mirror of
https://github.com/huggingface/diffusers.git
synced 2026-01-29 07:22:12 +03:00
1427 lines
61 KiB
Python
1427 lines
61 KiB
Python
# Copyright 2025 The HuggingFace Team. All rights reserved.
|
|
#
|
|
# Licensed under the Apache License, Version 2.0 (the "License");
|
|
# you may not use this file except in compliance with the License.
|
|
# You may obtain a copy of the License at
|
|
#
|
|
# http://www.apache.org/licenses/LICENSE-2.0
|
|
#
|
|
# Unless required by applicable law or agreed to in writing, software
|
|
# distributed under the License is distributed on an "AS IS" BASIS,
|
|
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
# See the License for the specific language governing permissions and
|
|
# limitations under the License.
|
|
|
|
from __future__ import annotations
|
|
|
|
import abc
|
|
import inspect
|
|
from typing import Any, Callable, Dict, List, Optional, Tuple, Union
|
|
|
|
import numpy as np
|
|
import torch
|
|
import torch.nn.functional as F
|
|
from packaging import version
|
|
from transformers import (
|
|
CLIPImageProcessor,
|
|
CLIPTextModel,
|
|
CLIPTokenizer,
|
|
CLIPVisionModelWithProjection,
|
|
)
|
|
|
|
from diffusers import AutoencoderKL, DiffusionPipeline, UNet2DConditionModel
|
|
from diffusers.configuration_utils import FrozenDict, deprecate
|
|
from diffusers.image_processor import VaeImageProcessor
|
|
from diffusers.loaders import (
|
|
FromSingleFileMixin,
|
|
IPAdapterMixin,
|
|
StableDiffusionLoraLoaderMixin,
|
|
TextualInversionLoaderMixin,
|
|
)
|
|
from diffusers.models.attention import Attention
|
|
from diffusers.models.lora import adjust_lora_scale_text_encoder
|
|
from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput
|
|
from diffusers.pipelines.stable_diffusion.safety_checker import (
|
|
StableDiffusionSafetyChecker,
|
|
)
|
|
from diffusers.schedulers import KarrasDiffusionSchedulers
|
|
from diffusers.utils import (
|
|
USE_PEFT_BACKEND,
|
|
logging,
|
|
scale_lora_layers,
|
|
unscale_lora_layers,
|
|
)
|
|
from diffusers.utils.torch_utils import randn_tensor
|
|
|
|
|
|
logger = logging.get_logger(__name__)
|
|
|
|
|
|
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.rescale_noise_cfg
|
|
def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0):
|
|
"""
|
|
Rescale `noise_cfg` according to `guidance_rescale`. Based on findings of [Common Diffusion Noise Schedules and
|
|
Sample Steps are Flawed](https://huggingface.co/papers/2305.08891). See Section 3.4
|
|
"""
|
|
std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True)
|
|
std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True)
|
|
# rescale the results from guidance (fixes overexposure)
|
|
noise_pred_rescaled = noise_cfg * (std_text / std_cfg)
|
|
# mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images
|
|
noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg
|
|
return noise_cfg
|
|
|
|
|
|
class Prompt2PromptPipeline(
|
|
DiffusionPipeline,
|
|
TextualInversionLoaderMixin,
|
|
StableDiffusionLoraLoaderMixin,
|
|
IPAdapterMixin,
|
|
FromSingleFileMixin,
|
|
):
|
|
r"""
|
|
Pipeline for text-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.).
|
|
|
|
The pipeline also inherits the following loading methods:
|
|
- [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
|
|
- [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`] for loading LoRA weights
|
|
- [`~loaders.StableDiffusionLoraLoaderMixin.save_lora_weights`] for saving LoRA weights
|
|
- [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files
|
|
- [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters
|
|
|
|
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 be one of
|
|
[`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
|
|
safety_checker ([`StableDiffusionSafetyChecker`]):
|
|
Classification module that estimates whether generated images could be considered offensive or harmful.
|
|
Please refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for 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`.
|
|
"""
|
|
|
|
model_cpu_offload_seq = "text_encoder->image_encoder->unet->vae"
|
|
_exclude_from_cpu_offload = ["safety_checker"]
|
|
_callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"]
|
|
_optional_components = ["safety_checker", "feature_extractor"]
|
|
|
|
def __init__(
|
|
self,
|
|
vae: AutoencoderKL,
|
|
text_encoder: CLIPTextModel,
|
|
tokenizer: CLIPTokenizer,
|
|
unet: UNet2DConditionModel,
|
|
scheduler: KarrasDiffusionSchedulers,
|
|
safety_checker: StableDiffusionSafetyChecker,
|
|
feature_extractor: CLIPImageProcessor,
|
|
image_encoder: CLIPVisionModelWithProjection = None,
|
|
requires_safety_checker: bool = True,
|
|
):
|
|
super().__init__()
|
|
|
|
if scheduler is not None and getattr(scheduler.config, "steps_offset", 1) != 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 scheduler is not None and getattr(scheduler.config, "clip_sample", False) is True:
|
|
deprecation_message = (
|
|
f"The configuration file of this scheduler: {scheduler} has not set the configuration `clip_sample`."
