883 lines
43 KiB
Python
883 lines
43 KiB
Python
# Copyright 2023 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
|
||
import warnings
|
||
from typing import Any, Callable, Dict, List, Optional, Union
|
||
|
||
import numpy as np
|
||
import PIL
|
||
import torch
|
||
from packaging import version
|
||
from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer
|
||
|
||
from diffusers.utils import is_accelerate_available, is_accelerate_version
|
||
|
||
from ...configuration_utils import FrozenDict
|
||
from ...image_processor import VaeImageProcessor
|
||
from ...loaders import LoraLoaderMixin, TextualInversionLoaderMixin
|
||
from ...models import AutoencoderKL, UNet2DConditionModel
|
||
from ...schedulers import DDIMScheduler
|
||
from ...utils import PIL_INTERPOLATION, deprecate, logging, randn_tensor
|
||
from ..pipeline_utils import DiffusionPipeline
|
||
from . import StableDiffusionPipelineOutput
|
||
from .safety_checker import StableDiffusionSafetyChecker
|
||
|
||
|
||
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
|
||
|
||
|
||
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.preprocess
|
||
def preprocess(image):
|
||
warnings.warn(
|
||
"The preprocess method is deprecated and will be removed in a future version. Please"
|
||
" use VaeImageProcessor.preprocess instead",
|
||
FutureWarning,
|
||
)
|
||
if isinstance(image, torch.Tensor):
|
||
return image
|
||
elif isinstance(image, PIL.Image.Image):
|
||
image = [image]
|
||
|
||
if isinstance(image[0], PIL.Image.Image):
|
||
w, h = image[0].size
|
||
w, h = (x - x % 8 for x in (w, h)) # resize to integer multiple of 8
|
||
|
||
image = [np.array(i.resize((w, h), resample=PIL_INTERPOLATION["lanczos"]))[None, :] for i in image]
|
||
image = np.concatenate(image, axis=0)
|
||
image = np.array(image).astype(np.float32) / 255.0
|
||
image = image.transpose(0, 3, 1, 2)
|
||
image = 2.0 * image - 1.0
|
||
image = torch.from_numpy(image)
|
||
elif isinstance(image[0], torch.Tensor):
|
||
image = torch.cat(image, dim=0)
|
||
return image
|
||
|
||
|
||
def posterior_sample(scheduler, latents, timestep, clean_latents, generator, eta):
|
||
# 1. get previous step value (=t-1)
|
||
prev_timestep = timestep - scheduler.config.num_train_timesteps // scheduler.num_inference_steps
|
||
|
||
if prev_timestep <= 0:
|
||
return clean_latents
|
||
|
||
# 2. compute alphas, betas
|
||
alpha_prod_t = scheduler.alphas_cumprod[timestep]
|
||
alpha_prod_t_prev = (
|
||
scheduler.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else scheduler.final_alpha_cumprod
|
||
)
|
||
|
||
variance = scheduler._get_variance(timestep, prev_timestep)
|
||
std_dev_t = eta * variance ** (0.5)
|
||
|
||
# direction pointing to x_t
|
||
e_t = (latents - alpha_prod_t ** (0.5) * clean_latents) / (1 - alpha_prod_t) ** (0.5)
|
||
dir_xt = (1.0 - alpha_prod_t_prev - std_dev_t**2) ** (0.5) * e_t
|
||
noise = std_dev_t * randn_tensor(
|
||
clean_latents.shape, dtype=clean_latents.dtype, device=clean_latents.device, generator=generator
|
||
)
|
||
prev_latents = alpha_prod_t_prev ** (0.5) * clean_latents + dir_xt + noise
|
||
|
||
return prev_latents
|
||
|
||
|
||
def compute_noise(scheduler, prev_latents, latents, timestep, noise_pred, eta):
|
||
# 1. get previous step value (=t-1)
|
||
prev_timestep = timestep - scheduler.config.num_train_timesteps // scheduler.num_inference_steps
|
||
|
||
# 2. compute alphas, betas
|
||
alpha_prod_t = scheduler.alphas_cumprod[timestep]
|
||
alpha_prod_t_prev = (
|
||
scheduler.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else scheduler.final_alpha_cumprod
|
||
)
|
||
|
||
beta_prod_t = 1 - alpha_prod_t
|
||
|
||
# 3. compute predicted original sample from predicted noise also called
|
||
# "predicted x_0" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
|
||
pred_original_sample = (latents - beta_prod_t ** (0.5) * noise_pred) / alpha_prod_t ** (0.5)
|
||
|
||
# 4. Clip "predicted x_0"
|
||
if scheduler.config.clip_sample:
