5.png)
### UnCLIP Image Interpolation Pipeline
This Diffusion Pipeline takes two images or an image_embeddings tensor of size 2 and interpolates between their embeddings using spherical interpolation ( slerp ). The input images/image_embeddings are converted to image embeddings by the pipeline's image_encoder and the interpolation is done on the resulting image_embeddings over the number of steps specified. Defaults to 5 steps.
```python
import torch
from diffusers import DiffusionPipeline
from PIL import Image
device = torch.device("cpu" if not torch.cuda.is_available() else "cuda")
dtype = torch.float16 if torch.cuda.is_available() else torch.bfloat16
pipe = DiffusionPipeline.from_pretrained(
"kakaobrain/karlo-v1-alpha-image-variations",
torch_dtype=dtype,
custom_pipeline="unclip_image_interpolation"
)
pipe.to(device)
images = [Image.open('./starry_night.jpg'), Image.open('./flowers.jpg')]
#For best results keep the prompts close in length to each other. Of course, feel free to try out with differing lengths.
generator = torch.Generator(device=device).manual_seed(42)
output = pipe(image = images ,steps = 6, generator = generator)
for i,image in enumerate(output.images):
image.save('starry_to_flowers_%s.jpg' % i)
```
The original images:-
![starry](https://huggingface.co/datasets/NagaSaiAbhinay/UnCLIPImageInterpolationSamples/resolve/main/starry_night.jpg)
![flowers](https://huggingface.co/datasets/NagaSaiAbhinay/UnCLIPImageInterpolationSamples/resolve/main/flowers.jpg)
The resulting images in order:-
![result0](https://huggingface.co/datasets/NagaSaiAbhinay/UnCLIPImageInterpolationSamples/resolve/main/starry_to_flowers_0.png)
![result1](https://huggingface.co/datasets/NagaSaiAbhinay/UnCLIPImageInterpolationSamples/resolve/main/starry_to_flowers_1.png)
![result2](https://huggingface.co/datasets/NagaSaiAbhinay/UnCLIPImageInterpolationSamples/resolve/main/starry_to_flowers_2.png)
![result3](https://huggingface.co/datasets/NagaSaiAbhinay/UnCLIPImageInterpolationSamples/resolve/main/starry_to_flowers_3.png)
![result4](https://huggingface.co/datasets/NagaSaiAbhinay/UnCLIPImageInterpolationSamples/resolve/main/starry_to_flowers_4.png)
![result5](https://huggingface.co/datasets/NagaSaiAbhinay/UnCLIPImageInterpolationSamples/resolve/main/starry_to_flowers_5.png)

493
examples/community/unclip_image_interpolation.py Normal file
View File

@@ -0,0 +1,493 @@
import inspect
from typing import List, Optional, Union
import PIL
import torch
from torch.nn import functional as F
from transformers import (
CLIPFeatureExtractor,
CLIPTextModelWithProjection,
CLIPTokenizer,
CLIPVisionModelWithProjection,
)
from diffusers import (
DiffusionPipeline,
ImagePipelineOutput,
UnCLIPScheduler,
UNet2DConditionModel,
UNet2DModel,
)
from diffusers.pipelines.unclip import UnCLIPTextProjModel
from diffusers.utils import is_accelerate_available, logging, randn_tensor
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
def slerp(val, low, high):
"""
Find the interpolation point between the 'low' and 'high' values for the given 'val'. See https://en.wikipedia.org/wiki/Slerp for more details on the topic.
"""
low_norm = low / torch.norm(low)
high_norm = high / torch.norm(high)
omega = torch.acos((low_norm * high_norm))
so = torch.sin(omega)
res = (torch.sin((1.0 - val) * omega) / so) * low + (torch.sin(val * omega) / so) * high
return res
class UnCLIPImageInterpolationPipeline(DiffusionPipeline):
"""
Pipeline to generate variations from an input image using unCLIP
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:
text_encoder ([`CLIPTextModelWithProjection`]):
Frozen text-encoder.
tokenizer (`CLIPTokenizer`):
Tokenizer of class
[CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
feature_extractor ([`CLIPFeatureExtractor`]):
Model that extracts features from generated images to be used as inputs for the `image_encoder`.
image_encoder ([`CLIPVisionModelWithProjection`]):
Frozen CLIP image-encoder. unCLIP 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.
text_proj ([`UnCLIPTextProjModel`]):
Utility class to prepare and combine the embeddings before they are passed to the decoder.
decoder ([`UNet2DConditionModel`]):
The decoder to invert the image embedding into an image.
super_res_first ([`UNet2DModel`]):
Super resolution unet. Used in all but the last step of the super resolution diffusion process.
super_res_last ([`UNet2DModel`]):
Super resolution unet. Used in the last step of the super resolution diffusion process.
decoder_scheduler ([`UnCLIPScheduler`]):
Scheduler used in the decoder denoising process. Just a modified DDPMScheduler.
super_res_scheduler ([`UnCLIPScheduler`]):
Scheduler used in the super resolution denoising process. Just a modified DDPMScheduler.