|
|
" `clip_sample` should be set to False in the configuration file. Please make sure to update the"
|
|
" config accordingly as not setting `clip_sample` in the config might lead to incorrect results in"
|
|
" future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it would be very"
|
|
" nice if you could open a Pull request for the `scheduler/scheduler_config.json` file"
|
|
)
|
|
deprecate("clip_sample not set", "1.0.0", deprecation_message, standard_warn=False)
|
|
new_config = dict(scheduler.config)
|
|
new_config["clip_sample"] = False
|
|
scheduler._internal_dict = FrozenDict(new_config)
|
|
|
|
if safety_checker is None 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 = (
|
|
unet is not None
|
|
and 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 = (
|
|
unet is not None and 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 your 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,
|
|
image_encoder=image_encoder,
|
|
)
|
|
self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8
|
|
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._encode_prompt
|
|
def _encode_prompt(
|
|
self,
|
|
prompt,
|
|
device,
|
|
num_images_per_prompt,
|
|
do_classifier_free_guidance,
|
|
negative_prompt=None,
|
|
prompt_embeds: Optional[torch.Tensor] = None,
|
|
negative_prompt_embeds: Optional[torch.Tensor] = None,
|
|
lora_scale: Optional[float] = None,
|
|
**kwargs,
|
|
):
|
|
deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple."
|
|
deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False)
|
|
|
|
prompt_embeds_tuple = 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,
|
|
lora_scale=lora_scale,
|
|
**kwargs,
|
|
)
|
|
|
|
# concatenate for backwards comp
|
|
prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]])
|
|
|
|
return prompt_embeds
|
|
|
|
# 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.Tensor] = None,
|
|
negative_prompt_embeds: Optional[torch.Tensor] = None,
|
|
lora_scale: Optional[float] = None,
|
|
clip_skip: Optional[int] = 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.Tensor`, *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.Tensor`, *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.
|
|
clip_skip (`int`, *optional*):
|
|
Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
|
|
the output of the pre-final layer will be used for computing the prompt embeddings.
|
|
"""
|
|
# 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, StableDiffusionLoraLoaderMixin):
|
|
self._lora_scale = lora_scale
|
|
|
|
# dynamically adjust the LoRA scale
|
|
if not USE_PEFT_BACKEND:
|
|
adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
|
|
else:
|
|
scale_lora_layers(self.text_encoder, 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: process 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
|
|
|
|
if clip_skip is None:
|
|
prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask)
|
|
prompt_embeds = prompt_embeds[0]
|
|
else:
|
|
prompt_embeds = self.text_encoder(
|
|
text_input_ids.to(device),
|
|
attention_mask=attention_mask,
|
|
output_hidden_states=True,
|
|
)
|
|
# Access the `hidden_states` first, that contains a tuple of
|
|
# all the hidden states from the encoder layers. Then index into
|
|
# the tuple to access the hidden states from the desired layer.
|
|
prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)]
|
|
# We also need to apply the final LayerNorm here to not mess with the
|
|
# representations. The `last_hidden_states` that we typically use for
|
|
# obtaining the final prompt representations passes through the LayerNorm
|
|
# layer.
|
|
prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds)
|
|
|
|
if self.text_encoder is not None:
|
|
prompt_embeds_dtype = self.text_encoder.dtype
|
|
elif self.unet is not None:
|
|
prompt_embeds_dtype = self.unet.dtype
|
|
else:
|
|
prompt_embeds_dtype = prompt_embeds.dtype
|
|
|
|
prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_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: process 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=prompt_embeds_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)
|
|
|
|
if isinstance(self, StableDiffusionLoraLoaderMixin) and USE_PEFT_BACKEND:
|
|
# Retrieve the original scale by scaling back the LoRA layers
|
|
unscale_lora_layers(self.text_encoder, lora_scale)
|
|
|
|
return prompt_embeds, negative_prompt_embeds
|
|
|
|
# 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.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://huggingface.co/papers/2010.02502
|
|
# and should be between [0, 1]
|
|
|
|
accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
|
|
extra_step_kwargs = {}
|
|
if accepts_eta:
|
|
extra_step_kwargs["eta"] = eta
|
|
|
|
# check if the scheduler accepts generator
|
|
accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
|
|
if accepts_generator:
|
|
extra_step_kwargs["generator"] = generator
|
|
return extra_step_kwargs
|
|
|
|
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.check_inputs
|
|
def check_inputs(
|
|
self,
|
|
prompt,
|
|
height,
|
|
width,
|
|
callback_steps,
|
|
negative_prompt=None,
|
|
prompt_embeds=None,
|
|
negative_prompt_embeds=None,
|
|
ip_adapter_image=None,
|
|
ip_adapter_image_embeds=None,
|
|
callback_on_step_end_tensor_inputs=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 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 callback_on_step_end_tensor_inputs is not None and not all(
|
|
k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
|
|
):
|
|
raise ValueError(
|
|
f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
|
|
)
|
|
|
|
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}."
|
|
)
|
|
|
|
if ip_adapter_image is not None and ip_adapter_image_embeds is not None:
|
|
raise ValueError(
|
|
"Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined."
|
|
)
|
|
|
|
# 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
|
|
|
|
@torch.no_grad()
|
|
def __call__(
|
|
self,
|
|
prompt: Union[str, List[str]],
|
|
height: Optional[int] = None,
|
|
width: Optional[int] = None,
|
|
num_inference_steps: int = 50,
|
|
guidance_scale: float = 7.5,
|
|
negative_prompt: Optional[Union[str, List[str]]] = None,
|
|
num_images_per_prompt: Optional[int] = 1,
|
|
eta: float = 0.0,
|
|
generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
|
|
latents: Optional[torch.Tensor] = None,
|
|
prompt_embeds: Optional[torch.Tensor] = None,
|
|
negative_prompt_embeds: Optional[torch.Tensor] = None,
|
|
output_type: Optional[str] = "pil",
|
|
return_dict: bool = True,
|
|
callback: Optional[Callable[[int, int, torch.Tensor], None]] = None,
|
|
callback_steps: Optional[int] = 1,
|
|
cross_attention_kwargs: Optional[Dict[str, Any]] = None,
|
|
guidance_rescale: float = 0.0,
|
|
):
|
|
r"""
|
|
Function invoked when calling the pipeline for generation.
