Skip to content

GitLab

Unverified Commit 62b3c9e0 authored by Will Berman's avatar Will Berman Committed by GitHub
Browse files

unCLIP variant (#2297) 



* pipeline_variant

* Add docs for when clip_stats_path is specified

* Update src/diffusers/pipelines/stable_diffusion/pipeline_stable_unclip.py
Co-authored-by: default avatarPatrick von Platen <patrick.v.platen@gmail.com>

* Update src/diffusers/pipelines/stable_diffusion/pipeline_stable_unclip.py
Co-authored-by: default avatarPatrick von Platen <patrick.v.platen@gmail.com>

* Update src/diffusers/pipelines/stable_diffusion/pipeline_stable_unclip_img2img.py
Co-authored-by: default avatarPatrick von Platen <patrick.v.platen@gmail.com>

* Update src/diffusers/pipelines/stable_diffusion/pipeline_stable_unclip_img2img.py
Co-authored-by: default avatarPatrick von Platen <patrick.v.platen@gmail.com>

* prepare_latents # Copied from re: @patrickvonplaten

* NoiseAugmentor->ImageNormalizer

* stable_unclip_prior default to None re: @patrickvonplaten

* prepare_prior_extra_step_kwargs

* prior denoising scale model input

* {DDIM,DDPM}Scheduler -> KarrasDiffusionSchedulers re: @patrickvonplaten

* docs

* Update docs/source/en/api/pipelines/stable_unclip.mdx
Co-authored-by: default avatarPatrick von Platen <patrick.v.platen@gmail.com>