|
||
pred_original_sample = torch.clamp(pred_original_sample, -1, 1)
|
||
|
||
# 5. compute variance: "sigma_t(η)" -> see formula (16)
|
||
# σ_t = sqrt((1 − α_t−1)/(1 − α_t)) * sqrt(1 − α_t/α_t−1)
|
||
variance = scheduler._get_variance(timestep, prev_timestep)
|
||
std_dev_t = eta * variance ** (0.5)
|
||
|
||
# 6. compute "direction pointing to x_t" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
|
||
pred_sample_direction = (1 - alpha_prod_t_prev - std_dev_t**2) ** (0.5) * noise_pred
|
||
|
||
noise = (prev_latents - (alpha_prod_t_prev ** (0.5) * pred_original_sample + pred_sample_direction)) / (
|
||
variance ** (0.5) * eta
|
||
)
|
||
return noise
|
||
|
||
|
||
class CycleDiffusionPipeline(DiffusionPipeline, TextualInversionLoaderMixin, LoraLoaderMixin):
|
||
r"""
|
||
Pipeline for text-guided image to image generation using Stable Diffusion.
|
||
|
||
This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
|
||
implemented for all pipelines (downloading, saving, running on a particular device, etc.).
|
||
|
||
Args:
|
||
vae ([`AutoencoderKL`]):
|
||
Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations.
|
||
text_encoder ([`~transformers.CLIPTextModel`]):
|
||
Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
|
||
tokenizer ([`~transformers.CLIPTokenizer`]):
|
||
A `CLIPTokenizer` to tokenize text.
|
||
unet ([`UNet2DConditionModel`]):
|
||
A `UNet2DConditionModel` to denoise the encoded image latents.
|
||
scheduler ([`SchedulerMixin`]):
|
||
A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can only be an
|
||
instance of [`DDIMScheduler`].
|
||
safety_checker ([`StableDiffusionSafetyChecker`]):
|
||
Classification module that estimates whether generated images could be considered offensive or harmful.
|
||
Please refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for more details
|
||
about a model's potential harms.
|
||
feature_extractor ([`~transformers.CLIPImageProcessor`]):
|
||
A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`.
|
||
"""
|
||
_optional_components = ["safety_checker", "feature_extractor"]
|
||
|
||
def __init__(
|
||
self,
|
||
vae: AutoencoderKL,
|
||
text_encoder: CLIPTextModel,
|
||
tokenizer: CLIPTokenizer,
|
||
unet: UNet2DConditionModel,
|
||
scheduler: DDIMScheduler,
|
||
safety_checker: StableDiffusionSafetyChecker,
|
||
feature_extractor: CLIPImageProcessor,
|
||
requires_safety_checker: bool = True,
|
||
):
|
||
super().__init__()
|
||
|
||
if hasattr(scheduler.config, "steps_offset") and scheduler.config.steps_offset != 1:
|
||
deprecation_message = (
|
||
f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`"
|
||
f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure "
|
||
"to update the config accordingly as leaving `steps_offset` might led to incorrect results"
|
||
" in future versions. If you have downloaded this checkpoint from the Hugging Face Hub,"
|
||
" it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`"
|
||
" file"
|
||
)
|
||
deprecate("steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False)
|
||
new_config = dict(scheduler.config)
|
||
new_config["steps_offset"] = 1
|
||
scheduler._internal_dict = FrozenDict(new_config)
|
||
|
||
if safety_checker is None and requires_safety_checker:
|
||
logger.warning(
|
||
f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
|
||
" that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
|
||
" results in services or applications open to the public. Both the diffusers team and Hugging Face"
|
||
" strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
|
||
" it only for use-cases that involve analyzing network behavior or auditing its results. For more"
|
||
" information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
|
||
)
|
||
|
||
if safety_checker is not None and feature_extractor is None:
|
||
raise ValueError(
|
||
"Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety"
|
||
" checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead."