|
|
|
|
Args:
|
|
prompt (`str` or `List[str]`):
|
|
The prompt or prompts to guide the image generation.
|
|
height (`int`, *optional*, defaults to 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 50):
|
|
The number of denoising steps. More denoising steps usually lead to a higher quality image at the
|
|
expense of slower inference.
|
|
guidance_scale (`float`, *optional*, defaults to 7.5):
|
|
Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://huggingface.co/papers/2207.12598).
|
|
`guidance_scale` is defined as `w` of equation 2. of [Imagen
|
|
Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting `guidance_scale >
|
|
1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
|
|
usually at the expense of lower image quality.
|
|
negative_prompt (`str` or `List[str]`, *optional*):
|
|
The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
|
|
if `guidance_scale` is less than `1`).
|
|
num_images_per_prompt (`int`, *optional*, defaults to 1):
|
|
The number of images to generate per prompt.
|
|
eta (`float`, *optional*, defaults to 0.0):
|
|
Corresponds to parameter eta (η) in the DDIM paper: https://huggingface.co/papers/2010.02502. Only applies to
|
|
[`schedulers.DDIMScheduler`], will be ignored for others.
|
|
generator (`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.Tensor`, *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 be generated by sampling using the supplied random `generator`.
|
|
output_type (`str`, *optional*, defaults to `"pil"`):
|
|
The output format of the generate image. Choose between
|
|
[PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
|
|
return_dict (`bool`, *optional*, defaults to `True`):
|
|
Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
|
|
plain tuple.
|
|
callback (`Callable`, *optional*):
|
|
A function that will be called every `callback_steps` steps during inference. The function will be
|
|
called with the following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`.
|
|
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 [`AttentionProcessor`] as defined in
|
|
[`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
|
|
|
|
The keyword arguments to configure the edit are:
|
|
- edit_type (`str`). The edit type to apply. Can be either of `replace`, `refine`, `reweight`.
|
|
- n_cross_replace (`int`): Number of diffusion steps in which cross attention should be replaced
|
|
- n_self_replace (`int`): Number of diffusion steps in which self attention should be replaced
|
|
- local_blend_words(`List[str]`, *optional*, default to `None`): Determines which area should be
|
|
changed. If None, then the whole image can be changed.
|
|
- equalizer_words(`List[str]`, *optional*, default to `None`): Required for edit type `reweight`.
|
|
Determines which words should be enhanced.
|
|
- equalizer_strengths (`List[float]`, *optional*, default to `None`) Required for edit type `reweight`.
|
|
Determines which how much the words in `equalizer_words` should be enhanced.
|
|
|
|
guidance_rescale (`float`, *optional*, defaults to 0.0):
|
|
Guidance rescale factor from [Common Diffusion Noise Schedules and Sample Steps are
|
|
Flawed](https://huggingface.co/papers/2305.08891). Guidance rescale factor should fix overexposure when
|
|
using zero terminal SNR.
|
|
|
|
Returns:
|
|
[`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
|
|
[`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple.
|
|
When returning a tuple, the first element is a list with the generated images, and the second element is a
|
|
list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work"
|
|
(nsfw) content, according to the `safety_checker`.
|
|
"""
|
|
|
|
self.controller = create_controller(
|
|
prompt,
|
|
cross_attention_kwargs,
|
|
num_inference_steps,
|
|
tokenizer=self.tokenizer,
|
|
device=self.device,
|
|
)
|
|
self.register_attention_control(self.controller) # add attention controller
|
|
|
|
# 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, height, width, callback_steps)
|
|
|
|
# 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]
|
|
|
|
device = self._execution_device
|
|
# here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
|
|
# of the Imagen paper: https://huggingface.co/papers/2205.11487 . `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,
|
|
negative_prompt,
|
|
prompt_embeds=prompt_embeds,
|
|
negative_prompt_embeds=negative_prompt_embeds,
|
|
lora_scale=text_encoder_lora_scale,
|
|
)
|
|
|
|
# 4. Prepare timesteps
|
|
self.scheduler.set_timesteps(num_inference_steps, device=device)
|
|
timesteps = self.scheduler.timesteps
|
|
|
|
# 5. Prepare latent variables
|
|
num_channels_latents = self.unet.config.in_channels
|
|
latents = self.prepare_latents(
|
|
batch_size * num_images_per_prompt,
|
|
num_channels_latents,
|
|
height,
|
|
width,
|
|
prompt_embeds.dtype,
|
|
device,
|
|
generator,
|
|
latents,
|
|
)
|
|
|
|
# 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
|
|
extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
|
|
|
|
# 7. Denoising loop
|
|
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
|
|
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).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)
|
|
|
|
if do_classifier_free_guidance and guidance_rescale > 0.0:
|
|
# Based on 3.4. in https://huggingface.co/papers/2305.08891
|
|
noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=guidance_rescale)
|
|
|
|
# compute the previous noisy sample x_t -> x_t-1
|
|
latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
|
|
|
|
# step callback
|
|
latents = self.controller.step_callback(latents)
|
|
|
|
# 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:
|
|
step_idx = i // getattr(self.scheduler, "order", 1)
|
|
callback(step_idx, t, latents)
|
|
|
|
# 8. 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
|
|
|
|
# 9. Run safety checker
|
|
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)
|
|
|
|
# Offload last model to CPU
|
|
if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
|
|
self.final_offload_hook.offload()
|
|
|
|
if not return_dict:
|
|
return (image, has_nsfw_concept)
|
|
|
|
return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)
|
|
|
|
def register_attention_control(self, controller):
|
|
attn_procs = {}
|
|
cross_att_count = 0
|
|
for name in self.unet.attn_processors.keys():
|
|
(None if name.endswith("attn1.processor") else self.unet.config.cross_attention_dim)
|
|
if name.startswith("mid_block"):
|
|
self.unet.config.block_out_channels[-1]
|
|
place_in_unet = "mid"
|
|
elif name.startswith("up_blocks"):
|
|
block_id = int(name[len("up_blocks.")])