---------
Co-authored-by: default avatarPatrick von Platen <patrick.v.platen@gmail.com>
parent e55687e1
Hide whitespace changes
Inline Side-by-side
Showing with 2589 additions and 17 deletions
docs/source/en/_toctree.yml
View file @ 62b3c9e0
......@@ -154,6 +154,8 @@
title: Stable Diffusion
- local: api/pipelines/stable_diffusion_2
title: Stable Diffusion 2
- local: api/pipelines/stable_unclip
title: Stable unCLIP
- local: api/pipelines/stochastic_karras_ve
title: Stochastic Karras VE
- local: api/pipelines/unclip
......
docs/source/en/api/pipelines/overview.mdx
View file @ 62b3c9e0
......@@ -64,6 +64,8 @@ available a colab notebook to directly try them out.
| [stable_diffusion_2](./stable_diffusion_2) | [**Stable Diffusion 2**](https://stability.ai/blog/stable-diffusion-v2-release) | Text-Guided Image Inpainting |
| [stable_diffusion_2](./stable_diffusion_2) | [**Stable Diffusion 2**](https://stability.ai/blog/stable-diffusion-v2-release) | Text-Guided Super Resolution Image-to-Image |
| [stable_diffusion_safe](./stable_diffusion_safe) | [**Safe Stable Diffusion**](https://arxiv.org/abs/2211.05105) | Text-Guided Generation | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/ml-research/safe-latent-diffusion/blob/main/examples/Safe%20Latent%20Diffusion.ipynb)
| [stable_unclip](./stable_unclip) | **Stable unCLIP** | Text-to-Image Generation |
| [stable_unclip](./stable_unclip) | **Stable unCLIP** | Image-to-Image Text-Guided Generation |
| [stochastic_karras_ve](./stochastic_karras_ve) | [**Elucidating the Design Space of Diffusion-Based Generative Models**](https://arxiv.org/abs/2206.00364) | Unconditional Image Generation |
| [unclip](./unclip) | [Hierarchical Text-Conditional Image Generation with CLIP Latents](https://arxiv.org/abs/2204.06125) | Text-to-Image Generation |
| [versatile_diffusion](./versatile_diffusion) | [Versatile Diffusion: Text, Images and Variations All in One Diffusion Model](https://arxiv.org/abs/2211.08332) | Text-to-Image Generation |
......
docs/source/en/api/pipelines/stable_diffusion/text2img.mdx
View file @ 62b3c9e0
......@@ -17,7 +17,7 @@ specific language governing permissions and limitations under the License.
The Stable Diffusion model was created by the researchers and engineers from [CompVis](https://github.com/CompVis), [Stability AI](https://stability.ai/), [runway](https://github.com/runwayml), and [LAION](https://laion.ai/). The [`StableDiffusionPipeline`] is capable of generating photo-realistic images given any text input using Stable Diffusion.
The original codebase can be found here:
- *Stable Diffusion V1*: [CampVis/stable-diffusion](https://github.com/CompVis/stable-diffusion)
- *Stable Diffusion V1*: [CompVis/stable-diffusion](https://github.com/CompVis/stable-diffusion)
- *Stable Diffusion v2*: [Stability-AI/stablediffusion](https://github.com/Stability-AI/stablediffusion)
Available Checkpoints are:
......
docs/source/en/api/pipelines/stable_unclip.mdx 0 → 100644
View file @ 62b3c9e0
<!--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.
-->
# Stable unCLIP
Stable unCLIP checkpoints are finetuned from [stable diffusion 2.1](./stable_diffusion_2) checkpoints to condition on CLIP image embeddings.
Stable unCLIP also still conditions on text embeddings. Given the two separate conditionings, stable unCLIP can be used
for text guided image variation. When combined with an unCLIP prior, it can also be used for full text to image generation.
## Tips
Stable unCLIP takes a `noise_level` as input during inference. `noise_level` determines how much noise is added
to the image embeddings. A higher `noise_level` increases variation in the final un-noised images. By default,
we do not add any additional noise to the image embeddings i.e. `noise_level = 0`.
### Available checkpoints:
TODO
### Text-to-Image Generation
```python
import torch
from diffusers import StableUnCLIPPipeline
pipe = StableUnCLIPPipeline.from_pretrained(
"fusing/stable-unclip-2-1-l", torch_dtype=torch.float16
) # TODO update model path
pipe = pipe.to("cuda")
prompt = "a photo of an astronaut riding a horse on mars"
images = pipe(prompt).images
images[0].save("astronaut_horse.png")
```
### Text guided Image-to-Image Variation
```python
import requests
import torch
from PIL import Image
from io import BytesIO
from diffusers import StableUnCLIPImg2ImgPipeline
pipe = StableUnCLIPImg2ImgPipeline.from_pretrained(
"fusing/stable-unclip-2-1-l-img2img", torch_dtype=torch.float16
) # TODO update model path
pipe = pipe.to("cuda")
url = "https://raw.githubusercontent.com/CompVis/stable-diffusion/main/assets/stable-samples/img2img/sketch-mountains-input.jpg"
response = requests.get(url)
init_image = Image.open(BytesIO(response.content)).convert("RGB")
init_image = init_image.resize((768, 512))
prompt = "A fantasy landscape, trending on artstation"
images = pipe(prompt, init_image).images
images[0].save("fantasy_landscape.png")
```
### StableUnCLIPPipeline
[[autodoc]] StableUnCLIPPipeline
- all
- __call__
- enable_attention_slicing
- disable_attention_slicing
- enable_vae_slicing
- disable_vae_slicing
- enable_xformers_memory_efficient_attention
- disable_xformers_memory_efficient_attention
### StableUnCLIPImg2ImgPipeline
[[autodoc]] StableUnCLIPImg2ImgPipeline
- all
- __call__
- enable_attention_slicing
- disable_attention_slicing
- enable_vae_slicing
- disable_vae_slicing
- enable_xformers_memory_efficient_attention
- disable_xformers_memory_efficient_attention
\ No newline at end of file
docs/source/en/index.mdx
View file @ 62b3c9e0
......@@ -54,6 +54,8 @@ available a colab notebook to directly try them out.
| [stable_diffusion_2](./api/pipelines/stable_diffusion_2) | [**Stable Diffusion 2**](https://stability.ai/blog/stable-diffusion-v2-release) | Text-Guided Image Inpainting |
| [stable_diffusion_2](./api/pipelines/stable_diffusion_2) | [**Stable Diffusion 2**](https://stability.ai/blog/stable-diffusion-v2-release) | Text-Guided Super Resolution Image-to-Image |
| [stable_diffusion_safe](./api/pipelines/stable_diffusion_safe) | [**Safe Stable Diffusion**](https://arxiv.org/abs/2211.05105) | Text-Guided Generation | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/ml-research/safe-latent-diffusion/blob/main/examples/Safe%20Latent%20Diffusion.ipynb)
| [stable_unclip](./stable_unclip) | **Stable unCLIP** | Text-to-Image Generation |
| [stable_unclip](./stable_unclip) | **Stable unCLIP** | Image-to-Image Text-Guided Generation |
| [stochastic_karras_ve](./api/pipelines/stochastic_karras_ve) | [**Elucidating the Design Space of Diffusion-Based Generative Models**](https://arxiv.org/abs/2206.00364) | Unconditional Image Generation |
| [unclip](./api/pipelines/unclip) | [Hierarchical Text-Conditional Image Generation with CLIP Latents](https://arxiv.org/abs/2204.06125) | Text-to-Image Generation |
| [versatile_diffusion](./api/pipelines/versatile_diffusion) | [Versatile Diffusion: Text, Images and Variations All in One Diffusion Model](https://arxiv.org/abs/2211.08332) | Text-to-Image Generation |
......
scripts/convert_original_stable_diffusion_to_diffusers.py
View file @ 62b3c9e0
......@@ -100,6 +100,26 @@ if __name__ == "__main__":
)
parser.add_argument("--dump_path", default=None, type=str, required=True, help="Path to the output model.")
parser.add_argument("--device", type=str, help="Device to use (e.g. cpu, cuda:0, cuda:1, etc.)")
parser.add_argument(
"--stable_unclip",
type=str,
default=None,
required=False,
help="Set if this is a stable unCLIP model. One of 'txt2img' or 'img2img'.",
)
parser.add_argument(
"--stable_unclip_prior",
type=str,
default=None,
required=False,
help="Set if this is a stable unCLIP txt2img model. Selects which prior to use. If `--stable_unclip` is set to `txt2img`, the karlo prior (https://huggingface.co/kakaobrain/karlo-v1-alpha/tree/main/prior) is selected by default.",
)
parser.add_argument(
"--clip_stats_path",
type=str,
help="Path to the clip stats file. Only required if the stable unclip model's config specifies `model.params.noise_aug_config.params.clip_stats_path`.",
required=False,
)
args = parser.parse_args()
pipe = load_pipeline_from_original_stable_diffusion_ckpt(
......@@ -114,5 +134,8 @@ if __name__ == "__main__":
upcast_attention=args.upcast_attention,
from_safetensors=args.from_safetensors,
device=args.device,
stable_unclip=args.stable_unclip,
stable_unclip_prior=args.stable_unclip_prior,
clip_stats_path=args.clip_stats_path,
)
pipe.save_pretrained(args.dump_path, safe_serialization=args.to_safetensors)
src/diffusers/__init__.py
View file @ 62b3c9e0
......@@ -119,6 +119,8 @@ else:
StableDiffusionPipeline,
StableDiffusionPipelineSafe,
StableDiffusionUpscalePipeline,
StableUnCLIPImg2ImgPipeline,
StableUnCLIPPipeline,
UnCLIPImageVariationPipeline,
UnCLIPPipeline,
VersatileDiffusionDualGuidedPipeline,
......
src/diffusers/models/unet_2d_condition.py
View file @ 62b3c9e0
......@@ -91,7 +91,7 @@ class UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin)
resnet_time_scale_shift (`str`, *optional*, defaults to `"default"`): Time scale shift config
for resnet blocks, see [`~models.resnet.ResnetBlock2D`]. Choose from `default` or `scale_shift`.
class_embed_type (`str`, *optional*, defaults to None): The type of class embedding to use which is ultimately
summed with the time embeddings. Choose from `None`, `"timestep"`, or `"identity"`.
summed with the time embeddings. Choose from `None`, `"timestep"`, `"identity"`, or `"projection"`.
num_class_embeds (`int`, *optional*, defaults to None):
Input dimension of the learnable embedding matrix to be projected to `time_embed_dim`, when performing
class conditioning with `class_embed_type` equal to `None`.
......@@ -102,7 +102,9 @@ class UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin)
time_cond_proj_dim (`int`, *optional*, default to `None`):
The dimension of `cond_proj` layer in timestep embedding.
conv_in_kernel (`int`, *optional*, default to `3`): The kernel size of `conv_in` layer.
conv_out_kernel (`int`, *optional*, default to `3`): the Kernel size of `conv_out` layer.
conv_out_kernel (`int`, *optional*, default to `3`): The kernel size of `conv_out` layer.
projection_class_embeddings_input_dim (`int`, *optional*): The dimension of the `class_labels` input when
using the "projection" `class_embed_type`. Required when using the "projection" `class_embed_type`.
"""
_supports_gradient_checkpointing = True
......@@ -145,6 +147,7 @@ class UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin)
time_cond_proj_dim: Optional[int] = None,
conv_in_kernel: int = 3,
conv_out_kernel: int = 3,
projection_class_embeddings_input_dim: Optional[int] = None,
):
super().__init__()
......@@ -211,6 +214,19 @@ class UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin)
self.class_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim)
elif class_embed_type == "identity":
self.class_embedding = nn.Identity(time_embed_dim, time_embed_dim)
elif class_embed_type == "projection":
if projection_class_embeddings_input_dim is None:
raise ValueError(
"`class_embed_type`: 'projection' requires `projection_class_embeddings_input_dim` be set"
)
# The projection `class_embed_type` is the same as the timestep `class_embed_type` except
# 1. the `class_labels` inputs are not first converted to sinusoidal embeddings
# 2. it projects from an arbitrary input dimension.
#
# Note that `TimestepEmbedding` is quite general, being mainly linear layers and activations.
# When used for embedding actual timesteps, the timesteps are first converted to sinusoidal embeddings.
# As a result, `TimestepEmbedding` can be passed arbitrary vectors.
self.class_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim)
else:
self.class_embedding = None
......
src/diffusers/pipelines/__init__.py
View file @ 62b3c9e0
......@@ -55,6 +55,8 @@ else:
StableDiffusionLatentUpscalePipeline,
StableDiffusionPipeline,
StableDiffusionUpscalePipeline,
StableUnCLIPImg2ImgPipeline,
StableUnCLIPPipeline,
)
from .stable_diffusion_safe import StableDiffusionPipelineSafe
from .unclip import UnCLIPImageVariationPipeline, UnCLIPPipeline
......
src/diffusers/pipelines/stable_diffusion/__init__.py
View file @ 62b3c9e0
......@@ -45,7 +45,10 @@ if is_transformers_available() and is_torch_available():