"""
decoder: UNet2DConditionModel
text_proj: UnCLIPTextProjModel
text_encoder: CLIPTextModelWithProjection
tokenizer: CLIPTokenizer
feature_extractor: CLIPFeatureExtractor
image_encoder: CLIPVisionModelWithProjection
super_res_first: UNet2DModel
super_res_last: UNet2DModel
decoder_scheduler: UnCLIPScheduler
super_res_scheduler: UnCLIPScheduler
# Copied from diffusers.pipelines.unclip.pipeline_unclip_image_variation.UnCLIPImageVariationPipeline.__init__
def __init__(
self,
decoder: UNet2DConditionModel,
text_encoder: CLIPTextModelWithProjection,
tokenizer: CLIPTokenizer,
text_proj: UnCLIPTextProjModel,
feature_extractor: CLIPFeatureExtractor,
image_encoder: CLIPVisionModelWithProjection,
super_res_first: UNet2DModel,
super_res_last: UNet2DModel,
decoder_scheduler: UnCLIPScheduler,
super_res_scheduler: UnCLIPScheduler,
):
super().__init__()
self.register_modules(
decoder=decoder,
text_encoder=text_encoder,
tokenizer=tokenizer,
text_proj=text_proj,
feature_extractor=feature_extractor,
image_encoder=image_encoder,
super_res_first=super_res_first,
super_res_last=super_res_last,
decoder_scheduler=decoder_scheduler,
super_res_scheduler=super_res_scheduler,
)
# Copied from diffusers.pipelines.unclip.pipeline_unclip.UnCLIPPipeline.prepare_latents
def prepare_latents(self, shape, dtype, device, generator, latents, scheduler):
if latents is None:
latents = randn_tensor(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)
latents = latents * scheduler.init_noise_sigma
return latents
# Copied from diffusers.pipelines.unclip.pipeline_unclip_image_variation.UnCLIPImageVariationPipeline._encode_prompt
def _encode_prompt(self, prompt, device, num_images_per_prompt, do_classifier_free_guidance):
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,
return_tensors="pt",
)
text_input_ids = text_inputs.input_ids
text_mask = text_inputs.attention_mask.bool().to(device)
text_encoder_output = self.text_encoder(text_input_ids.to(device))
prompt_embeds = text_encoder_output.text_embeds
text_encoder_hidden_states = text_encoder_output.last_hidden_state
prompt_embeds = prompt_embeds.repeat_interleave(num_images_per_prompt, dim=0)
text_encoder_hidden_states = text_encoder_hidden_states.repeat_interleave(num_images_per_prompt, dim=0)
text_mask = text_mask.repeat_interleave(num_images_per_prompt, dim=0)
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",
)
uncond_text_mask = uncond_input.attention_mask.bool().to(device)
negative_prompt_embeds_text_encoder_output = self.text_encoder(uncond_input.input_ids.to(device))
negative_prompt_embeds = negative_prompt_embeds_text_encoder_output.text_embeds
uncond_text_encoder_hidden_states = negative_prompt_embeds_text_encoder_output.last_hidden_state
# duplicate unconditional embeddings for each generation per prompt, using mps friendly method
seq_len = negative_prompt_embeds.shape[1]
negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt)
negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len)
seq_len = uncond_text_encoder_hidden_states.shape[1]
uncond_text_encoder_hidden_states = uncond_text_encoder_hidden_states.repeat(1, num_images_per_prompt, 1)
uncond_text_encoder_hidden_states = uncond_text_encoder_hidden_states.view(
batch_size * num_images_per_prompt, seq_len, -1
)
uncond_text_mask = uncond_text_mask.repeat_interleave(num_images_per_prompt, dim=0)
# done duplicates
# 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
prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
text_encoder_hidden_states = torch.cat([uncond_text_encoder_hidden_states, text_encoder_hidden_states])
text_mask = torch.cat([uncond_text_mask, text_mask])
return prompt_embeds, text_encoder_hidden_states, text_mask
# Copied from diffusers.pipelines.unclip.pipeline_unclip_image_variation.UnCLIPImageVariationPipeline._encode_image
def _encode_image(self, image, device, num_images_per_prompt, image_embeddings: Optional[torch.Tensor] = None):
dtype = next(self.image_encoder.parameters()).dtype
if image_embeddings is None:
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.repeat_interleave(num_images_per_prompt, dim=0)
return image_embeddings
# Copied from diffusers.pipelines.unclip.pipeline_unclip_image_variation.UnCLIPImageVariationPipeline.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, the pipeline's
models 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}")
models = [
self.decoder,
self.text_proj,
self.text_encoder,
self.super_res_first,
self.super_res_last,
]
for cpu_offloaded_model in models:
if cpu_offloaded_model is not None:
cpu_offload(cpu_offloaded_model, device)
@property
# Copied from diffusers.pipelines.unclip.pipeline_unclip.UnCLIPPipeline._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.decoder, "_hf_hook"):
return self.device
for module in self.decoder.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
@torch.no_grad()
def __call__(
self,
image: Optional[Union[List[PIL.Image.Image], torch.FloatTensor]] = None,
steps: int = 5,
decoder_num_inference_steps: int = 25,
super_res_num_inference_steps: int = 7,
generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
image_embeddings: Optional[torch.Tensor] = None,
decoder_latents: Optional[torch.FloatTensor] = None,
super_res_latents: Optional[torch.FloatTensor] = None,
decoder_guidance_scale: float = 8.0,
output_type: Optional[str] = "pil",
return_dict: bool = True,
):
"""
Function invoked when calling the pipeline for generation.