|
|
list(reversed(self.unet.config.block_out_channels))[block_id]
|
|
place_in_unet = "up"
|
|
elif name.startswith("down_blocks"):
|
|
block_id = int(name[len("down_blocks.")])
|
|
self.unet.config.block_out_channels[block_id]
|
|
place_in_unet = "down"
|
|
else:
|
|
continue
|
|
cross_att_count += 1
|
|
attn_procs[name] = P2PCrossAttnProcessor(controller=controller, place_in_unet=place_in_unet)
|
|
|
|
self.unet.set_attn_processor(attn_procs)
|
|
controller.num_att_layers = cross_att_count
|
|
|
|
|
|
class P2PCrossAttnProcessor:
|
|
def __init__(self, controller, place_in_unet):
|
|
super().__init__()
|
|
self.controller = controller
|
|
self.place_in_unet = place_in_unet
|
|
|
|
def __call__(
|
|
self,
|
|
attn: Attention,
|
|
hidden_states,
|
|
encoder_hidden_states=None,
|
|
attention_mask=None,
|
|
):
|
|
batch_size, sequence_length, _ = hidden_states.shape
|
|
attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
|
|
|
|
query = attn.to_q(hidden_states)
|
|
|
|
is_cross = encoder_hidden_states is not None
|
|
encoder_hidden_states = encoder_hidden_states if encoder_hidden_states is not None else hidden_states
|
|
key = attn.to_k(encoder_hidden_states)
|
|
value = attn.to_v(encoder_hidden_states)
|
|
|
|
query = attn.head_to_batch_dim(query)
|
|
key = attn.head_to_batch_dim(key)
|
|
value = attn.head_to_batch_dim(value)
|
|
|
|
attention_probs = attn.get_attention_scores(query, key, attention_mask)
|
|
|
|
# one line change
|
|
self.controller(attention_probs, is_cross, self.place_in_unet)
|
|
|
|
hidden_states = torch.bmm(attention_probs, value)
|
|
hidden_states = attn.batch_to_head_dim(hidden_states)
|
|
|
|
# linear proj
|
|
hidden_states = attn.to_out[0](hidden_states)
|
|
# dropout
|
|
hidden_states = attn.to_out[1](hidden_states)
|
|
|
|
return hidden_states
|
|
|
|
|
|
def create_controller(
|
|
prompts: List[str],
|
|
cross_attention_kwargs: Dict,
|
|
num_inference_steps: int,
|
|
tokenizer,
|
|
device,
|
|
) -> AttentionControl:
|
|
edit_type = cross_attention_kwargs.get("edit_type", None)
|
|
local_blend_words = cross_attention_kwargs.get("local_blend_words", None)
|
|
equalizer_words = cross_attention_kwargs.get("equalizer_words", None)
|
|
equalizer_strengths = cross_attention_kwargs.get("equalizer_strengths", None)
|
|
n_cross_replace = cross_attention_kwargs.get("n_cross_replace", 0.4)
|
|
n_self_replace = cross_attention_kwargs.get("n_self_replace", 0.4)
|
|
|
|
# only replace
|
|
if edit_type == "replace" and local_blend_words is None:
|
|
return AttentionReplace(
|
|
prompts,
|
|
num_inference_steps,
|
|
n_cross_replace,
|
|
n_self_replace,
|
|
tokenizer=tokenizer,
|
|
device=device,
|
|
)
|
|
|
|
# replace + localblend
|
|
if edit_type == "replace" and local_blend_words is not None:
|
|
lb = LocalBlend(prompts, local_blend_words, tokenizer=tokenizer, device=device)
|
|
return AttentionReplace(
|
|
prompts,
|
|
num_inference_steps,
|
|
n_cross_replace,
|
|
n_self_replace,
|
|
lb,
|
|
tokenizer=tokenizer,
|
|
device=device,
|
|
)
|
|
|
|
# only refine
|
|
if edit_type == "refine" and local_blend_words is None:
|
|
return AttentionRefine(
|
|
prompts,
|
|
num_inference_steps,
|
|
n_cross_replace,
|
|
n_self_replace,
|
|
tokenizer=tokenizer,
|
|
device=device,
|
|
)
|
|
|
|
# refine + localblend
|
|
if edit_type == "refine" and local_blend_words is not None:
|
|
lb = LocalBlend(prompts, local_blend_words, tokenizer=tokenizer, device=device)
|
|
return AttentionRefine(
|
|
prompts,
|
|
num_inference_steps,
|
|
n_cross_replace,
|
|
n_self_replace,
|
|
lb,
|
|
tokenizer=tokenizer,
|
|
device=device,
|
|
)
|
|
|
|
# reweight
|
|
if edit_type == "reweight":
|
|
assert equalizer_words is not None and equalizer_strengths is not None, (
|
|
"To use reweight edit, please specify equalizer_words and equalizer_strengths."