from .pipeline_stable_diffusion_instruct_pix2pix import StableDiffusionInstructPix2PixPipeline
from .pipeline_stable_diffusion_latent_upscale import StableDiffusionLatentUpscalePipeline
from .pipeline_stable_diffusion_upscale import StableDiffusionUpscalePipeline
from .pipeline_stable_unclip import StableUnCLIPPipeline
from .pipeline_stable_unclip_img2img import StableUnCLIPImg2ImgPipeline
from .safety_checker import StableDiffusionSafetyChecker
from .stable_unclip_image_normalizer import StableUnCLIPImageNormalizer
try:
if not (is_transformers_available() and is_torch_available() and is_transformers_version(">=", "4.25.0")):
......
src/diffusers/pipelines/stable_diffusion/convert_from_ckpt.py
View file @ 62b3c9e0
......@@ -21,11 +21,21 @@ from typing import Optional
import requests
import torch
from transformers import AutoFeatureExtractor, BertTokenizerFast, CLIPTextModel, CLIPTokenizer, CLIPVisionConfig
from transformers import (
AutoFeatureExtractor,
BertTokenizerFast,
CLIPImageProcessor,
CLIPTextModel,
CLIPTextModelWithProjection,
CLIPTokenizer,
CLIPVisionConfig,
CLIPVisionModelWithProjection,
)
from diffusers import (
AutoencoderKL,
DDIMScheduler,
DDPMScheduler,
DPMSolverMultistepScheduler,
EulerAncestralDiscreteScheduler,
EulerDiscreteScheduler,
......@@ -33,12 +43,17 @@ from diffusers import (
LDMTextToImagePipeline,
LMSDiscreteScheduler,
PNDMScheduler,
PriorTransformer,
StableDiffusionPipeline,
StableUnCLIPImg2ImgPipeline,
StableUnCLIPPipeline,
UnCLIPScheduler,
UNet2DConditionModel,
)
from diffusers.pipelines.latent_diffusion.pipeline_latent_diffusion import LDMBertConfig, LDMBertModel
from diffusers.pipelines.paint_by_example import PaintByExampleImageEncoder, PaintByExamplePipeline
from diffusers.pipelines.stable_diffusion import StableDiffusionSafetyChecker
from diffusers.pipelines.stable_diffusion.stable_unclip_image_normalizer import StableUnCLIPImageNormalizer
from ...utils import is_omegaconf_available, is_safetensors_available, logging
from ...utils.import_utils import BACKENDS_MAPPING
......@@ -243,6 +258,17 @@ def create_unet_diffusers_config(original_config, image_size: int):
if head_dim is None:
head_dim = [5, 10, 20, 20]
class_embed_type = None
projection_class_embeddings_input_dim = None
if "num_classes" in unet_params:
if unet_params.num_classes == "sequential":
class_embed_type = "projection"
assert "adm_in_channels" in unet_params
projection_class_embeddings_input_dim = unet_params.adm_in_channels
else:
raise NotImplementedError(f"Unknown conditional unet num_classes config: {unet_params.num_classes}")
config = dict(
sample_size=image_size // vae_scale_factor,
in_channels=unet_params.in_channels,
......@@ -254,6 +280,8 @@ def create_unet_diffusers_config(original_config, image_size: int):
cross_attention_dim=unet_params.context_dim,
attention_head_dim=head_dim,
use_linear_projection=use_linear_projection,
class_embed_type=class_embed_type,
projection_class_embeddings_input_dim=projection_class_embeddings_input_dim,
)
return config
......@@ -342,6 +370,17 @@ def convert_ldm_unet_checkpoint(checkpoint, config, path=None, extract_ema=False
new_checkpoint["time_embedding.linear_2.weight"] = unet_state_dict["time_embed.2.weight"]
new_checkpoint["time_embedding.linear_2.bias"] = unet_state_dict["time_embed.2.bias"]
if config["class_embed_type"] is None:
# No parameters to port
...
elif config["class_embed_type"] == "timestep" or config["class_embed_type"] == "projection":
new_checkpoint["class_embedding.linear_1.weight"] = unet_state_dict["label_emb.0.0.weight"]
new_checkpoint["class_embedding.linear_1.bias"] = unet_state_dict["label_emb.0.0.bias"]
new_checkpoint["class_embedding.linear_2.weight"] = unet_state_dict["label_emb.0.2.weight"]
new_checkpoint["class_embedding.linear_2.bias"] = unet_state_dict["label_emb.0.2.bias"]
else:
raise NotImplementedError(f"Not implemented `class_embed_type`: {config['class_embed_type']}")
new_checkpoint["conv_in.weight"] = unet_state_dict["input_blocks.0.0.weight"]
new_checkpoint["conv_in.bias"] = unet_state_dict["input_blocks.0.0.bias"]
......@@ -780,6 +819,84 @@ def convert_open_clip_checkpoint(checkpoint):
return text_model
def stable_unclip_image_encoder(original_config):
"""
Returns the image processor and clip image encoder for the img2img unclip pipeline.
We currently know of two types of stable unclip models which separately use the clip and the openclip image
encoders.
"""
image_embedder_config = original_config.model.params.embedder_config
sd_clip_image_embedder_class = image_embedder_config.target
sd_clip_image_embedder_class = sd_clip_image_embedder_class.split(".")[-1]
if sd_clip_image_embedder_class == "ClipImageEmbedder":
clip_model_name = image_embedder_config.params.model
if clip_model_name == "ViT-L/14":
feature_extractor = CLIPImageProcessor()
image_encoder = CLIPVisionModelWithProjection.from_pretrained("openai/clip-vit-large-patch14")
else:
raise NotImplementedError(f"Unknown CLIP checkpoint name in stable diffusion checkpoint {clip_model_name}")
elif sd_clip_image_embedder_class == "FrozenOpenCLIPImageEmbedder":
feature_extractor = CLIPImageProcessor()
image_encoder = CLIPVisionModelWithProjection.from_pretrained("laion/CLIP-ViT-H-14-laion2B-s32B-b79K")
else:
raise NotImplementedError(
f"Unknown CLIP image embedder class in stable diffusion checkpoint {sd_clip_image_embedder_class}"
)
return feature_extractor, image_encoder
def stable_unclip_image_noising_components(
original_config, clip_stats_path: Optional[str] = None, device: Optional[str] = None
):
"""
Returns the noising components for the img2img and txt2img unclip pipelines.
Converts the stability noise augmentor into
1. a `StableUnCLIPImageNormalizer` for holding the CLIP stats
2. a `DDPMScheduler` for holding the noise schedule
If the noise augmentor config specifies a clip stats path, the `clip_stats_path` must be provided.
"""
noise_aug_config = original_config.model.params.noise_aug_config
noise_aug_class = noise_aug_config.target
noise_aug_class = noise_aug_class.split(".")[-1]
if noise_aug_class == "CLIPEmbeddingNoiseAugmentation":
noise_aug_config = noise_aug_config.params
embedding_dim = noise_aug_config.timestep_dim
max_noise_level = noise_aug_config.noise_schedule_config.timesteps
beta_schedule = noise_aug_config.noise_schedule_config.beta_schedule
image_normalizer = StableUnCLIPImageNormalizer(embedding_dim=embedding_dim)
image_noising_scheduler = DDPMScheduler(num_train_timesteps=max_noise_level, beta_schedule=beta_schedule)
if "clip_stats_path" in noise_aug_config:
if clip_stats_path is None:
raise ValueError("This stable unclip config requires a `clip_stats_path`")
clip_mean, clip_std = torch.load(clip_stats_path, map_location=device)
clip_mean = clip_mean[None, :]
clip_std = clip_std[None, :]
clip_stats_state_dict = {
"mean": clip_mean,
"std": clip_std,
}
image_normalizer.load_state_dict(clip_stats_state_dict)
else:
raise NotImplementedError(f"Unknown noise augmentor class: {noise_aug_class}")
return image_normalizer, image_noising_scheduler
def load_pipeline_from_original_stable_diffusion_ckpt(
checkpoint_path: str,
original_config_file: str = None,
......@@ -792,6 +909,9 @@ def load_pipeline_from_original_stable_diffusion_ckpt(
upcast_attention: Optional[bool] = None,
device: str = None,
from_safetensors: bool = False,
stable_unclip: Optional[str] = None,
stable_unclip_prior: Optional[str] = None,
clip_stats_path: Optional[str] = None,
) -> StableDiffusionPipeline:
"""
Load a Stable Diffusion pipeline object from a CompVis-style `.ckpt`/`.safetensors` file and (ideally) a `.yaml`
......@@ -976,16 +1096,73 @@ def load_pipeline_from_original_stable_diffusion_ckpt(
if model_type == "FrozenOpenCLIPEmbedder":
text_model = convert_open_clip_checkpoint(checkpoint)
tokenizer = CLIPTokenizer.from_pretrained("stabilityai/stable-diffusion-2", subfolder="tokenizer")
pipe = StableDiffusionPipeline(
vae=vae,
text_encoder=text_model,
tokenizer=tokenizer,
unet=unet,
scheduler=scheduler,
safety_checker=None,
feature_extractor=None,
requires_safety_checker=False,
)
if stable_unclip is None:
pipe = StableDiffusionPipeline(
vae=vae,
text_encoder=text_model,
tokenizer=tokenizer,
unet=unet,
scheduler=scheduler,
safety_checker=None,
feature_extractor=None,
requires_safety_checker=False,
)
else:
image_normalizer, image_noising_scheduler = stable_unclip_image_noising_components(
original_config, clip_stats_path=clip_stats_path, device=device
)
if stable_unclip == "img2img":
feature_extractor, image_encoder = stable_unclip_image_encoder(original_config)
pipe = StableUnCLIPImg2ImgPipeline(
# image encoding components
feature_extractor=feature_extractor,
image_encoder=image_encoder,
# image noising components
image_normalizer=image_normalizer,
image_noising_scheduler=image_noising_scheduler,
# regular denoising components
tokenizer=tokenizer,
text_encoder=text_model,
unet=unet,
scheduler=scheduler,
# vae
vae=vae,
)
elif stable_unclip == "txt2img":
if stable_unclip_prior is None or stable_unclip_prior == "karlo":
karlo_model = "kakaobrain/karlo-v1-alpha"
prior = PriorTransformer.from_pretrained(karlo_model, subfolder="prior")
prior_tokenizer = CLIPTokenizer.from_pretrained("openai/clip-vit-large-patch14")
prior_text_model = CLIPTextModelWithProjection.from_pretrained("openai/clip-vit-large-patch14")
prior_scheduler = UnCLIPScheduler.from_pretrained(karlo_model, subfolder="prior_scheduler")
prior_scheduler = DDPMScheduler.from_config(prior_scheduler.config)
else:
raise NotImplementedError(f"unknown prior for stable unclip model: {stable_unclip_prior}")
pipe = StableUnCLIPPipeline(
# prior components
prior_tokenizer=prior_tokenizer,
prior_text_encoder=prior_text_model,
prior=prior,
prior_scheduler=prior_scheduler,
# image noising components
image_normalizer=image_normalizer,
image_noising_scheduler=image_noising_scheduler,
# regular denoising components
tokenizer=tokenizer,
text_encoder=text_model,
unet=unet,
scheduler=scheduler,
# vae
vae=vae,
)
else:
raise NotImplementedError(f"unknown `stable_unclip` type: {stable_unclip}")
elif model_type == "PaintByExample":
vision_model = convert_paint_by_example_checkpoint(checkpoint)
tokenizer = CLIPTokenizer.from_pretrained("openai/clip-vit-large-patch14")
......
src/diffusers/pipelines/stable_diffusion/pipeline_stable_unclip.py 0 → 100644
View file @ 62b3c9e0
# 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 Any, Callable, Dict, List, Optional, Tuple, Union
import torch
from transformers import CLIPTextModel, CLIPTextModelWithProjection, CLIPTokenizer
from transformers.models.clip.modeling_clip import CLIPTextModelOutput
from ...models import AutoencoderKL, PriorTransformer, UNet2DConditionModel
from ...models.embeddings import get_timestep_embedding
from ...schedulers import KarrasDiffusionSchedulers
from ...utils import is_accelerate_available, logging, randn_tensor, replace_example_docstring
from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput
from .stable_unclip_image_normalizer import StableUnCLIPImageNormalizer
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
EXAMPLE_DOC_STRING = """
Examples:
```py
>>> import torch
>>> from diffusers import StableUnCLIPPipeline
>>> pipe = StableUnCLIPPipeline.from_pretrained(
... "fusing/stable-unclip-2-1-l", torch_dtype=torch.float16
... ) # TODO update model path
>>> pipe = pipe.to("cuda")
>>> prompt = "a photo of an astronaut riding a horse on mars"
>>> images = pipe(prompt).images
>>> images[0].save("astronaut_horse.png")
```
"""
class StableUnCLIPPipeline(DiffusionPipeline):
"""
Pipeline for text-to-image generation using stable 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:
prior_tokenizer ([`CLIPTokenizer`]):
Tokenizer of class
[CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
prior_text_encoder ([`CLIPTextModelWithProjection`]):
Frozen text-encoder.
prior ([`PriorTransformer`]):
The canonincal unCLIP prior to approximate the image embedding from the text embedding.
prior_scheduler ([`KarrasDiffusionSchedulers`]):
Scheduler used in the prior denoising process.
image_normalizer ([`StableUnCLIPImageNormalizer`]):
Used to normalize the predicted image embeddings before the noise is applied and un-normalize the image
embeddings after the noise has been applied.
image_noising_scheduler ([`KarrasDiffusionSchedulers`]):