Args:
image (`List[PIL.Image.Image]` or `torch.FloatTensor`):
The images to use for the image interpolation. Only accepts a list of two PIL Images or If you provide a tensor, it needs to comply with the
configuration of
[this](https://huggingface.co/fusing/karlo-image-variations-diffusers/blob/main/feature_extractor/preprocessor_config.json)
`CLIPFeatureExtractor` while still having a shape of two in the 0th dimension. Can be left to `None` only when `image_embeddings` are passed.
steps (`int`, *optional*, defaults to 5):
The number of interpolation images to generate.
decoder_num_inference_steps (`int`, *optional*, defaults to 25):
The number of denoising steps for the decoder. More denoising steps usually lead to a higher quality
image at the expense of slower inference.
super_res_num_inference_steps (`int`, *optional*, defaults to 7):
The number of denoising steps for super resolution. More denoising steps usually lead to a higher
quality image at the expense of slower inference.
generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
to make generation deterministic.
image_embeddings (`torch.Tensor`, *optional*):
Pre-defined image embeddings that can be derived from the image encoder. Pre-defined image embeddings
can be passed for tasks like image interpolations. `image` can the be left to `None`.
decoder_latents (`torch.FloatTensor` of shape (batch size, channels, height, width), *optional*):
Pre-generated noisy latents to be used as inputs for the decoder.
super_res_latents (`torch.FloatTensor` of shape (batch size, channels, super res height, super res width), *optional*):
Pre-generated noisy latents to be used as inputs for the decoder.
decoder_guidance_scale (`float`, *optional*, defaults to 4.0):
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.
output_type (`str`, *optional*, defaults to `"pil"`):
The output format of the generated 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.ImagePipelineOutput`] instead of a plain tuple.
"""
batch_size = steps
device = self._execution_device
if isinstance(image, List):
if len(image) != 2:
raise AssertionError(
f"Expected 'image' List to be of size 2, but passed 'image' length is {len(image)}"
)
elif not (isinstance(image[0], PIL.Image.Image) and isinstance(image[0], PIL.Image.Image)):
raise AssertionError(
f"Expected 'image' List to contain PIL.Image.Image, but passed 'image' contents are {type(image[0])} and {type(image[1])}"
)
elif isinstance(image, torch.FloatTensor):
if image.shape[0] != 2:
raise AssertionError(
f"Expected 'image' to be torch.FloatTensor of shape 2 in 0th dimension, but passed 'image' size is {image.shape[0]}"
)
elif isinstance(image_embeddings, torch.Tensor):
if image_embeddings.shape[0] != 2:
raise AssertionError(
f"Expected 'image_embeddings' to be torch.FloatTensor of shape 2 in 0th dimension, but passed 'image_embeddings' shape is {image_embeddings.shape[0]}"
)
else:
raise AssertionError(
f"Expected 'image' or 'image_embeddings' to be not None with types List[PIL.Image] or Torch.FloatTensor respectively. Received {type(image)} and {type(image_embeddings)} repsectively"
)
original_image_embeddings = self._encode_image(
image=image, device=device, num_images_per_prompt=1, image_embeddings=image_embeddings
)
image_embeddings = []
for interp_step in torch.linspace(0, 1, steps):
temp_image_embeddings = slerp(
interp_step, original_image_embeddings[0], original_image_embeddings[1]
).unsqueeze(0)
image_embeddings.append(temp_image_embeddings)
image_embeddings = torch.cat(image_embeddings).to(device)
do_classifier_free_guidance = decoder_guidance_scale > 1.0
prompt_embeds, text_encoder_hidden_states, text_mask = self._encode_prompt(
prompt=["" for i in range(steps)],
device=device,
num_images_per_prompt=1,
do_classifier_free_guidance=do_classifier_free_guidance,
)
text_encoder_hidden_states, additive_clip_time_embeddings = self.text_proj(
image_embeddings=image_embeddings,
prompt_embeds=prompt_embeds,
text_encoder_hidden_states=text_encoder_hidden_states,