|
|
)
|
|
assert len(equalizer_words) == len(equalizer_strengths), (
|
|
"equalizer_words and equalizer_strengths must be of same length."
|
|
)
|
|
equalizer = get_equalizer(prompts[1], equalizer_words, equalizer_strengths, tokenizer=tokenizer)
|
|
return AttentionReweight(
|
|
prompts,
|
|
num_inference_steps,
|
|
n_cross_replace,
|
|
n_self_replace,
|
|
tokenizer=tokenizer,
|
|
device=device,
|
|
equalizer=equalizer,
|
|
)
|
|
|
|
raise ValueError(f"Edit type {edit_type} not recognized. Use one of: replace, refine, reweight.")
|
|
|
|
|
|
class AttentionControl(abc.ABC):
|
|
def step_callback(self, x_t):
|
|
return x_t
|
|
|
|
def between_steps(self):
|
|
return
|
|
|
|
@property
|
|
def num_uncond_att_layers(self):
|
|
return 0
|
|
|
|
@abc.abstractmethod
|
|
def forward(self, attn, is_cross: bool, place_in_unet: str):
|
|
raise NotImplementedError
|
|
|
|
def __call__(self, attn, is_cross: bool, place_in_unet: str):
|
|
if self.cur_att_layer >= self.num_uncond_att_layers:
|
|
h = attn.shape[0]
|
|
attn[h // 2 :] = self.forward(attn[h // 2 :], is_cross, place_in_unet)
|
|
self.cur_att_layer += 1
|
|
if self.cur_att_layer == self.num_att_layers + self.num_uncond_att_layers:
|
|
self.cur_att_layer = 0
|
|
self.cur_step += 1
|
|
self.between_steps()
|
|
return attn
|
|
|
|
def reset(self):
|
|
self.cur_step = 0
|
|
self.cur_att_layer = 0
|
|
|
|
def __init__(self):
|
|
self.cur_step = 0
|
|
self.num_att_layers = -1
|
|
self.cur_att_layer = 0
|
|
|
|
|
|
class EmptyControl(AttentionControl):
|
|
def forward(self, attn, is_cross: bool, place_in_unet: str):
|
|
return attn
|
|
|
|
|
|
class AttentionStore(AttentionControl):
|
|
@staticmethod
|
|
def get_empty_store():
|
|
return {
|
|
"down_cross": [],
|
|
"mid_cross": [],
|
|
"up_cross": [],
|
|
"down_self": [],
|
|
"mid_self": [],
|
|
"up_self": [],
|
|
}
|
|
|
|
def forward(self, attn, is_cross: bool, place_in_unet: str):
|
|
key = f"{place_in_unet}_{'cross' if is_cross else 'self'}"
|
|
if attn.shape[1] <= 32**2: # avoid memory overhead
|
|
self.step_store[key].append(attn)
|
|
return attn
|
|
|
|
def between_steps(self):
|
|
if len(self.attention_store) == 0:
|
|
self.attention_store = self.step_store
|
|
else:
|
|
for key in self.attention_store:
|
|
for i in range(len(self.attention_store[key])):
|
|
self.attention_store[key][i] += self.step_store[key][i]
|
|
self.step_store = self.get_empty_store()
|
|
|
|
def get_average_attention(self):
|
|
average_attention = {
|
|
key: [item / self.cur_step for item in self.attention_store[key]] for key in self.attention_store
|
|
}
|
|
return average_attention
|
|
|
|
def reset(self):
|
|
super(AttentionStore, self).reset()
|
|
self.step_store = self.get_empty_store()
|
|
self.attention_store = {}
|
|
|
|
def __init__(self):
|
|
super(AttentionStore, self).__init__()
|
|
self.step_store = self.get_empty_store()
|
|
self.attention_store = {}
|
|
|
|
|
|
class LocalBlend:
|
|
def __call__(self, x_t, attention_store):
|
|
k = 1
|
|
maps = attention_store["down_cross"][2:4] + attention_store["up_cross"][:3]
|
|
maps = [item.reshape(self.alpha_layers.shape[0], -1, 1, 16, 16, self.max_num_words) for item in maps]
|
|
maps = torch.cat(maps, dim=1)
|
|
maps = (maps * self.alpha_layers).sum(-1).mean(1)
|
|
mask = F.max_pool2d(maps, (k * 2 + 1, k * 2 + 1), (1, 1), padding=(k, k))
|
|
mask = F.interpolate(mask, size=(x_t.shape[2:]))
|
|
mask = mask / mask.max(2, keepdims=True)[0].max(3, keepdims=True)[0]
|
|
mask = mask.gt(self.threshold)
|
|
mask = (mask[:1] + mask[1:]).float()
|
|
x_t = x_t[:1] + mask * (x_t - x_t[:1])
|
|
return x_t
|
|
|
|
def __init__(
|
|
self,
|
|
prompts: List[str],
|
|
words: [List[List[str]]],
|
|
tokenizer,
|
|
device,
|
|
threshold=0.3,
|
|
max_num_words=77,
|
|
):
|
|
self.max_num_words = 77
|
|
|
|
alpha_layers = torch.zeros(len(prompts), 1, 1, 1, 1, self.max_num_words)
|
|