Noise schedule for adding noise to the predicted image embeddings. The amount of noise to add is determined
by `noise_level` in `StableUnCLIPPipeline.__call__`.
tokenizer (`CLIPTokenizer`):
Tokenizer of class
[CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
text_encoder ([`CLIPTextModel`]):
Frozen text-encoder.
unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents.
scheduler ([`KarrasDiffusionSchedulers`]):
A scheduler to be used in combination with `unet` to denoise the encoded image latents.
vae ([`AutoencoderKL`]):
Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
"""
# prior components
prior_tokenizer: CLIPTokenizer
prior_text_encoder: CLIPTextModelWithProjection
prior: PriorTransformer
prior_scheduler: KarrasDiffusionSchedulers
# image noising components
image_normalizer: StableUnCLIPImageNormalizer
image_noising_scheduler: KarrasDiffusionSchedulers
# regular denoising components
tokenizer: CLIPTokenizer
text_encoder: CLIPTextModel
unet: UNet2DConditionModel
scheduler: KarrasDiffusionSchedulers
vae: AutoencoderKL
def __init__(
self,
# prior components
prior_tokenizer: CLIPTokenizer,
prior_text_encoder: CLIPTextModelWithProjection,
prior: PriorTransformer,
prior_scheduler: KarrasDiffusionSchedulers,
# image noising components
image_normalizer: StableUnCLIPImageNormalizer,
image_noising_scheduler: KarrasDiffusionSchedulers,
# regular denoising components
tokenizer: CLIPTokenizer,
text_encoder: CLIPTextModelWithProjection,
unet: UNet2DConditionModel,
scheduler: KarrasDiffusionSchedulers,
# vae
vae: AutoencoderKL,
):
super().__init__()
self.register_modules(
prior_tokenizer=prior_tokenizer,
prior_text_encoder=prior_text_encoder,
prior=prior,
prior_scheduler=prior_scheduler,
image_normalizer=image_normalizer,
image_noising_scheduler=image_noising_scheduler,
tokenizer=tokenizer,
text_encoder=text_encoder,
unet=unet,
scheduler=scheduler,
vae=vae,
)
self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_slicing
def enable_vae_slicing(self):
r"""
Enable sliced VAE decoding.
When this option is enabled, the VAE will split the input tensor in slices to compute decoding in several
steps. This is useful to save some memory and allow larger batch sizes.
"""
self.vae.enable_slicing()
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_slicing
def disable_vae_slicing(self):
r"""
Disable sliced VAE decoding. If `enable_vae_slicing` was previously invoked, this method will go back to
computing decoding in one step.
"""
self.vae.disable_slicing()
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}")
# TODO: self.prior.post_process_latents and self.image_noiser.{scale,unscale} are not covered by the offload hooks, so they fails if added to the list
models = [
self.prior_text_encoder,
self.text_encoder,
self.unet,
self.vae,
]
for cpu_offloaded_model in models:
if cpu_offloaded_model is not None:
cpu_offload(cpu_offloaded_model, device)
@property
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._execution_device
def _execution_device(self):
r"""
Returns the device on which the pipeline's models will be executed. After calling
`pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module
hooks.
"""
if self.device != torch.device("meta") or not hasattr(self.unet, "_hf_hook"):
return self.device
for module in self.unet.modules():
if (
hasattr(module, "_hf_hook")
and hasattr(module._hf_hook, "execution_device")
and module._hf_hook.execution_device is not None
):
return torch.device(module._hf_hook.execution_device)
return self.device
# Copied from diffusers.pipelines.unclip.pipeline_unclip.UnCLIPPipeline._encode_prompt with _encode_prompt->_encode_prior_prompt, tokenizer->prior_tokenizer, text_encoder->prior_text_encoder
def _encode_prior_prompt(
self,
prompt,
device,
num_images_per_prompt,
do_classifier_free_guidance,
text_model_output: Optional[Union[CLIPTextModelOutput, Tuple]] = None,
text_attention_mask: Optional[torch.Tensor] = None,
):
if text_model_output is None:
batch_size = len(prompt) if isinstance(prompt, list) else 1
# get prompt text embeddings
text_inputs = self.prior_tokenizer(
prompt,
padding="max_length",
max_length=self.prior_tokenizer.model_max_length,
truncation=True,
return_tensors="pt",
)
text_input_ids = text_inputs.input_ids
text_mask = text_inputs.attention_mask.bool().to(device)
untruncated_ids = self.prior_tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
text_input_ids, untruncated_ids
):
removed_text = self.prior_tokenizer.batch_decode(
untruncated_ids[:, self.prior_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.prior_tokenizer.model_max_length} tokens: {removed_text}"
)
text_input_ids = text_input_ids[:, : self.prior_tokenizer.model_max_length]
prior_text_encoder_output = self.prior_text_encoder(text_input_ids.to(device))
prompt_embeds = prior_text_encoder_output.text_embeds
prior_text_encoder_hidden_states = prior_text_encoder_output.last_hidden_state
else:
batch_size = text_model_output[0].shape[0]
prompt_embeds, prior_text_encoder_hidden_states = text_model_output[0], text_model_output[1]
text_mask = text_attention_mask
prompt_embeds = prompt_embeds.repeat_interleave(num_images_per_prompt, dim=0)
prior_text_encoder_hidden_states = prior_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
uncond_input = self.prior_tokenizer(
uncond_tokens,
padding="max_length",
max_length=self.prior_tokenizer.model_max_length,
truncation=True,
return_tensors="pt",
)
uncond_text_mask = uncond_input.attention_mask.bool().to(device)
negative_prompt_embeds_prior_text_encoder_output = self.prior_text_encoder(
uncond_input.input_ids.to(device)
)
negative_prompt_embeds = negative_prompt_embeds_prior_text_encoder_output.text_embeds
uncond_prior_text_encoder_hidden_states = (
negative_prompt_embeds_prior_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_prior_text_encoder_hidden_states.shape[1]
uncond_prior_text_encoder_hidden_states = uncond_prior_text_encoder_hidden_states.repeat(
1, num_images_per_prompt, 1
)
uncond_prior_text_encoder_hidden_states = uncond_prior_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])
prior_text_encoder_hidden_states = torch.cat(
[uncond_prior_text_encoder_hidden_states, prior_text_encoder_hidden_states]
)
text_mask = torch.cat([uncond_text_mask, text_mask])
return prompt_embeds, prior_text_encoder_hidden_states, text_mask
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt
def _encode_prompt(
self,
prompt,
device,
num_images_per_prompt,
do_classifier_free_guidance,
negative_prompt=None,
prompt_embeds: Optional[torch.FloatTensor] = None,
negative_prompt_embeds: Optional[torch.FloatTensor] = None,
):
r"""
Encodes the prompt into text encoder hidden states.
Args:
prompt (`str` or `List[str]`, *optional*):
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]`, *optional*):
The prompt or prompts not to guide the image generation. If not defined, one has to pass
`negative_prompt_embeds`. instead. If not defined, one has to pass `negative_prompt_embeds`. instead.
Ignored when not using guidance (i.e., ignored if `guidance_scale` is less than `1`).
prompt_embeds (`torch.FloatTensor`, *optional*):
Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
provided, text embeddings will be generated from `prompt` input argument.
negative_prompt_embeds (`torch.FloatTensor`, *optional*):
Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
argument.
"""
if prompt is not None and isinstance(prompt, str):
batch_size = 1
elif prompt is not None and isinstance(prompt, list):
batch_size = len(prompt)
else:
batch_size = prompt_embeds.shape[0]
if prompt_embeds is None:
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="longest", return_tensors="pt").input_ids
if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and 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
prompt_embeds = self.text_encoder(
text_input_ids.to(device),
attention_mask=attention_mask,
)
prompt_embeds = prompt_embeds[0]
prompt_embeds = prompt_embeds.to(dtype=self.text_encoder.dtype, device=device)
bs_embed, seq_len, _ = prompt_embeds.shape
# duplicate text embeddings for each generation per prompt, using mps friendly method
prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
# get unconditional embeddings for classifier free guidance
if do_classifier_free_guidance and negative_prompt_embeds is None:
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 = prompt_embeds.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
negative_prompt_embeds = self.text_encoder(
uncond_input.input_ids.to(device),
attention_mask=attention_mask,
)
negative_prompt_embeds = negative_prompt_embeds[0]
if do_classifier_free_guidance:
# 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.to(dtype=self.text_encoder.dtype, device=device)
negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
negative_prompt_embeds = negative_prompt_embeds.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
prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
return prompt_embeds
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents
def decode_latents(self, latents):
latents = 1 / self.vae.config.scaling_factor * 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 bfloat16
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 with prepare_extra_step_kwargs->prepare_prior_extra_step_kwargs, scheduler->prior_scheduler
def prepare_prior_extra_step_kwargs(self, generator, eta):
# prepare extra kwargs for the prior_scheduler step, since not all prior_schedulers have the same signature
# eta (η) is only used with the DDIMScheduler, it will be ignored for other prior_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.prior_scheduler.step).parameters.keys())
extra_step_kwargs = {}
if accepts_eta:
extra_step_kwargs["eta"] = eta
# check if the prior_scheduler accepts generator
accepts_generator = "generator" in set(inspect.signature(self.prior_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.prepare_extra_step_kwargs
def prepare_extra_step_kwargs(self, generator, eta):
# prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
# eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
# eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
# and should be between [0, 1]
accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
extra_step_kwargs = {}
if accepts_eta:
extra_step_kwargs["eta"] = eta
# check if the scheduler accepts generator
accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
if accepts_generator:
extra_step_kwargs["generator"] = generator
return extra_step_kwargs
def check_inputs(
self,
prompt,
height,
width,
callback_steps,
noise_level,
negative_prompt=None,
prompt_embeds=None,
negative_prompt_embeds=None,
):
if height % 8 != 0 or width % 8 != 0:
raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
if (callback_steps is None) or (
callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
):
raise ValueError(
f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
f" {type(callback_steps)}."
)
if prompt is not None and prompt_embeds is not None:
raise ValueError(
"Provide either `prompt` or `prompt_embeds`. Please make sure to define only one of the two."
)
if prompt is None and prompt_embeds is None:
raise ValueError(
"Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
)
if prompt is not None and (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 negative_prompt is not None and negative_prompt_embeds is not None:
raise ValueError(
"Provide either `negative_prompt` or `negative_prompt_embeds`. Cannot leave both `negative_prompt` and `negative_prompt_embeds` undefined."
)
if prompt is not None and negative_prompt is not None:
if 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)}."
)
if prompt_embeds is not None and negative_prompt_embeds is not None:
if prompt_embeds.shape != negative_prompt_embeds.shape:
raise ValueError(
"`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
f" {negative_prompt_embeds.shape}."
)
if noise_level < 0 or noise_level >= self.image_noising_scheduler.config.num_train_timesteps:
raise ValueError(
f"`noise_level` must be between 0 and {self.image_noising_scheduler.config.num_train_timesteps - 1}, inclusive."
)
# 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
def noise_image_embeddings(
self,
image_embeds: torch.Tensor,
noise_level: int,
noise: Optional[torch.FloatTensor] = None,
generator: Optional[torch.Generator] = None,
):
"""