|
||
)
|
||
is_unet_version_less_0_9_0 = hasattr(unet.config, "_diffusers_version") and version.parse(
|
||
version.parse(unet.config._diffusers_version).base_version
|
||
) < version.parse("0.9.0.dev0")
|
||
is_unet_sample_size_less_64 = hasattr(unet.config, "sample_size") and unet.config.sample_size < 64
|
||
if is_unet_version_less_0_9_0 and is_unet_sample_size_less_64:
|
||
deprecation_message = (
|
||
"The configuration file of the unet has set the default `sample_size` to smaller than"
|
||
" 64 which seems highly unlikely .If you're checkpoint is a fine-tuned version of any of the"
|
||
" following: \n- CompVis/stable-diffusion-v1-4 \n- CompVis/stable-diffusion-v1-3 \n-"
|
||
" CompVis/stable-diffusion-v1-2 \n- CompVis/stable-diffusion-v1-1 \n- runwayml/stable-diffusion-v1-5"
|
||
" \n- runwayml/stable-diffusion-inpainting \n you should change 'sample_size' to 64 in the"
|
||
" configuration file. Please make sure to update the config accordingly as leaving `sample_size=32`"
|
||
" in the config might lead to incorrect results in future versions. If you have downloaded this"
|
||
" checkpoint from the Hugging Face Hub, it would be very nice if you could open a Pull request for"
|
||
" the `unet/config.json` file"
|
||
)
|
||
deprecate("sample_size<64", "1.0.0", deprecation_message, standard_warn=False)
|
||
new_config = dict(unet.config)
|
||
new_config["sample_size"] = 64
|
||
unet._internal_dict = FrozenDict(new_config)
|
||
|
||
self.register_modules(
|
||
vae=vae,
|
||
text_encoder=text_encoder,
|
||
tokenizer=tokenizer,
|
||
unet=unet,
|
||
scheduler=scheduler,
|
||
safety_checker=safety_checker,
|
||
feature_extractor=feature_extractor,
|
||
)
|
||
self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
|
||
self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
|
||
self.register_to_config(requires_safety_checker=requires_safety_checker)
|
||
|
||
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_model_cpu_offload
|
||
def enable_model_cpu_offload(self, gpu_id=0):
|
||
r"""
|
||
Offload all models to CPU to reduce memory usage with a low impact on performance. Moves one whole model at a
|
||
time to the GPU when its `forward` method is called, and the model remains in GPU until the next model runs.
|
||
Memory savings are lower than using `enable_sequential_cpu_offload`, but performance is much better due to the
|
||
iterative execution of the `unet`.
|
||
"""
|
||
if is_accelerate_available() and is_accelerate_version(">=", "0.17.0.dev0"):
|
||
from accelerate import cpu_offload_with_hook
|
||
else:
|
||
raise ImportError("`enable_model_cpu_offload` requires `accelerate v0.17.0` or higher.")
|
||
|
||
device = torch.device(f"cuda:{gpu_id}")
|
||
|
||
if self.device.type != "cpu":
|
||
self.to("cpu", silence_dtype_warnings=True)
|
||
torch.cuda.empty_cache() # otherwise we don't see the memory savings (but they probably exist)
|
||
|
||
hook = None
|
||
for cpu_offloaded_model in [self.text_encoder, self.unet, self.vae]:
|
||
_, hook = cpu_offload_with_hook(cpu_offloaded_model, device, prev_module_hook=hook)
|
||
|
||
if self.safety_checker is not None:
|
||
_, hook = cpu_offload_with_hook(self.safety_checker, device, prev_module_hook=hook)
|
||
|
||
# We'll offload the last model manually.
|
||
self.final_offload_hook = hook
|
||
|
||
# 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,
|
||
lora_scale: Optional[float] = 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. 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.
|
||
lora_scale (`float`, *optional*):
|
||
A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
|
||
"""
|
||
# set lora scale so that monkey patched LoRA
|
||
# function of text encoder can correctly access it
|
||
if lora_scale is not None and isinstance(self, LoraLoaderMixin):
|
||
self._lora_scale = lora_scale
|
||
|
||
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:
|
||
# textual inversion: procecss multi-vector tokens if necessary
|
||
if isinstance(self, TextualInversionLoaderMixin):
|
||
prompt = self.maybe_convert_prompt(prompt, self.tokenizer)
|
||
|
||
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 prompt is not None and 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
|
||
|
||
# textual inversion: procecss multi-vector tokens if necessary
|
||
if isinstance(self, TextualInversionLoaderMixin):
|
||
uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer)
|
||
|
||
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_img2img.StableDiffusionImg2ImgPipeline.check_inputs
|
||
def check_inputs(
|
||
self, prompt, strength, callback_steps, negative_prompt=None, prompt_embeds=None, negative_prompt_embeds=None
|
||
):
|
||
if strength < 0 or strength > 1:
|
||
raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}")
|
||
|
||
if (callback_steps is None) or (
|
||
callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
|
||
):
|
||
raise ValueError(
|
||
f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
|
||
f" {type(callback_steps)}."