for i, (prompt, words_) in enumerate(zip(prompts, words)):
|
|
if isinstance(words_, str):
|
|
words_ = [words_]
|
|
for word in words_:
|
|
ind = get_word_inds(prompt, word, tokenizer)
|
|
alpha_layers[i, :, :, :, :, ind] = 1
|
|
self.alpha_layers = alpha_layers.to(device)
|
|
self.threshold = threshold
|
|
|
|
|
|
class AttentionControlEdit(AttentionStore, abc.ABC):
|
|
def step_callback(self, x_t):
|
|
if self.local_blend is not None:
|
|
x_t = self.local_blend(x_t, self.attention_store)
|
|
return x_t
|
|
|
|
def replace_self_attention(self, attn_base, att_replace):
|
|
if att_replace.shape[2] <= 16**2:
|
|
return attn_base.unsqueeze(0).expand(att_replace.shape[0], *attn_base.shape)
|
|
else:
|
|
return att_replace
|
|
|
|
@abc.abstractmethod
|
|
def replace_cross_attention(self, attn_base, att_replace):
|
|
raise NotImplementedError
|
|
|
|
def forward(self, attn, is_cross: bool, place_in_unet: str):
|
|
super(AttentionControlEdit, self).forward(attn, is_cross, place_in_unet)
|
|
# FIXME not replace correctly
|
|
if is_cross or (self.num_self_replace[0] <= self.cur_step < self.num_self_replace[1]):
|
|
h = attn.shape[0] // (self.batch_size)
|
|
attn = attn.reshape(self.batch_size, h, *attn.shape[1:])
|
|
attn_base, attn_repalce = attn[0], attn[1:]
|
|
if is_cross:
|
|
alpha_words = self.cross_replace_alpha[self.cur_step]
|
|
attn_repalce_new = (
|
|
self.replace_cross_attention(attn_base, attn_repalce) * alpha_words
|
|
+ (1 - alpha_words) * attn_repalce
|
|
)
|
|
attn[1:] = attn_repalce_new
|
|
else:
|
|
attn[1:] = self.replace_self_attention(attn_base, attn_repalce)
|
|
attn = attn.reshape(self.batch_size * h, *attn.shape[2:])
|
|
return attn
|
|
|
|
def __init__(
|
|
self,
|
|
prompts,
|
|
num_steps: int,
|
|
cross_replace_steps: Union[float, Tuple[float, float], Dict[str, Tuple[float, float]]],
|
|
self_replace_steps: Union[float, Tuple[float, float]],
|
|
local_blend: Optional[LocalBlend],
|
|
tokenizer,
|
|
device,
|
|
):
|
|
super(AttentionControlEdit, self).__init__()
|
|
# add tokenizer and device here
|
|
|
|
self.tokenizer = tokenizer
|
|
self.device = device
|
|
|
|
self.batch_size = len(prompts)
|
|
self.cross_replace_alpha = get_time_words_attention_alpha(
|
|
prompts, num_steps, cross_replace_steps, self.tokenizer
|
|
).to(self.device)
|
|
if isinstance(self_replace_steps, float):
|
|
self_replace_steps = 0, self_replace_steps
|
|
self.num_self_replace = int(num_steps * self_replace_steps[0]), int(num_steps * self_replace_steps[1])
|
|
self.local_blend = local_blend # 在外面定义后传进来
|
|
|
|
|
|
class AttentionReplace(AttentionControlEdit):
|
|
def replace_cross_attention(self, attn_base, att_replace):
|
|
return torch.einsum("hpw,bwn->bhpn", attn_base, self.mapper)
|
|
|
|
def __init__(
|
|
self,
|
|
prompts,
|
|
num_steps: int,
|
|
cross_replace_steps: float,
|
|
self_replace_steps: float,
|
|
local_blend: Optional[LocalBlend] = None,
|
|
tokenizer=None,
|
|
device=None,
|
|
):
|
|
super(AttentionReplace, self).__init__(
|
|
prompts,
|
|
num_steps,
|
|
cross_replace_steps,
|
|
self_replace_steps,
|
|
local_blend,
|
|
tokenizer,
|
|
device,
|
|
)
|
|
self.mapper = get_replacement_mapper(prompts, self.tokenizer).to(self.device)
|
|
|
|
|
|
class AttentionRefine(AttentionControlEdit):
|
|
def replace_cross_attention(self, attn_base, att_replace):
|
|
attn_base_replace = attn_base[:, :, self.mapper].permute(2, 0, 1, 3)
|
|
attn_replace = attn_base_replace * self.alphas + att_replace * (1 - self.alphas)
|
|
return attn_replace
|
|
|
|
def __init__(
|
|
self,
|
|
prompts,
|
|
num_steps: int,
|
|
cross_replace_steps: float,
|
|
self_replace_steps: float,
|
|
local_blend: Optional[LocalBlend] = None,
|
|
tokenizer=None,
|
|
device=None,
|
|
):
|
|
super(AttentionRefine, self).__init__(
|
|
prompts,
|
|
num_steps,
|
|
cross_replace_steps,
|
|
self_replace_steps,
|
|
local_blend,
|
|
tokenizer,
|
|
device,
|
|