Add noise to the image embeddings. The amount of noise is controlled by a `noise_level` input. A higher
`noise_level` increases the variance in the final un-noised images.
The noise is applied in two ways
1. A noise schedule is applied directly to the embeddings
2. A vector of sinusoidal time embeddings are appended to the output.
In both cases, the amount of noise is controlled by the same `noise_level`.
The embeddings are normalized before the noise is applied and un-normalized after the noise is applied.
"""
if noise is None:
noise = randn_tensor(
image_embeds.shape, generator=generator, device=image_embeds.device, dtype=image_embeds.dtype
)
noise_level = torch.tensor([noise_level] * image_embeds.shape[0], device=image_embeds.device)
image_embeds = self.image_normalizer.scale(image_embeds)
image_embeds = self.image_noising_scheduler.add_noise(image_embeds, timesteps=noise_level, noise=noise)
image_embeds = self.image_normalizer.unscale(image_embeds)
noise_level = get_timestep_embedding(
timesteps=noise_level, embedding_dim=image_embeds.shape[-1], flip_sin_to_cos=True, downscale_freq_shift=0
)
# `get_timestep_embeddings` does not contain any weights and will always return f32 tensors,
# but we might actually be running in fp16. so we need to cast here.
# there might be better ways to encapsulate this.
noise_level = noise_level.to(image_embeds.dtype)
image_embeds = torch.cat((image_embeds, noise_level), 1)
return image_embeds
@torch.no_grad()
@replace_example_docstring(EXAMPLE_DOC_STRING)
def __call__(
self,
# regular denoising process args
prompt: Optional[Union[str, List[str]]] = None,
height: Optional[int] = None,
width: Optional[int] = None,
num_inference_steps: int = 20,
guidance_scale: float = 10.0,
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,
prompt_embeds: Optional[torch.FloatTensor] = None,
negative_prompt_embeds: 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,
cross_attention_kwargs: Optional[Dict[str, Any]] = None,
noise_level: int = 0,
# prior args
prior_num_inference_steps: int = 25,
prior_guidance_scale: float = 4.0,
prior_latents: Optional[torch.FloatTensor] = None,
):
"""
Function invoked when calling the pipeline for generation.
Args:
prompt (`str` or `List[str]`, *optional*):
The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
instead.
height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
The height in pixels of the generated image.
width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
The width in pixels of the generated image.
num_inference_steps (`int`, *optional*, defaults to 20):
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 10.0):
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.
negative_prompt (`str` or `List[str]`, *optional*):
The prompt or prompts not to guide the image generation. If not defined, one has to pass
`negative_prompt_embeds`. instead. If not defined, one has to pass `negative_prompt_embeds`. instead.
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` 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.
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`.
prompt_embeds (`torch.FloatTensor`, *optional*):
Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
provided, text embeddings will be generated from `prompt` input argument.
negative_prompt_embeds (`torch.FloatTensor`, *optional*):
Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
argument.
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.
cross_attention_kwargs (`dict`, *optional*):
A kwargs dictionary that if specified is passed along to the `AttnProcessor` as defined under
`self.processor` in
[diffusers.cross_attention](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/cross_attention.py).
noise_level (`int`, *optional*, defaults to `0`):
The amount of noise to add to the image embeddings. A higher `noise_level` increases the variance in
the final un-noised images. See `StableUnCLIPPipeline.noise_image_embeddings` for details.
prior_num_inference_steps (`int`, *optional*, defaults to 25):
The number of denoising steps in the prior denoising process. More denoising steps usually lead to a
higher quality image at the expense of slower inference.
prior_guidance_scale (`float`, *optional*, defaults to 4.0):
Guidance scale for the prior denoising process as defined in [Classifier-Free Diffusion
Guidance](https://arxiv.org/abs/2207.12598). `prior_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.
prior_latents (`torch.FloatTensor`, *optional*):
Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
embedding generation in the prior denoising process. 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`.
Examples:
Returns:
[`~pipelines.ImagePipelineOutput`] or `tuple`: [`~ pipeline_utils.ImagePipelineOutput`] if `return_dict` is
True, otherwise a `tuple`. When returning a tuple, the first element is a list with the generated images.
"""
# 0. Default height and width to unet
height = height or self.unet.config.sample_size * self.vae_scale_factor
width = width or self.unet.config.sample_size * self.vae_scale_factor
# 1. Check inputs. Raise error if not correct
self.check_inputs(
prompt=prompt,
height=height,
width=width,
callback_steps=callback_steps,
noise_level=noise_level,
negative_prompt=negative_prompt,
prompt_embeds=prompt_embeds,
negative_prompt_embeds=negative_prompt_embeds,
)
# 2. Define call parameters
if prompt is not None and isinstance(prompt, str):
batch_size = 1
elif prompt is not None and isinstance(prompt, list):
batch_size = len(prompt)
else:
batch_size = prompt_embeds.shape[0]
batch_size = batch_size * num_images_per_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.
prior_do_classifier_free_guidance = prior_guidance_scale > 1.0
# 3. Encode input prompt
prior_prompt_embeds, prior_text_encoder_hidden_states, prior_text_mask = self._encode_prior_prompt(
prompt=prompt,
device=device,
num_images_per_prompt=num_images_per_prompt,
do_classifier_free_guidance=prior_do_classifier_free_guidance,
)
# 4. Prepare prior timesteps
self.prior_scheduler.set_timesteps(prior_num_inference_steps, device=device)
prior_timesteps_tensor = self.prior_scheduler.timesteps
# 5. Prepare prior latent variables
embedding_dim = self.prior.config.embedding_dim
prior_latents = self.prepare_latents(
(batch_size, embedding_dim),
prior_prompt_embeds.dtype,
device,
generator,
prior_latents,
self.prior_scheduler,
)
# 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
prior_extra_step_kwargs = self.prepare_prior_extra_step_kwargs(generator, eta)
# 7. Prior denoising loop
for i, t in enumerate(self.progress_bar(prior_timesteps_tensor)):
# expand the latents if we are doing classifier free guidance
latent_model_input = torch.cat([prior_latents] * 2) if prior_do_classifier_free_guidance else prior_latents
latent_model_input = self.prior_scheduler.scale_model_input(latent_model_input, t)
predicted_image_embedding = self.prior(
latent_model_input,
timestep=t,
proj_embedding=prior_prompt_embeds,
encoder_hidden_states=prior_text_encoder_hidden_states,
attention_mask=prior_text_mask,
).predicted_image_embedding
if prior_do_classifier_free_guidance:
predicted_image_embedding_uncond, predicted_image_embedding_text = predicted_image_embedding.chunk(2)
predicted_image_embedding = predicted_image_embedding_uncond + prior_guidance_scale * (
predicted_image_embedding_text - predicted_image_embedding_uncond
)
prior_latents = self.prior_scheduler.step(
predicted_image_embedding,
timestep=t,
sample=prior_latents,
**prior_extra_step_kwargs,
).prev_sample
if callback is not None and i % callback_steps == 0:
callback(i, t, prior_latents)
prior_latents = self.prior.post_process_latents(prior_latents)
image_embeds = prior_latents
# done prior
# 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
# 8. Encode input prompt
prompt_embeds = self._encode_prompt(
prompt=prompt,
device=device,
num_images_per_prompt=num_images_per_prompt,
do_classifier_free_guidance=do_classifier_free_guidance,
negative_prompt=negative_prompt,
prompt_embeds=prompt_embeds,
negative_prompt_embeds=negative_prompt_embeds,
)
# 9. Prepare image embeddings
image_embeds = self.noise_image_embeddings(
image_embeds=image_embeds,
noise_level=noise_level,
generator=generator,
)
if do_classifier_free_guidance:
negative_prompt_embeds = torch.zeros_like(image_embeds)
# For classifier free guidance, we need to do two forward passes.
# Here we concatenate the unconditional and text embeddings into a single batch
# to avoid doing two forward passes
image_embeds = torch.cat([negative_prompt_embeds, image_embeds])
# 10. Prepare timesteps
self.scheduler.set_timesteps(num_inference_steps, device=device)
timesteps = self.scheduler.timesteps
# 11. Prepare latent variables
num_channels_latents = self.unet.in_channels
shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
latents = self.prepare_latents(
shape=shape,
dtype=prompt_embeds.dtype,
device=device,
generator=generator,
latents=latents,
scheduler=self.scheduler,
)
# 12. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
# 13. Denoising loop
for i, t in enumerate(self.progress_bar(timesteps)):
latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
# predict the noise residual
noise_pred = self.unet(
latent_model_input,
t,
encoder_hidden_states=prompt_embeds,
class_labels=image_embeds,
cross_attention_kwargs=cross_attention_kwargs,
).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
if callback is not None and i % callback_steps == 0:
callback(i, t, latents)
# 14. Post-processing
image = self.decode_latents(latents)
# 15. 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/pipelines/stable_diffusion/pipeline_stable_unclip_img2img.py 0 → 100644
View file @ 62b3c9e0
# 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 Any, Callable, Dict, List, Optional, Union
import PIL
import torch
from transformers import CLIPFeatureExtractor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection
from diffusers.utils.import_utils import is_accelerate_available
from ...models import AutoencoderKL, UNet2DConditionModel
from ...models.embeddings import get_timestep_embedding
from ...schedulers import KarrasDiffusionSchedulers
from ...utils import logging, randn_tensor, replace_example_docstring
from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput
from .stable_unclip_image_normalizer import StableUnCLIPImageNormalizer
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
EXAMPLE_DOC_STRING = """
Examples:
```py
>>> import requests
>>> import torch
>>> from PIL import Image
>>> from io import BytesIO
>>> from diffusers import StableUnCLIPImg2ImgPipeline
>>> pipe = StableUnCLIPImg2ImgPipeline.from_pretrained(
... "fusing/stable-unclip-2-1-l-img2img", torch_dtype=torch.float16
... ) # TODO update model path
>>> pipe = pipe.to("cuda")
>>> url = "https://raw.githubusercontent.com/CompVis/stable-diffusion/main/assets/stable-samples/img2img/sketch-mountains-input.jpg"
>>> response = requests.get(url)
>>> init_image = Image.open(BytesIO(response.content)).convert("RGB")
>>> init_image = init_image.resize((768, 512))
>>> prompt = "A fantasy landscape, trending on artstation"
>>> images = pipe(prompt, init_image).images
>>> images[0].save("fantasy_landscape.png")
```
"""
class StableUnCLIPImg2ImgPipeline(DiffusionPipeline):
"""
Pipeline for text-guided image to image generation using stable 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:
feature_extractor ([`CLIPFeatureExtractor`]):
Feature extractor for image pre-processing before being encoded.
image_encoder ([`CLIPVisionModelWithProjection`]):
CLIP vision model for encoding images.
image_normalizer ([`StableUnCLIPImageNormalizer`]):
Used to normalize the predicted image embeddings before the noise is applied and un-normalize the image
embeddings after the noise has been applied.
image_noising_scheduler ([`KarrasDiffusionSchedulers`]):
Noise schedule for adding noise to the predicted image embeddings. The amount of noise to add is determined
by `noise_level` in `StableUnCLIPPipeline.__call__`.
tokenizer (`CLIPTokenizer`):
Tokenizer of class
[CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
text_encoder ([`CLIPTextModel`]):
Frozen text-encoder.
unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents.
scheduler ([`KarrasDiffusionSchedulers`]):