|
||
)
|
||
|
||
if prompt is not None and prompt_embeds is not None:
|
||
raise ValueError(
|
||
f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
|
||
" only forward one of the two."
|
||
)
|
||
elif prompt is None and prompt_embeds is None:
|
||
raise ValueError(
|
||
"Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
|
||
)
|
||
elif 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(
|
||
f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
|
||
f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
|
||
)
|
||
|
||
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}."
|
||
)
|
||
|
||
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
|
||
def prepare_extra_step_kwargs(self, generator, eta):
|
||
# prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
|
||
# eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
|
||
# eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
|
||
# and should be between [0, 1]
|
||
|
||
accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
|
||
extra_step_kwargs = {}
|
||
if accepts_eta:
|
||
extra_step_kwargs["eta"] = eta
|
||
|
||
# check if the scheduler accepts generator
|
||
accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
|
||
if accepts_generator:
|
||
extra_step_kwargs["generator"] = generator
|
||
return extra_step_kwargs
|
||
|
||
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker
|
||
def run_safety_checker(self, image, device, dtype):
|
||
if self.safety_checker is None:
|
||
has_nsfw_concept = None
|
||
else:
|
||
if torch.is_tensor(image):
|
||
feature_extractor_input = self.image_processor.postprocess(image, output_type="pil")
|
||
else:
|
||
feature_extractor_input = self.image_processor.numpy_to_pil(image)
|
||
safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device)
|
||
image, has_nsfw_concept = self.safety_checker(
|
||
images=image, clip_input=safety_checker_input.pixel_values.to(dtype)
|
||
)
|
||
return image, has_nsfw_concept
|
||
|
||
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents
|
||
def decode_latents(self, latents):
|
||
warnings.warn(
|
||
"The decode_latents method is deprecated and will be removed in a future version. Please"
|
||
" use VaeImageProcessor instead",
|
||
FutureWarning,
|
||
)
|
||
latents = 1 / self.vae.config.scaling_factor * latents
|
||
image = self.vae.decode(latents, return_dict=False)[0]
|
||
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_img2img.StableDiffusionImg2ImgPipeline.get_timesteps
|
||
def get_timesteps(self, num_inference_steps, strength, device):
|
||
# get the original timestep using init_timestep
|
||
init_timestep = min(int(num_inference_steps * strength), num_inference_steps)
|
||
|
||
t_start = max(num_inference_steps - init_timestep, 0)
|
||
timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :]
|
||
|
||
return timesteps, num_inference_steps - t_start
|
||
|
||
def prepare_latents(self, image, timestep, batch_size, num_images_per_prompt, dtype, device, generator=None):
|
||
image = image.to(device=device, dtype=dtype)
|
||
|
||
batch_size = image.shape[0]
|
||
|
||
if image.shape[1] == 4:
|
||
init_latents = image
|
||
|
||
else:
|
||
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 isinstance(generator, list):
|
||
init_latents = [
|
||
self.vae.encode(image[i : i + 1]).latent_dist.sample(generator[i]) for i in range(batch_size)
|
||
]
|
||
init_latents = torch.cat(init_latents, dim=0)
|
||
else:
|
||
init_latents = self.vae.encode(image).latent_dist.sample(generator)
|
||
|
||
init_latents = self.vae.config.scaling_factor * init_latents
|
||
|
||
if batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] == 0:
|
||
# expand init_latents for batch_size
|
||
deprecation_message = (
|
||
f"You have passed {batch_size} text prompts (`prompt`), but only {init_latents.shape[0]} initial"
|
||
" images (`image`). Initial images are now duplicating to match the number of text prompts. Note"
|
||
" that this behavior is deprecated and will be removed in a version 1.0.0. Please make sure to update"
|
||
" your script to pass as many initial images as text prompts to suppress this warning."