do_classifier_free_guidance=do_classifier_free_guidance,
)
if device.type == "mps":
# HACK: MPS: There is a panic when padding bool tensors,
# so cast to int tensor for the pad and back to bool afterwards
text_mask = text_mask.type(torch.int)
decoder_text_mask = F.pad(text_mask, (self.text_proj.clip_extra_context_tokens, 0), value=1)
decoder_text_mask = decoder_text_mask.type(torch.bool)
else:
decoder_text_mask = F.pad(text_mask, (self.text_proj.clip_extra_context_tokens, 0), value=True)
self.decoder_scheduler.set_timesteps(decoder_num_inference_steps, device=device)
decoder_timesteps_tensor = self.decoder_scheduler.timesteps
num_channels_latents = self.decoder.in_channels
height = self.decoder.sample_size
width = self.decoder.sample_size
decoder_latents = self.prepare_latents(
(batch_size, num_channels_latents, height, width),
text_encoder_hidden_states.dtype,
device,
generator,
decoder_latents,
self.decoder_scheduler,
)
for i, t in enumerate(self.progress_bar(decoder_timesteps_tensor)):
# expand the latents if we are doing classifier free guidance
latent_model_input = torch.cat([decoder_latents] * 2) if do_classifier_free_guidance else decoder_latents
noise_pred = self.decoder(
sample=latent_model_input,
timestep=t,
encoder_hidden_states=text_encoder_hidden_states,
class_labels=additive_clip_time_embeddings,
attention_mask=decoder_text_mask,
).sample
if do_classifier_free_guidance:
noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
noise_pred_uncond, _ = noise_pred_uncond.split(latent_model_input.shape[1], dim=1)
noise_pred_text, predicted_variance = noise_pred_text.split(latent_model_input.shape[1], dim=1)
noise_pred = noise_pred_uncond + decoder_guidance_scale * (noise_pred_text - noise_pred_uncond)
noise_pred = torch.cat([noise_pred, predicted_variance], dim=1)
if i + 1 == decoder_timesteps_tensor.shape[0]:
prev_timestep = None
else:
prev_timestep = decoder_timesteps_tensor[i + 1]
# compute the previous noisy sample x_t -> x_t-1
decoder_latents = self.decoder_scheduler.step(
noise_pred, t, decoder_latents, prev_timestep=prev_timestep, generator=generator
).prev_sample
decoder_latents = decoder_latents.clamp(-1, 1)
image_small = decoder_latents
# done decoder
# super res
self.super_res_scheduler.set_timesteps(super_res_num_inference_steps, device=device)
super_res_timesteps_tensor = self.super_res_scheduler.timesteps
channels = self.super_res_first.in_channels // 2
height = self.super_res_first.sample_size
width = self.super_res_first.sample_size
super_res_latents = self.prepare_latents(
(batch_size, channels, height, width),
image_small.dtype,
device,
generator,
super_res_latents,
self.super_res_scheduler,
)
if device.type == "mps":
# MPS does not support many interpolations
image_upscaled = F.interpolate(image_small, size=[height, width])
else:
interpolate_antialias = {}
if "antialias" in inspect.signature(F.interpolate).parameters:
interpolate_antialias["antialias"] = True
image_upscaled = F.interpolate(
image_small, size=[height, width], mode="bicubic", align_corners=False, **interpolate_antialias
)
for i, t in enumerate(self.progress_bar(super_res_timesteps_tensor)):
# no classifier free guidance
if i == super_res_timesteps_tensor.shape[0] - 1:
unet = self.super_res_last
else:
unet = self.super_res_first
latent_model_input = torch.cat([super_res_latents, image_upscaled], dim=1)
noise_pred = unet(
sample=latent_model_input,
timestep=t,
).sample
if i + 1 == super_res_timesteps_tensor.shape[0]:
prev_timestep = None
else:
prev_timestep = super_res_timesteps_tensor[i + 1]
# compute the previous noisy sample x_t -> x_t-1
super_res_latents = self.super_res_scheduler.step(
noise_pred, t, super_res_latents, prev_timestep=prev_timestep, generator=generator
).prev_sample
image = super_res_latents
# done super res
# post processing
image = image * 0.5 + 0.5
image = image.clamp(0, 1)
image = image.cpu().permute(0, 2, 3, 1).float().numpy()
if output_type == "pil":
image = self.numpy_to_pil(image)
if not return_dict:
return (image,)
return ImagePipelineOutput(images=image)