)
|
|
self.mapper, alphas = get_refinement_mapper(prompts, self.tokenizer)
|
|
self.mapper, alphas = self.mapper.to(self.device), alphas.to(self.device)
|
|
self.alphas = alphas.reshape(alphas.shape[0], 1, 1, alphas.shape[1])
|
|
|
|
|
|
class AttentionReweight(AttentionControlEdit):
|
|
def replace_cross_attention(self, attn_base, att_replace):
|
|
if self.prev_controller is not None:
|
|
attn_base = self.prev_controller.replace_cross_attention(attn_base, att_replace)
|
|
attn_replace = attn_base[None, :, :, :] * self.equalizer[:, None, None, :]
|
|
return attn_replace
|
|
|
|
def __init__(
|
|
self,
|
|
prompts,
|
|
num_steps: int,
|
|
cross_replace_steps: float,
|
|
self_replace_steps: float,
|
|
equalizer,
|
|
local_blend: Optional[LocalBlend] = None,
|
|
controller: Optional[AttentionControlEdit] = None,
|
|
tokenizer=None,
|
|
device=None,
|
|
):
|
|
super(AttentionReweight, self).__init__(
|
|
prompts,
|
|
num_steps,
|
|
cross_replace_steps,
|
|
self_replace_steps,
|
|
local_blend,
|
|
tokenizer,
|
|
device,
|
|
)
|
|
self.equalizer = equalizer.to(self.device)
|
|
self.prev_controller = controller
|
|
|
|
|
|
### util functions for all Edits
|
|
def update_alpha_time_word(
|
|
alpha,
|
|
bounds: Union[float, Tuple[float, float]],
|
|
prompt_ind: int,
|
|
word_inds: Optional[torch.Tensor] = None,
|
|
):
|
|
if isinstance(bounds, float):
|
|
bounds = 0, bounds
|
|
start, end = int(bounds[0] * alpha.shape[0]), int(bounds[1] * alpha.shape[0])
|
|
if word_inds is None:
|
|
word_inds = torch.arange(alpha.shape[2])
|
|
alpha[:start, prompt_ind, word_inds] = 0
|
|
alpha[start:end, prompt_ind, word_inds] = 1
|
|
alpha[end:, prompt_ind, word_inds] = 0
|
|
return alpha
|
|
|
|
|
|
def get_time_words_attention_alpha(
|
|
prompts,
|
|
num_steps,
|
|
cross_replace_steps: Union[float, Dict[str, Tuple[float, float]]],
|
|
tokenizer,
|
|
max_num_words=77,
|
|
):
|
|
if not isinstance(cross_replace_steps, dict):
|
|
cross_replace_steps = {"default_": cross_replace_steps}
|
|
if "default_" not in cross_replace_steps:
|
|
cross_replace_steps["default_"] = (0.0, 1.0)
|
|
alpha_time_words = torch.zeros(num_steps + 1, len(prompts) - 1, max_num_words)
|
|
for i in range(len(prompts) - 1):
|
|
alpha_time_words = update_alpha_time_word(alpha_time_words, cross_replace_steps["default_"], i)
|
|
for key, item in cross_replace_steps.items():
|
|
if key != "default_":
|
|
inds = [get_word_inds(prompts[i], key, tokenizer) for i in range(1, len(prompts))]
|
|
for i, ind in enumerate(inds):
|
|
if len(ind) > 0:
|
|
alpha_time_words = update_alpha_time_word(alpha_time_words, item, i, ind)
|
|
alpha_time_words = alpha_time_words.reshape(num_steps + 1, len(prompts) - 1, 1, 1, max_num_words)
|
|
return alpha_time_words
|
|
|
|
|
|
### util functions for LocalBlend and ReplacementEdit
|
|
def get_word_inds(text: str, word_place: int, tokenizer):
|
|
split_text = text.split(" ")
|
|
if isinstance(word_place, str):
|
|
word_place = [i for i, word in enumerate(split_text) if word_place == word]
|
|
elif isinstance(word_place, int):
|
|
word_place = [word_place]
|
|
out = []
|
|
if len(word_place) > 0:
|
|
words_encode = [tokenizer.decode([item]).strip("#") for item in tokenizer.encode(text)][1:-1]
|
|
cur_len, ptr = 0, 0
|
|
|
|
for i in range(len(words_encode)):
|
|
cur_len += len(words_encode[i])
|
|
if ptr in word_place:
|
|
out.append(i + 1)
|
|
if cur_len >= len(split_text[ptr]):
|
|
ptr += 1
|
|
cur_len = 0
|
|
return np.array(out)
|
|
|
|
|
|
### util functions for ReplacementEdit
|
|
def get_replacement_mapper_(x: str, y: str, tokenizer, max_len=77):
|
|
words_x = x.split(" ")
|
|
words_y = y.split(" ")
|
|
if len(words_x) != len(words_y):
|
|
raise ValueError(
|
|
f"attention replacement edit can only be applied on prompts with the same length"
|
|
f" but prompt A has {len(words_x)} words and prompt B has {len(words_y)} words."