|
||
)
|
||
deprecate("len(prompt) != len(image)", "1.0.0", deprecation_message, standard_warn=False)
|
||
additional_image_per_prompt = batch_size // init_latents.shape[0]
|
||
init_latents = torch.cat([init_latents] * additional_image_per_prompt * num_images_per_prompt, dim=0)
|
||
elif batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] != 0:
|
||
raise ValueError(
|
||
f"Cannot duplicate `image` of batch size {init_latents.shape[0]} to {batch_size} text prompts."
|
||
)
|
||
else:
|
||
init_latents = torch.cat([init_latents] * num_images_per_prompt, dim=0)
|
||
|
||
# add noise to latents using the timestep
|
||
shape = init_latents.shape
|
||
noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
|
||
|
||
# get latents
|
||
clean_latents = init_latents
|
||
init_latents = self.scheduler.add_noise(init_latents, noise, timestep)
|
||
latents = init_latents
|
||
|
||
return latents, clean_latents
|
||
|
||
@torch.no_grad()
|
||
def __call__(
|
||
self,
|
||
prompt: Union[str, List[str]],
|
||
source_prompt: Union[str, List[str]],
|
||
image: Union[
|
||
torch.FloatTensor,
|
||
PIL.Image.Image,
|
||
np.ndarray,
|
||
List[torch.FloatTensor],
|
||
List[PIL.Image.Image],
|
||
List[np.ndarray],
|
||
] = None,
|
||
strength: float = 0.8,
|
||
num_inference_steps: Optional[int] = 50,
|
||
guidance_scale: Optional[float] = 7.5,
|
||
source_guidance_scale: Optional[float] = 1,
|
||
num_images_per_prompt: Optional[int] = 1,
|
||
eta: Optional[float] = 0.1,
|
||
generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
|
||
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: int = 1,
|
||
cross_attention_kwargs: Optional[Dict[str, Any]] = None,
|
||
):
|
||
r"""
|
||
The call function to the pipeline for generation.
|
||
|
||
Args:
|
||
prompt (`str` or `List[str]`):
|
||
The prompt or prompts to guide the image generation.
|
||
image (`torch.FloatTensor` `np.ndarray`, `PIL.Image.Image`, `List[torch.FloatTensor]`, `List[PIL.Image.Image]`, or `List[np.ndarray]`):
|
||
`Image` or tensor representing an image batch to be used as the starting point. Can also accept image
|
||
latents as `image`, but if passing latents directly it is not encoded again.
|
||
strength (`float`, *optional*, defaults to 0.8):
|
||
Indicates extent to transform the reference `image`. Must be between 0 and 1. `image` is used as a
|
||
starting point and more noise is added the higher the `strength`. The number of denoising steps depends
|
||
on the amount of noise initially added. When `strength` is 1, added noise is maximum and the denoising
|
||
process runs for the full number of iterations specified in `num_inference_steps`. A value of 1
|
||
essentially ignores `image`.
|
||
num_inference_steps (`int`, *optional*, defaults to 50):
|
||
The number of denoising steps. More denoising steps usually lead to a higher quality image at the
|
||
expense of slower inference. This parameter is modulated by `strength`.
|
||
guidance_scale (`float`, *optional*, defaults to 7.5):
|
||
A higher guidance scale value encourages the model to generate images closely linked to the text
|
||
`prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
|
||
source_guidance_scale (`float`, *optional*, defaults to 1):
|
||
Guidance scale for the source prompt. This is useful to control the amount of influence the source
|
||
prompt has for encoding.
|
||
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) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
|
||
to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
|
||
generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
|
||
A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
|
||
generation deterministic.
|
||
prompt_embeds (`torch.FloatTensor`, *optional*):
|
||
Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
|
||
provided, text embeddings are generated from the `prompt` input argument.
|
||
negative_prompt_embeds (`torch.FloatTensor`, *optional*):
|
||
Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
|
||
not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
|
||
output_type (`str`, *optional*, defaults to `"pil"`):
|
||
The output format of the generated image. Choose between `PIL.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 calls every `callback_steps` steps during inference. The function is 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 is called. If not specified, the callback is called at
|
||
every step.
|
||
cross_attention_kwargs (`dict`, *optional*):
|
||
A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in
|
||
[`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/cross_attention.py).