A scheduler to be used in combination with `unet` to denoise the encoded image latents.
vae ([`AutoencoderKL`]):
Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
"""
# image encoding components
feature_extractor: CLIPFeatureExtractor
image_encoder: CLIPVisionModelWithProjection
# image noising components
image_normalizer: StableUnCLIPImageNormalizer
image_noising_scheduler: KarrasDiffusionSchedulers
# regular denoising components
tokenizer: CLIPTokenizer
text_encoder: CLIPTextModel
unet: UNet2DConditionModel
scheduler: KarrasDiffusionSchedulers
vae: AutoencoderKL
def __init__(
self,
# image encoding components
feature_extractor: CLIPFeatureExtractor,
image_encoder: CLIPVisionModelWithProjection,
# image noising components
image_normalizer: StableUnCLIPImageNormalizer,
image_noising_scheduler: KarrasDiffusionSchedulers,
# regular denoising components
tokenizer: CLIPTokenizer,
text_encoder: CLIPTextModel,
unet: UNet2DConditionModel,
scheduler: KarrasDiffusionSchedulers,
# vae
vae: AutoencoderKL,
):
super().__init__()
self.register_modules(
feature_extractor=feature_extractor,
image_encoder=image_encoder,
image_normalizer=image_normalizer,
image_noising_scheduler=image_noising_scheduler,
tokenizer=tokenizer,
text_encoder=text_encoder,
unet=unet,
scheduler=scheduler,
vae=vae,
)
self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_slicing
def enable_vae_slicing(self):
r"""
Enable sliced VAE decoding.
When this option is enabled, the VAE will split the input tensor in slices to compute decoding in several
steps. This is useful to save some memory and allow larger batch sizes.
"""
self.vae.enable_slicing()
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_slicing
def disable_vae_slicing(self):
r"""
Disable sliced VAE decoding. If `enable_vae_slicing` was previously invoked, this method will go back to
computing decoding in one step.
"""
self.vae.disable_slicing()
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}")
# TODO: self.image_noiser.{scale,unscale} are not covered by the offload hooks, so they fails if added to the list
models = [
self.image_encoder,
self.text_encoder,
self.unet,
self.vae,
]
for cpu_offloaded_model in models:
if cpu_offloaded_model is not None:
cpu_offload(cpu_offloaded_model, device)
@property
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._execution_device
def _execution_device(self):
r"""
Returns the device on which the pipeline's models will be executed. After calling
`pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module
hooks.
"""
if self.device != torch.device("meta") or not hasattr(self.unet, "_hf_hook"):
return self.device
for module in self.unet.modules():
if (
hasattr(module, "_hf_hook")
and hasattr(module._hf_hook, "execution_device")
and module._hf_hook.execution_device is not None
):
return torch.device(module._hf_hook.execution_device)
return self.device
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt
def _encode_prompt(
self,
prompt,
device,
num_images_per_prompt,
do_classifier_free_guidance,
negative_prompt=None,
prompt_embeds: Optional[torch.FloatTensor] = None,
negative_prompt_embeds: Optional[torch.FloatTensor] = None,
):
r"""
Encodes the prompt into text encoder hidden states.
Args:
prompt (`str` or `List[str]`, *optional*):
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]`, *optional*):
The prompt or prompts not to guide the image generation. If not defined, one has to pass
`negative_prompt_embeds`. instead. If not defined, one has to pass `negative_prompt_embeds`. instead.
Ignored when not using guidance (i.e., ignored if `guidance_scale` is less than `1`).
prompt_embeds (`torch.FloatTensor`, *optional*):
Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
provided, text embeddings will be generated from `prompt` input argument.
negative_prompt_embeds (`torch.FloatTensor`, *optional*):
Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
argument.
"""
if prompt is not None and isinstance(prompt, str):
batch_size = 1
elif prompt is not None and isinstance(prompt, list):
batch_size = len(prompt)
else:
batch_size = prompt_embeds.shape[0]
if prompt_embeds is None:
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="longest", return_tensors="pt").input_ids
if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and 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
prompt_embeds = self.text_encoder(
text_input_ids.to(device),
attention_mask=attention_mask,
)
prompt_embeds = prompt_embeds[0]
prompt_embeds = prompt_embeds.to(dtype=self.text_encoder.dtype, device=device)
bs_embed, seq_len, _ = prompt_embeds.shape
# duplicate text embeddings for each generation per prompt, using mps friendly method
prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
# get unconditional embeddings for classifier free guidance
if do_classifier_free_guidance and negative_prompt_embeds is None:
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 = prompt_embeds.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
negative_prompt_embeds = self.text_encoder(
uncond_input.input_ids.to(device),
attention_mask=attention_mask,
)
negative_prompt_embeds = negative_prompt_embeds[0]
if do_classifier_free_guidance:
# 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.to(dtype=self.text_encoder.dtype, device=device)
negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
negative_prompt_embeds = negative_prompt_embeds.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
prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
return prompt_embeds
def _encode_image(
self,
image,
device,
batch_size,
num_images_per_prompt,
do_classifier_free_guidance,
noise_level,
generator,
image_embeds,
):
dtype = next(self.image_encoder.parameters()).dtype
if isinstance(image, PIL.Image.Image):
# the image embedding should repeated so it matches the total batch size of the prompt
repeat_by = batch_size
else:
# assume the image input is already properly batched and just needs to be repeated so
# it matches the num_images_per_prompt.
#
# NOTE(will) this is probably missing a few number of side cases. I.e. batched/non-batched
# `image_embeds`. If those happen to be common use cases, let's think harder about
# what the expected dimensions of inputs should be and how we handle the encoding.
repeat_by = num_images_per_prompt
if not image_embeds:
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_embeds = self.image_encoder(image).image_embeds
image_embeds = self.noise_image_embeddings(
image_embeds=image_embeds,
noise_level=noise_level,
generator=generator,
)
# duplicate image embeddings for each generation per prompt, using mps friendly method
image_embeds = image_embeds.unsqueeze(1)
bs_embed, seq_len, _ = image_embeds.shape
image_embeds = image_embeds.repeat(1, repeat_by, 1)
image_embeds = image_embeds.view(bs_embed * repeat_by, seq_len, -1)
image_embeds = image_embeds.squeeze(1)
if do_classifier_free_guidance:
negative_prompt_embeds = torch.zeros_like(image_embeds)
# For classifier free guidance, we need to do two forward passes.
# Here we concatenate the unconditional and text embeddings into a single batch
# to avoid doing two forward passes
image_embeds = torch.cat([negative_prompt_embeds, image_embeds])
return image_embeds
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents
def decode_latents(self, latents):
latents = 1 / self.vae.config.scaling_factor * 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 bfloat16
image = image.cpu().permute(0, 2, 3, 1).float().numpy()
return image
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
def prepare_extra_step_kwargs(self, generator, eta):
# prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
# eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
# eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
# and should be between [0, 1]
accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
extra_step_kwargs = {}
if accepts_eta:
extra_step_kwargs["eta"] = eta
# check if the scheduler accepts generator
accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
if accepts_generator:
extra_step_kwargs["generator"] = generator
return extra_step_kwargs
def check_inputs(
self,
prompt,
image,
height,
width,
callback_steps,
noise_level,
negative_prompt=None,
prompt_embeds=None,
negative_prompt_embeds=None,
image_embeds=None,
):
if height % 8 != 0 or width % 8 != 0:
raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
if (callback_steps is None) or (
callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
):
raise ValueError(
f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
f" {type(callback_steps)}."
)
if prompt is not None and prompt_embeds is not None:
raise ValueError(
"Provide either `prompt` or `prompt_embeds`. Please make sure to define only one of the two."
)
if prompt is None and prompt_embeds is None:
raise ValueError(
"Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
)
if prompt is not None and (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 negative_prompt is not None and negative_prompt_embeds is not None:
raise ValueError(
"Provide either `negative_prompt` or `negative_prompt_embeds`. Cannot leave both `negative_prompt` and `negative_prompt_embeds` undefined."
)
if prompt is not None and negative_prompt is not None:
if 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)}."
)
if prompt_embeds is not None and negative_prompt_embeds is not None:
if prompt_embeds.shape != negative_prompt_embeds.shape:
raise ValueError(
"`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
f" {negative_prompt_embeds.shape}."
)
if noise_level < 0 or noise_level >= self.image_noising_scheduler.config.num_train_timesteps:
raise ValueError(
f"`noise_level` must be between 0 and {self.image_noising_scheduler.config.num_train_timesteps - 1}, inclusive."
)
if image is not None and image_embeds is not None:
raise ValueError(
"Provide either `image` or `image_embeds`. Please make sure to define only one of the two."
)
if image is None and image_embeds is None:
raise ValueError(
"Provide either `image` or `image_embeds`. Cannot leave both `image` and `image_embeds` undefined."
)
if image is not None:
if (
not isinstance(image, torch.Tensor)
and not isinstance(image, PIL.Image.Image)
and not isinstance(image, list)
):
raise ValueError(
"`image` has to be of type `torch.FloatTensor` or `PIL.Image.Image` or `List[PIL.Image.Image]` but is"
f" {type(image)}"
)
# 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 // self.vae_scale_factor, width // self.vae_scale_factor)
if isinstance(generator, list) and len(generator) != batch_size:
raise ValueError(
f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
f" size of {batch_size}. Make sure the batch size matches the length of the generators."
)
if latents is None:
latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
else:
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
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_unclip.StableUnCLIPPipeline.noise_image_embeddings
def noise_image_embeddings(
self,
image_embeds: torch.Tensor,
noise_level: int,
noise: Optional[torch.FloatTensor] = None,
generator: Optional[torch.Generator] = None,
):
"""
Add noise to the image embeddings. The amount of noise is controlled by a `noise_level` input. A higher
`noise_level` increases the variance in the final un-noised images.
The noise is applied in two ways
1. A noise schedule is applied directly to the embeddings
2. A vector of sinusoidal time embeddings are appended to the output.
In both cases, the amount of noise is controlled by the same `noise_level`.
The embeddings are normalized before the noise is applied and un-normalized after the noise is applied.
"""
if noise is None:
noise = randn_tensor(
image_embeds.shape, generator=generator, device=image_embeds.device, dtype=image_embeds.dtype
)
noise_level = torch.tensor([noise_level] * image_embeds.shape[0], device=image_embeds.device)
image_embeds = self.image_normalizer.scale(image_embeds)
image_embeds = self.image_noising_scheduler.add_noise(image_embeds, timesteps=noise_level, noise=noise)
image_embeds = self.image_normalizer.unscale(image_embeds)
noise_level = get_timestep_embedding(
timesteps=noise_level, embedding_dim=image_embeds.shape[-1], flip_sin_to_cos=True, downscale_freq_shift=0
)