|
|
)
|
|
inds_replace = [i for i in range(len(words_y)) if words_y[i] != words_x[i]]
|
|
inds_source = [get_word_inds(x, i, tokenizer) for i in inds_replace]
|
|
inds_target = [get_word_inds(y, i, tokenizer) for i in inds_replace]
|
|
mapper = np.zeros((max_len, max_len))
|
|
i = j = 0
|
|
cur_inds = 0
|
|
while i < max_len and j < max_len:
|
|
if cur_inds < len(inds_source) and inds_source[cur_inds][0] == i:
|
|
inds_source_, inds_target_ = inds_source[cur_inds], inds_target[cur_inds]
|
|
if len(inds_source_) == len(inds_target_):
|
|
mapper[inds_source_, inds_target_] = 1
|
|
else:
|
|
ratio = 1 / len(inds_target_)
|
|
for i_t in inds_target_:
|
|
mapper[inds_source_, i_t] = ratio
|
|
cur_inds += 1
|
|
i += len(inds_source_)
|
|
j += len(inds_target_)
|
|
elif cur_inds < len(inds_source):
|
|
mapper[i, j] = 1
|
|
i += 1
|
|
j += 1
|
|
else:
|
|
mapper[j, j] = 1
|
|
i += 1
|
|
j += 1
|
|
|
|
return torch.from_numpy(mapper).float()
|
|
|
|
|
|
def get_replacement_mapper(prompts, tokenizer, max_len=77):
|
|
x_seq = prompts[0]
|
|
mappers = []
|
|
for i in range(1, len(prompts)):
|
|
mapper = get_replacement_mapper_(x_seq, prompts[i], tokenizer, max_len)
|
|
mappers.append(mapper)
|
|
return torch.stack(mappers)
|
|
|
|
|
|
### util functions for ReweightEdit
|
|
def get_equalizer(
|
|
text: str,
|
|
word_select: Union[int, Tuple[int, ...]],
|
|
values: Union[List[float], Tuple[float, ...]],
|
|
tokenizer,
|
|
):
|
|
if isinstance(word_select, (int, str)):
|
|
word_select = (word_select,)
|
|
equalizer = torch.ones(len(values), 77)
|
|
values = torch.tensor(values, dtype=torch.float32)
|
|
for word in word_select:
|
|
inds = get_word_inds(text, word, tokenizer)
|
|
equalizer[:, inds] = values
|
|
return equalizer
|
|
|
|
|
|
### util functions for RefinementEdit
|
|
class ScoreParams:
|
|
def __init__(self, gap, match, mismatch):
|
|
self.gap = gap
|
|
self.match = match
|
|
self.mismatch = mismatch
|
|
|
|
def mis_match_char(self, x, y):
|
|
if x != y:
|
|
return self.mismatch
|
|
else:
|
|
return self.match
|
|
|
|
|
|
def get_matrix(size_x, size_y, gap):
|
|
matrix = np.zeros((size_x + 1, size_y + 1), dtype=np.int32)
|
|
matrix[0, 1:] = (np.arange(size_y) + 1) * gap
|
|
matrix[1:, 0] = (np.arange(size_x) + 1) * gap
|
|
return matrix
|
|
|
|
|
|
def get_traceback_matrix(size_x, size_y):
|
|
matrix = np.zeros((size_x + 1, size_y + 1), dtype=np.int32)
|
|
matrix[0, 1:] = 1
|
|
matrix[1:, 0] = 2
|
|
matrix[0, 0] = 4
|
|
return matrix
|
|
|
|
|
|
def global_align(x, y, score):
|
|
matrix = get_matrix(len(x), len(y), score.gap)
|
|
trace_back = get_traceback_matrix(len(x), len(y))
|
|
for i in range(1, len(x) + 1):
|
|
for j in range(1, len(y) + 1):
|
|
left = matrix[i, j - 1] + score.gap
|
|
up = matrix[i - 1, j] + score.gap
|
|
diag = matrix[i - 1, j - 1] + score.mis_match_char(x[i - 1], y[j - 1])
|
|
matrix[i, j] = max(left, up, diag)
|
|
if matrix[i, j] == left:
|
|
trace_back[i, j] = 1
|
|
elif matrix[i, j] == up:
|
|
trace_back[i, j] = 2
|
|
else:
|
|
trace_back[i, j] = 3
|
|
return matrix, trace_back
|
|
|
|
|
|
def get_aligned_sequences(x, y, trace_back):
|
|
x_seq = []
|
|
y_seq = []
|
|
i = len(x)
|
|
j = len(y)
|
|
mapper_y_to_x = []
|
|
while i > 0 or j > 0:
|
|
if trace_back[i, j] == 3:
|
|
x_seq.append(x[i - 1])
|
|
y_seq.append(y[j - 1])
|
|
i = i - 1
|
|
j = j - 1
|
|
mapper_y_to_x.append((j, i))
|
|
elif trace_back[i][j] == 1:
|
|
x_seq.append("-")
|
|
y_seq.append(y[j - 1])
|
|
j = j - 1
|
|
mapper_y_to_x.append((j, -1))
|
|
elif trace_back[i][j] == 2:
|
|
x_seq.append(x[i - 1])
|
|
y_seq.append("-")
|
|
i = i - 1
|
|
elif trace_back[i][j] == 4:
|
|
break
|
|
mapper_y_to_x.reverse()
|
|
return x_seq, y_seq, torch.tensor(mapper_y_to_x, dtype=torch.int64)
|
|
|
|
|
|
def get_mapper(x: str, y: str, tokenizer, max_len=77):
|
|
x_seq = tokenizer.encode(x)
|
|
y_seq = tokenizer.encode(y)
|
|
score = ScoreParams(0, 1, -1)
|
|
matrix, trace_back = global_align(x_seq, y_seq, score)
|
|
mapper_base = get_aligned_sequences(x_seq, y_seq, trace_back)[-1]
|
|
alphas = torch.ones(max_len)
|
|
alphas[: mapper_base.shape[0]] = mapper_base[:, 1].ne(-1).float()
|
|
mapper = torch.zeros(max_len, dtype=torch.int64)
|
|
mapper[: mapper_base.shape[0]] = mapper_base[:, 1]
|
|
mapper[mapper_base.shape[0] :] = len(y_seq) + torch.arange(max_len - len(y_seq))
|
|
return mapper, alphas
|
|
|
|
|
|
def get_refinement_mapper(prompts, tokenizer, max_len=77):
|
|
x_seq = prompts[0]
|
|
mappers, alphas = [], []
|
|
for i in range(1, len(prompts)):
|
|
mapper, alpha = get_mapper(x_seq, prompts[i], tokenizer, max_len)
|
|
mappers.append(mapper)
|
|
alphas.append(alpha)
|
|
return torch.stack(mappers), torch.stack(alphas)
|