|
||
|
||
Example:
|
||
|
||
```py
|
||
import requests
|
||
import torch
|
||
from PIL import Image
|
||
from io import BytesIO
|
||
|
||
from diffusers import CycleDiffusionPipeline, DDIMScheduler
|
||
|
||
# load the pipeline
|
||
# make sure you're logged in with `huggingface-cli login`
|
||
model_id_or_path = "CompVis/stable-diffusion-v1-4"
|
||
scheduler = DDIMScheduler.from_pretrained(model_id_or_path, subfolder="scheduler")
|
||
pipe = CycleDiffusionPipeline.from_pretrained(model_id_or_path, scheduler=scheduler).to("cuda")
|
||
|
||
# let's download an initial image
|
||
url = "https://raw.githubusercontent.com/ChenWu98/cycle-diffusion/main/data/dalle2/An%20astronaut%20riding%20a%20horse.png"
|
||
response = requests.get(url)
|
||
init_image = Image.open(BytesIO(response.content)).convert("RGB")
|
||
init_image = init_image.resize((512, 512))
|
||
init_image.save("horse.png")
|
||
|
||
# let's specify a prompt
|
||
source_prompt = "An astronaut riding a horse"
|
||
prompt = "An astronaut riding an elephant"
|
||
|
||
# call the pipeline
|
||
image = pipe(
|
||
prompt=prompt,
|
||
source_prompt=source_prompt,
|
||
image=init_image,
|
||
num_inference_steps=100,
|
||
eta=0.1,
|
||
strength=0.8,
|
||
guidance_scale=2,
|
||
source_guidance_scale=1,
|
||
).images[0]
|
||
|
||
image.save("horse_to_elephant.png")
|
||
|
||
# let's try another example
|
||
# See more samples at the original repo: https://github.com/ChenWu98/cycle-diffusion
|
||
url = (
|
||
"https://raw.githubusercontent.com/ChenWu98/cycle-diffusion/main/data/dalle2/A%20black%20colored%20car.png"
|
||
)
|
||
response = requests.get(url)
|
||
init_image = Image.open(BytesIO(response.content)).convert("RGB")
|
||
init_image = init_image.resize((512, 512))
|
||
init_image.save("black.png")
|
||
|
||
source_prompt = "A black colored car"
|
||
prompt = "A blue colored car"
|
||
|
||
# call the pipeline
|
||
torch.manual_seed(0)
|
||
image = pipe(
|
||
prompt=prompt,
|
||
source_prompt=source_prompt,
|
||
image=init_image,
|
||
num_inference_steps=100,
|
||
eta=0.1,
|
||
strength=0.85,
|
||
guidance_scale=3,
|
||
source_guidance_scale=1,
|
||
).images[0]
|
||
|
||
image.save("black_to_blue.png")
|
||
```
|
||
|
||
Returns:
|
||
[`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
|
||
If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned,
|
||
otherwise a `tuple` is returned where the first element is a list with the generated images and the
|
||
second element is a list of `bool`s indicating whether the corresponding generated image contains
|
||
"not-safe-for-work" (nsfw) content.
|
||
"""
|
||
# 1. Check inputs
|
||
self.check_inputs(prompt, strength, callback_steps)
|
||
|
||
# 2. Define call parameters
|
||
batch_size = 1 if isinstance(prompt, str) else len(prompt)
|
||
device = self._execution_device
|
||
# here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
|
||
# of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
|
||
# corresponds to doing no classifier free guidance.
|
||
do_classifier_free_guidance = guidance_scale > 1.0
|
||
|
||
# 3. Encode input prompt
|
||
text_encoder_lora_scale = (
|
||
cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None
|
||
)
|
||
prompt_embeds = self._encode_prompt(
|
||
prompt,
|
||
device,
|
||
num_images_per_prompt,
|
||
do_classifier_free_guidance,
|
||
prompt_embeds=prompt_embeds,
|
||
lora_scale=text_encoder_lora_scale,
|
||
)
|
||
source_prompt_embeds = self._encode_prompt(
|
||
source_prompt, device, num_images_per_prompt, do_classifier_free_guidance, None
|
||
)
|
||
|
||
# 4. Preprocess image
|
||
image = self.image_processor.preprocess(image)
|
||
|
||
# 5. Prepare timesteps
|
||
self.scheduler.set_timesteps(num_inference_steps, device=device)
|
||
timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength, device)
|
||
latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt)
|
||
|
||
# 6. Prepare latent variables
|
||
latents, clean_latents = self.prepare_latents(
|
||
image, latent_timestep, batch_size, num_images_per_prompt, prompt_embeds.dtype, device, generator
|
||
)
|
||
source_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)
|
||
generator = extra_step_kwargs.pop("generator", None)
|
||
|
||
# 8. Denoising loop
|
||
num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
|
||
with self.progress_bar(total=num_inference_steps) as progress_bar:
|
||
for i, t in enumerate(timesteps):
|
||
# expand the latents if we are doing classifier free guidance
|
||
latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
|
||
source_latent_model_input = (
|
||
torch.cat([source_latents] * 2) if do_classifier_free_guidance else source_latents
|
||
)
|
||
latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
|
||
source_latent_model_input = self.scheduler.scale_model_input(source_latent_model_input, t)
|
||
|
||
# predict the noise residual
|
||
if do_classifier_free_guidance:
|
||
concat_latent_model_input = torch.stack(
|
||
[
|
||
source_latent_model_input[0],
|
||
latent_model_input[0],
|
||
source_latent_model_input[1],
|
||
latent_model_input[1],
|
||
],
|
||
dim=0,
|
||
)
|
||
concat_prompt_embeds = torch.stack(
|
||
[
|
||
source_prompt_embeds[0],
|
||
prompt_embeds[0],
|
||
source_prompt_embeds[1],
|
||
prompt_embeds[1],
|
||
],
|
||
dim=0,
|
||
)
|
||
else:
|
||
concat_latent_model_input = torch.cat(
|
||
[
|
||
source_latent_model_input,
|
||
latent_model_input,
|
||
],
|
||
dim=0,
|
||
)
|
||
concat_prompt_embeds = torch.cat(
|
||
[
|
||
source_prompt_embeds,
|
||
prompt_embeds,
|
||
],
|
||
dim=0,
|
||
)
|
||
|
||
concat_noise_pred = self.unet(
|
||
concat_latent_model_input,
|
||
t,
|
||
cross_attention_kwargs=cross_attention_kwargs,
|
||
encoder_hidden_states=concat_prompt_embeds,
|
||
).sample
|
||
|
||
# perform guidance
|
||
if do_classifier_free_guidance:
|
||
(
|
||
source_noise_pred_uncond,
|
||
noise_pred_uncond,
|
||
source_noise_pred_text,
|
||
noise_pred_text,
|
||
) = concat_noise_pred.chunk(4, dim=0)
|
||
|
||
noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
|
||
source_noise_pred = source_noise_pred_uncond + source_guidance_scale * (
|
||
source_noise_pred_text - source_noise_pred_uncond
|
||
)
|
||
|
||
else:
|
||
(source_noise_pred, noise_pred) = concat_noise_pred.chunk(2, dim=0)
|
||
|
||
# Sample source_latents from the posterior distribution.
|
||
prev_source_latents = posterior_sample(
|
||
self.scheduler, source_latents, t, clean_latents, generator=generator, **extra_step_kwargs
|
||
)
|
||
# Compute noise.
|
||
noise = compute_noise(
|
||
self.scheduler, prev_source_latents, source_latents, t, source_noise_pred, **extra_step_kwargs
|
||
)
|
||
source_latents = prev_source_latents
|
||
|
||
# compute the previous noisy sample x_t -> x_t-1
|
||
latents = self.scheduler.step(
|
||
noise_pred, t, latents, variance_noise=noise, **extra_step_kwargs
|
||
).prev_sample
|
||
|
||
# call the callback, if provided
|
||
if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
|
||
progress_bar.update()
|
||
if callback is not None and i % callback_steps == 0:
|
||
callback(i, t, latents)
|
||
|
||
# 9. Post-processing
|
||
if not output_type == "latent":
|
||
image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
|
||
image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype)
|
||
else:
|
||
image = latents
|
||
has_nsfw_concept = None
|
||
|
||
if has_nsfw_concept is None:
|
||
do_denormalize = [True] * image.shape[0]
|
||
else:
|
||
do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept]
|
||
|
||
image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize)
|
||
|
||
if not return_dict:
|
||
return (image, has_nsfw_concept)
|
||
|
||
return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)
|