# `get_timestep_embeddings` does not contain any weights and will always return f32 tensors,
# but we might actually be running in fp16. so we need to cast here.
# there might be better ways to encapsulate this.
noise_level = noise_level.to(image_embeds.dtype)
image_embeds = torch.cat((image_embeds, noise_level), 1)
return image_embeds
@torch.no_grad()
@replace_example_docstring(EXAMPLE_DOC_STRING)
def __call__(
self,
prompt: Union[str, List[str]] = None,
image: Union[torch.FloatTensor, PIL.Image.Image] = None,
height: Optional[int] = None,
width: Optional[int] = None,
num_inference_steps: int = 20,
guidance_scale: float = 10,
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,
prompt_embeds: Optional[torch.FloatTensor] = None,
negative_prompt_embeds: 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,
cross_attention_kwargs: Optional[Dict[str, Any]] = None,
noise_level: int = 0,
image_embeds: Optional[torch.FloatTensor] = None,
):
r"""
Function invoked when calling the pipeline for generation.
Args:
prompt (`str` or `List[str]`, *optional*):
The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
instead.
image (`torch.FloatTensor` or `PIL.Image.Image`):
`Image`, or tensor representing an image batch. The image will be encoded to its CLIP embedding which
the unet will be conditioned on. Note that the image is _not_ encoded by the vae and then used as the
latents in the denoising process such as in the standard stable diffusion text guided image variation
process.
height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
The height in pixels of the generated image.
width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
The width in pixels of the generated image.
num_inference_steps (`int`, *optional*, defaults to 20):
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 10.0):
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.
negative_prompt (`str` or `List[str]`, *optional*):
The prompt or prompts not to guide the image generation. If not defined, one has to pass
`negative_prompt_embeds`. instead. If not defined, one has to pass `negative_prompt_embeds`. instead.
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` 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.
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`.
prompt_embeds (`torch.FloatTensor`, *optional*):
Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
provided, text embeddings will be generated from `prompt` input argument.
negative_prompt_embeds (`torch.FloatTensor`, *optional*):
Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
argument.
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.
cross_attention_kwargs (`dict`, *optional*):
A kwargs dictionary that if specified is passed along to the `AttnProcessor` as defined under
`self.processor` in
[diffusers.cross_attention](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/cross_attention.py).
noise_level (`int`, *optional*, defaults to `0`):
The amount of noise to add to the image embeddings. A higher `noise_level` increases the variance in
the final un-noised images. See `StableUnCLIPPipeline.noise_image_embeddings` for details.
image_embeds (`torch.FloatTensor`, *optional*):
Pre-generated CLIP embeddings to condition the unet on. Note that these are not latents to be used in
the denoising process. If you want to provide pre-generated latents, pass them to `__call__` as
`latents`.
Examples:
Returns:
[`~pipelines.ImagePipelineOutput`] or `tuple`: [`~ pipeline_utils.ImagePipelineOutput`] if `return_dict` is
True, otherwise a `tuple`. When returning a tuple, the first element is a list with the generated images.
"""
# 0. Default height and width to unet
height = height or self.unet.config.sample_size * self.vae_scale_factor
width = width or self.unet.config.sample_size * self.vae_scale_factor
# 1. Check inputs. Raise error if not correct
self.check_inputs(
prompt=prompt,
image=image,
height=height,
width=width,
callback_steps=callback_steps,
noise_level=noise_level,
negative_prompt=negative_prompt,
prompt_embeds=prompt_embeds,
negative_prompt_embeds=negative_prompt_embeds,
image_embeds=image_embeds,
)
# 2. Define call parameters
if prompt is not None and isinstance(prompt, str):
batch_size = 1
elif prompt is not None and isinstance(prompt, list):
batch_size = len(prompt)
else:
batch_size = prompt_embeds.shape[0]
batch_size = batch_size * num_images_per_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
prompt_embeds = self._encode_prompt(
prompt=prompt,
device=device,
num_images_per_prompt=num_images_per_prompt,
do_classifier_free_guidance=do_classifier_free_guidance,
negative_prompt=negative_prompt,
prompt_embeds=prompt_embeds,
negative_prompt_embeds=negative_prompt_embeds,
)
# 4. Encoder input image
noise_level = torch.tensor([noise_level], device=device)
image_embeds = self._encode_image(
image=image,
device=device,
batch_size=batch_size,
num_images_per_prompt=num_images_per_prompt,
do_classifier_free_guidance=do_classifier_free_guidance,
noise_level=noise_level,
generator=generator,
image_embeds=image_embeds,
)
# 5. Prepare timesteps
self.scheduler.set_timesteps(num_inference_steps, device=device)
timesteps = self.scheduler.timesteps
# 6. Prepare latent variables
num_channels_latents = self.unet.in_channels
latents = self.prepare_latents(
batch_size=batch_size,
num_channels_latents=num_channels_latents,
height=height,
width=width,
dtype=prompt_embeds.dtype,
device=device,
generator=generator,
latents=latents,
)
# 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
# 8. Denoising loop
for i, t in enumerate(self.progress_bar(timesteps)):
latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
# predict the noise residual
noise_pred = self.unet(
latent_model_input,
t,
encoder_hidden_states=prompt_embeds,
class_labels=image_embeds,
cross_attention_kwargs=cross_attention_kwargs,
).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
if callback is not None and i % callback_steps == 0:
callback(i, t, latents)
# 9. Post-processing
image = self.decode_latents(latents)
# 10. Convert to PIL
if output_type == "pil":
image = self.numpy_to_pil(image)
if not return_dict:
return (image,)
return ImagePipelineOutput(images=image)
src/diffusers/pipelines/stable_diffusion/stable_unclip_image_normalizer.py 0 → 100644
View file @ 62b3c9e0
# Copyright 2022 The HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import torch
from torch import nn
from ...configuration_utils import ConfigMixin, register_to_config
from ...models.modeling_utils import ModelMixin
class StableUnCLIPImageNormalizer(ModelMixin, ConfigMixin):
"""
This class is used to hold the mean and standard deviation of the CLIP embedder used in stable unCLIP.
It is used to normalize the image embeddings before the noise is applied and un-normalize the noised image
embeddings.
"""
@register_to_config
def __init__(
self,
embedding_dim: int = 768,
):
super().__init__()
self.mean = nn.Parameter(torch.zeros(1, embedding_dim))
self.std = nn.Parameter(torch.ones(1, embedding_dim))
def scale(self, embeds):
embeds = (embeds - self.mean) * 1.0 / self.std
return embeds
def unscale(self, embeds):
embeds = (embeds * self.std) + self.mean
return embeds
src/diffusers/pipelines/versatile_diffusion/modeling_text_unet.py
View file @ 62b3c9e0
......@@ -172,7 +172,7 @@ class UNetFlatConditionModel(ModelMixin, ConfigMixin):
resnet_time_scale_shift (`str`, *optional*, defaults to `"default"`): Time scale shift config
for resnet blocks, see [`~models.resnet.ResnetBlockFlat`]. Choose from `default` or `scale_shift`.
class_embed_type (`str`, *optional*, defaults to None): The type of class embedding to use which is ultimately
summed with the time embeddings. Choose from `None`, `"timestep"`, or `"identity"`.
summed with the time embeddings. Choose from `None`, `"timestep"`, `"identity"`, or `"projection"`.
num_class_embeds (`int`, *optional*, defaults to None):
Input dimension of the learnable embedding matrix to be projected to `time_embed_dim`, when performing
class conditioning with `class_embed_type` equal to `None`.
......@@ -183,7 +183,9 @@ class UNetFlatConditionModel(ModelMixin, ConfigMixin):
time_cond_proj_dim (`int`, *optional*, default to `None`):
The dimension of `cond_proj` layer in timestep embedding.
conv_in_kernel (`int`, *optional*, default to `3`): The kernel size of `conv_in` layer.
conv_out_kernel (`int`, *optional*, default to `3`): the Kernel size of `conv_out` layer.
conv_out_kernel (`int`, *optional*, default to `3`): The kernel size of `conv_out` layer.
projection_class_embeddings_input_dim (`int`, *optional*): The dimension of the `class_labels` input when
using the "projection" `class_embed_type`. Required when using the "projection" `class_embed_type`.
"""
_supports_gradient_checkpointing = True
......@@ -231,6 +233,7 @@ class UNetFlatConditionModel(ModelMixin, ConfigMixin):
time_cond_proj_dim: Optional[int] = None,
conv_in_kernel: int = 3,
conv_out_kernel: int = 3,
projection_class_embeddings_input_dim: Optional[int] = None,
):
super().__init__()
......@@ -301,6 +304,19 @@ class UNetFlatConditionModel(ModelMixin, ConfigMixin):
self.class_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim)
elif class_embed_type == "identity":
self.class_embedding = nn.Identity(time_embed_dim, time_embed_dim)
elif class_embed_type == "projection":
if projection_class_embeddings_input_dim is None:
raise ValueError(
"`class_embed_type`: 'projection' requires `projection_class_embeddings_input_dim` be set"
)
# The projection `class_embed_type` is the same as the timestep `class_embed_type` except
# 1. the `class_labels` inputs are not first converted to sinusoidal embeddings
# 2. it projects from an arbitrary input dimension.
#
# Note that `TimestepEmbedding` is quite general, being mainly linear layers and activations.
# When used for embedding actual timesteps, the timesteps are first converted to sinusoidal embeddings.
# As a result, `TimestepEmbedding` can be passed arbitrary vectors.
self.class_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim)
else:
self.class_embedding = None
......
src/diffusers/schedulers/scheduling_ddpm.py
View file @ 62b3c9e0
......@@ -119,6 +119,7 @@ class DDPMScheduler(SchedulerMixin, ConfigMixin):
variance_type: str = "fixed_small",
clip_sample: bool = True,
prediction_type: str = "epsilon",
clip_sample_range: Optional[float] = 1.0,
):
if trained_betas is not None:
self.betas = torch.tensor(trained_betas, dtype=torch.float32)
......@@ -284,7 +285,9 @@ class DDPMScheduler(SchedulerMixin, ConfigMixin):
# 3. Clip "predicted x_0"
if self.config.clip_sample:
pred_original_sample = torch.clamp(pred_original_sample, -1, 1)
pred_original_sample = torch.clamp(
pred_original_sample, -self.config.clip_sample_range, self.config.clip_sample_range
)
# 4. Compute coefficients for pred_original_sample x_0 and current sample x_t
# See formula (7) from https://arxiv.org/pdf/2006.11239.pdf
......
src/diffusers/utils/dummy_torch_and_transformers_objects.py
View file @ 62b3c9e0
......@@ -227,6 +227,36 @@ class StableDiffusionUpscalePipeline(metaclass=DummyObject):
requires_backends(cls, ["torch", "transformers"])
class StableUnCLIPImg2ImgPipeline(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 StableUnCLIPPipeline(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 UnCLIPImageVariationPipeline(metaclass=DummyObject):
_backends = ["torch", "transformers"]
......
tests/pipelines/stable_unclip/test_stable_unclip.py 0 → 100644
View file @ 62b3c9e0
import gc
import unittest
import torch
from transformers import CLIPTextConfig, CLIPTextModel, CLIPTextModelWithProjection, CLIPTokenizer
from diffusers import (
AutoencoderKL,
DDIMScheduler,
DDPMScheduler,
PriorTransformer,
StableUnCLIPPipeline,
UNet2DConditionModel,
)
from diffusers.pipelines.stable_diffusion.stable_unclip_image_normalizer import StableUnCLIPImageNormalizer
from diffusers.utils.testing_utils import load_numpy, require_torch_gpu, slow, torch_device
from ...test_pipelines_common import PipelineTesterMixin, assert_mean_pixel_difference
class StableUnCLIPPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
pipeline_class = StableUnCLIPPipeline
# TODO(will) Expected attn_bias.stride(1) == 0 to be true, but got false
test_xformers_attention = False
def get_dummy_components(self):
embedder_hidden_size = 32
embedder_projection_dim = embedder_hidden_size
# prior components
torch.manual_seed(0)
prior_tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
torch.manual_seed(0)
prior_text_encoder = CLIPTextModelWithProjection(
CLIPTextConfig(
bos_token_id=0,
eos_token_id=2,
hidden_size=embedder_hidden_size,
projection_dim=embedder_projection_dim,
intermediate_size=37,
layer_norm_eps=1e-05,
num_attention_heads=4,
num_hidden_layers=5,
pad_token_id=1,
vocab_size=1000,
)
)
torch.manual_seed(0)
prior = PriorTransformer(
num_attention_heads=2,
attention_head_dim=12,
embedding_dim=embedder_projection_dim,
num_layers=1,
)
torch.manual_seed(0)
prior_scheduler = DDPMScheduler(
variance_type="fixed_small_log",
prediction_type="sample",
num_train_timesteps=1000,
clip_sample=True,
clip_sample_range=5.0,
beta_schedule="squaredcos_cap_v2",
)
# regular denoising components
torch.manual_seed(0)
image_normalizer = StableUnCLIPImageNormalizer(embedding_dim=embedder_hidden_size)
image_noising_scheduler = DDPMScheduler(beta_schedule="squaredcos_cap_v2")
torch.manual_seed(0)
tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
torch.manual_seed(0)
text_encoder = CLIPTextModel(
CLIPTextConfig(
bos_token_id=0,
eos_token_id=2,
hidden_size=embedder_hidden_size,
projection_dim=32,
intermediate_size=37,
layer_norm_eps=1e-05,
num_attention_heads=4,
num_hidden_layers=5,
pad_token_id=1,
vocab_size=1000,
)
)
torch.manual_seed(0)
unet = UNet2DConditionModel(
sample_size=32,
in_channels=4,
out_channels=4,
down_block_types=("CrossAttnDownBlock2D", "DownBlock2D"),
up_block_types=("UpBlock2D", "CrossAttnUpBlock2D"),
block_out_channels=(32, 64),
attention_head_dim=(2, 4),
class_embed_type="projection",
# The class embeddings are the noise augmented image embeddings.
# I.e. the image embeddings concated with the noised embeddings of the same dimension
projection_class_embeddings_input_dim=embedder_projection_dim * 2,
cross_attention_dim=embedder_hidden_size,
layers_per_block=1,
upcast_attention=True,
use_linear_projection=True,
)
torch.manual_seed(0)
scheduler = DDIMScheduler(
beta_schedule="scaled_linear",
beta_start=0.00085,
beta_end=0.012,
prediction_type="v_prediction",
set_alpha_to_one=False,
steps_offset=1,
)
torch.manual_seed(0)
vae = AutoencoderKL()
components = {
# prior components
"prior_tokenizer": prior_tokenizer,
"prior_text_encoder": prior_text_encoder,
"prior": prior,
"prior_scheduler": prior_scheduler,
# image noising components
"image_normalizer": image_normalizer,
"image_noising_scheduler": image_noising_scheduler,
# regular denoising components
"tokenizer": tokenizer,
"text_encoder": text_encoder,
"unet": unet,
"scheduler": scheduler,
"vae": vae,
}
return components
def get_dummy_inputs(self, device, seed=0):
if str(device).startswith("mps"):
generator = torch.manual_seed(seed)
else:
generator = torch.Generator(device=device).manual_seed(seed)
inputs = {
"prompt": "A painting of a squirrel eating a burger",
"generator": generator,
"num_inference_steps": 2,
"prior_num_inference_steps": 2,
"output_type": "numpy",
}
return inputs
# Overriding PipelineTesterMixin::test_attention_slicing_forward_pass
# because UnCLIP GPU undeterminism requires a looser check.
def test_attention_slicing_forward_pass(self):
test_max_difference = torch_device == "cpu"
self._test_attention_slicing_forward_pass(test_max_difference=test_max_difference)
# Overriding PipelineTesterMixin::test_inference_batch_single_identical
# because UnCLIP undeterminism requires a looser check.
def test_inference_batch_single_identical(self):
test_max_difference = torch_device in ["cpu", "mps"]
self._test_inference_batch_single_identical(test_max_difference=test_max_difference)
@slow
@require_torch_gpu
class StableUnCLIPPipelineIntegrationTests(unittest.TestCase):
def tearDown(self):
# clean up the VRAM after each test
super().tearDown()
gc.collect()
torch.cuda.empty_cache()
def test_stable_unclip(self):
expected_image = load_numpy(
"https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/stable_unclip/stable_unclip_2_1_l_anime_turtle_fp16.npy"
)
pipe = StableUnCLIPPipeline.from_pretrained("fusing/stable-unclip-2-1-l", torch_dtype=torch.float16)
pipe.to(torch_device)
pipe.set_progress_bar_config(disable=None)
generator = torch.Generator(device="cpu").manual_seed(0)
output = pipe("anime turle", generator=generator, output_type="np")
image = output.images[0]
assert image.shape == (768, 768, 3)
assert_mean_pixel_difference(image, expected_image)
def test_stable_unclip_pipeline_with_sequential_cpu_offloading(self):
torch.cuda.empty_cache()
torch.cuda.reset_max_memory_allocated()
torch.cuda.reset_peak_memory_stats()
pipe = StableUnCLIPPipeline.from_pretrained("fusing/stable-unclip-2-1-l", torch_dtype=torch.float16)
pipe = pipe.to(torch_device)
pipe.set_progress_bar_config(disable=None)
pipe.enable_attention_slicing()
pipe.enable_sequential_cpu_offload()
_ = pipe(
"anime turtle",
prior_num_inference_steps=2,
num_inference_steps=2,
output_type="np",
)
mem_bytes = torch.cuda.max_memory_allocated()
# make sure that less than 7 GB is allocated
assert mem_bytes < 7 * 10**9
tests/pipelines/stable_unclip/test_stable_unclip_img2img.py 0 → 100644
View file @ 62b3c9e0
import gc
import random
import unittest
import torch
from transformers import (
CLIPFeatureExtractor,
CLIPTextConfig,
CLIPTextModel,
CLIPTokenizer,
CLIPVisionConfig,
CLIPVisionModelWithProjection,
)
from diffusers import AutoencoderKL, DDIMScheduler, DDPMScheduler, StableUnCLIPImg2ImgPipeline, UNet2DConditionModel
from diffusers.pipelines.pipeline_utils import DiffusionPipeline
from diffusers.pipelines.stable_diffusion.stable_unclip_image_normalizer import StableUnCLIPImageNormalizer
from diffusers.utils.import_utils import is_xformers_available
from diffusers.utils.testing_utils import floats_tensor, load_image, load_numpy, require_torch_gpu, slow, torch_device
from ...test_pipelines_common import PipelineTesterMixin, assert_mean_pixel_difference
class StableUnCLIPImg2ImgPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
pipeline_class = StableUnCLIPImg2ImgPipeline
def get_dummy_components(self):
embedder_hidden_size = 32
embedder_projection_dim = embedder_hidden_size
# image encoding components
feature_extractor = CLIPFeatureExtractor(crop_size=32, size=32)
image_encoder = CLIPVisionModelWithProjection(
CLIPVisionConfig(
hidden_size=embedder_hidden_size,
projection_dim=embedder_projection_dim,
num_hidden_layers=5,
num_attention_heads=4,
image_size=32,
intermediate_size=37,
patch_size=1,
)
)
# regular denoising components
torch.manual_seed(0)
image_normalizer = StableUnCLIPImageNormalizer(embedding_dim=embedder_hidden_size)
image_noising_scheduler = DDPMScheduler(beta_schedule="squaredcos_cap_v2")
torch.manual_seed(0)
tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
torch.manual_seed(0)
text_encoder = CLIPTextModel(
CLIPTextConfig(
bos_token_id=0,
eos_token_id=2,
hidden_size=embedder_hidden_size,
projection_dim=32,
intermediate_size=37,
layer_norm_eps=1e-05,
num_attention_heads=4,
num_hidden_layers=5,
pad_token_id=1,
vocab_size=1000,
)
)
torch.manual_seed(0)
unet = UNet2DConditionModel(
sample_size=32,
in_channels=4,
out_channels=4,
down_block_types=("CrossAttnDownBlock2D", "DownBlock2D"),
up_block_types=("UpBlock2D", "CrossAttnUpBlock2D"),
block_out_channels=(32, 64),
attention_head_dim=(2, 4),
class_embed_type="projection",
# The class embeddings are the noise augmented image embeddings.
# I.e. the image embeddings concated with the noised embeddings of the same dimension
projection_class_embeddings_input_dim=embedder_projection_dim * 2,
cross_attention_dim=embedder_hidden_size,
layers_per_block=1,
upcast_attention=True,
use_linear_projection=True,
)
torch.manual_seed(0)
scheduler = DDIMScheduler(
beta_schedule="scaled_linear",
beta_start=0.00085,
beta_end=0.012,
prediction_type="v_prediction",
set_alpha_to_one=False,
steps_offset=1,
)
torch.manual_seed(0)
vae = AutoencoderKL()
components = {
# image encoding components
"feature_extractor": feature_extractor,
"image_encoder": image_encoder,
# image noising components
"image_normalizer": image_normalizer,
"image_noising_scheduler": image_noising_scheduler,
# regular denoising components
"tokenizer": tokenizer,
"text_encoder": text_encoder,
"unet": unet,
"scheduler": scheduler,
"vae": vae,
}
return components
def get_dummy_inputs(self, device, seed=0, pil_image=True):
if str(device).startswith("mps"):
generator = torch.manual_seed(seed)
else:
generator = torch.Generator(device=device).manual_seed(seed)
input_image = floats_tensor((1, 3, 32, 32), rng=random.Random(seed)).to(device)
if pil_image:
input_image = input_image * 0.5 + 0.5
input_image = input_image.clamp(0, 1)
input_image = input_image.cpu().permute(0, 2, 3, 1).float().numpy()
input_image = DiffusionPipeline.numpy_to_pil(input_image)[0]
return {
"prompt": "An anime racoon running a marathon",
"image": input_image,
"generator": generator,
"num_inference_steps": 2,
"output_type": "np",
}
# Overriding PipelineTesterMixin::test_attention_slicing_forward_pass
# because GPU undeterminism requires a looser check.
def test_attention_slicing_forward_pass(self):
test_max_difference = torch_device in ["cpu", "mps"]
self._test_attention_slicing_forward_pass(test_max_difference=test_max_difference)
# Overriding PipelineTesterMixin::test_inference_batch_single_identical
# because undeterminism requires a looser check.
def test_inference_batch_single_identical(self):
test_max_difference = torch_device in ["cpu", "mps"]
self._test_inference_batch_single_identical(test_max_difference=test_max_difference)
@unittest.skipIf(
torch_device != "cuda" or not is_xformers_available(),
reason="XFormers attention is only available with CUDA and `xformers` installed",
)
def test_xformers_attention_forwardGenerator_pass(self):
self._test_xformers_attention_forwardGenerator_pass(test_max_difference=False)
@slow
@require_torch_gpu
class StableUnCLIPImg2ImgPipelineIntegrationTests(unittest.TestCase):
def tearDown(self):
# clean up the VRAM after each test
super().tearDown()
gc.collect()
torch.cuda.empty_cache()
def test_stable_unclip_l_img2img(self):
input_image = load_image(
"https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/stable_unclip/turtle.png"
)
expected_image = load_numpy(
"https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/stable_unclip/stable_unclip_2_1_l_img2img_anime_turtle_fp16.npy"
)
pipe = StableUnCLIPImg2ImgPipeline.from_pretrained(
"fusing/stable-unclip-2-1-l-img2img", torch_dtype=torch.float16
)
pipe.to(torch_device)
pipe.set_progress_bar_config(disable=None)
generator = torch.Generator(device="cpu").manual_seed(0)
output = pipe("anime turle", image=input_image, generator=generator, output_type="np")
image = output.images[0]
assert image.shape == (768, 768, 3)
assert_mean_pixel_difference(image, expected_image)
def test_stable_unclip_h_img2img(self):
input_image = load_image(
"https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/stable_unclip/turtle.png"
)
expected_image = load_numpy(
"https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/stable_unclip/stable_unclip_2_1_h_img2img_anime_turtle_fp16.npy"
)
pipe = StableUnCLIPImg2ImgPipeline.from_pretrained(
"fusing/stable-unclip-2-1-h-img2img", torch_dtype=torch.float16
)
pipe.to(torch_device)
pipe.set_progress_bar_config(disable=None)
generator = torch.Generator(device="cpu").manual_seed(0)
output = pipe("anime turle", image=input_image, generator=generator, output_type="np")
image = output.images[0]
assert image.shape == (768, 768, 3)
assert_mean_pixel_difference(image, expected_image)
def test_stable_unclip_img2img_pipeline_with_sequential_cpu_offloading(self):
input_image = load_image(
"https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/stable_unclip/turtle.png"
)
torch.cuda.empty_cache()
torch.cuda.reset_max_memory_allocated()
torch.cuda.reset_peak_memory_stats()
pipe = StableUnCLIPImg2ImgPipeline.from_pretrained(
"fusing/stable-unclip-2-1-h-img2img", torch_dtype=torch.float16
)
pipe = pipe.to(torch_device)
pipe.set_progress_bar_config(disable=None)
pipe.enable_attention_slicing()
pipe.enable_sequential_cpu_offload()
_ = pipe(
"anime turtle",
image=input_image,
num_inference_steps=2,
output_type="np",
)
mem_bytes = torch.cuda.max_memory_allocated()
# make sure that less than 7 GB is allocated
assert mem_bytes < 7 * 10**9
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment