Skip to content

GitLab

Commit 0a1fba1c authored by chenpangpang's avatar chenpangpang
Browse files

feat: 初始提交

parent 2ed6785e
Pipeline #1561 failed with stages
in 0 seconds
Hide whitespace changes
Inline Side-by-side
Showing with 3787 additions and 0 deletions
Kolors-FaceID/kolors/pipelines/pipeline_stable_diffusion_xl_chatglm_256_inpainting.py 0 → 100644
View file @ 0a1fba1c
# Copyright 2024 The HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import inspect
from typing import Any, Callable, Dict, List, Optional, Tuple, Union
import numpy as np
import PIL.Image
import torch
from transformers import (
CLIPImageProcessor,
CLIPTextModel,
CLIPTextModelWithProjection,
CLIPTokenizer,
CLIPVisionModelWithProjection,
)
from diffusers.callbacks import MultiPipelineCallbacks, PipelineCallback
from diffusers.image_processor import PipelineImageInput, VaeImageProcessor
from diffusers.loaders import (
FromSingleFileMixin,
IPAdapterMixin,
StableDiffusionXLLoraLoaderMixin,
TextualInversionLoaderMixin,
)
from diffusers.models import AutoencoderKL, ImageProjection, UNet2DConditionModel
from diffusers.models.attention_processor import (
AttnProcessor2_0,
LoRAAttnProcessor2_0,
LoRAXFormersAttnProcessor,
XFormersAttnProcessor,
)
from diffusers.models.lora import adjust_lora_scale_text_encoder
from diffusers.schedulers import KarrasDiffusionSchedulers
from diffusers.utils import (
USE_PEFT_BACKEND,
deprecate,
is_invisible_watermark_available,
is_torch_xla_available,
logging,
replace_example_docstring,
scale_lora_layers,
unscale_lora_layers,
)
from diffusers.utils.torch_utils import randn_tensor
from diffusers.pipelines.pipeline_utils import DiffusionPipeline, StableDiffusionMixin
from diffusers.pipelines.stable_diffusion_xl.pipeline_output import StableDiffusionXLPipelineOutput
if is_invisible_watermark_available():
from .watermark import StableDiffusionXLWatermarker
if is_torch_xla_available():
import torch_xla.core.xla_model as xm
XLA_AVAILABLE = True
else:
XLA_AVAILABLE = False
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
EXAMPLE_DOC_STRING = """
Examples:
```py
>>> import torch
>>> from diffusers import StableDiffusionXLInpaintPipeline
>>> from diffusers.utils import load_image
>>> pipe = StableDiffusionXLInpaintPipeline.from_pretrained(
... "stabilityai/stable-diffusion-xl-base-1.0",
... torch_dtype=torch.float16,
... variant="fp16",
... use_safetensors=True,
... )
>>> pipe.to("cuda")
>>> img_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo.png"
>>> mask_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo_mask.png"
>>> init_image = load_image(img_url).convert("RGB")
>>> mask_image = load_image(mask_url).convert("RGB")
>>> prompt = "A majestic tiger sitting on a bench"
>>> image = pipe(
... prompt=prompt, image=init_image, mask_image=mask_image, num_inference_steps=50, strength=0.80
... ).images[0]
```
"""
# 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://arxiv.org/pdf/2305.08891.pdf). 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
def mask_pil_to_torch(mask, height, width):
# preprocess mask
if isinstance(mask, (PIL.Image.Image, np.ndarray)):
mask = [mask]
if isinstance(mask, list) and isinstance(mask[0], PIL.Image.Image):
mask = [i.resize((width, height), resample=PIL.Image.LANCZOS) for i in mask]
mask = np.concatenate([np.array(m.convert("L"))[None, None, :] for m in mask], axis=0)
mask = mask.astype(np.float32) / 255.0
elif isinstance(mask, list) and isinstance(mask[0], np.ndarray):
mask = np.concatenate([m[None, None, :] for m in mask], axis=0)
mask = torch.from_numpy(mask)
return mask
def prepare_mask_and_masked_image(image, mask, height, width, return_image: bool = False):
"""
Prepares a pair (image, mask) to be consumed by the Stable Diffusion pipeline. This means that those inputs will be
converted to ``torch.Tensor`` with shapes ``batch x channels x height x width`` where ``channels`` is ``3`` for the
``image`` and ``1`` for the ``mask``.
The ``image`` will be converted to ``torch.float32`` and normalized to be in ``[-1, 1]``. The ``mask`` will be
binarized (``mask > 0.5``) and cast to ``torch.float32`` too.
Args:
image (Union[np.array, PIL.Image, torch.Tensor]): The image to inpaint.
It can be a ``PIL.Image``, or a ``height x width x 3`` ``np.array`` or a ``channels x height x width``
``torch.Tensor`` or a ``batch x channels x height x width`` ``torch.Tensor``.
mask (_type_): The mask to apply to the image, i.e. regions to inpaint.
It can be a ``PIL.Image``, or a ``height x width`` ``np.array`` or a ``1 x height x width``
``torch.Tensor`` or a ``batch x 1 x height x width`` ``torch.Tensor``.
Raises:
ValueError: ``torch.Tensor`` images should be in the ``[-1, 1]`` range. ValueError: ``torch.Tensor`` mask
should be in the ``[0, 1]`` range. ValueError: ``mask`` and ``image`` should have the same spatial dimensions.
TypeError: ``mask`` is a ``torch.Tensor`` but ``image`` is not
(ot the other way around).
Returns:
tuple[torch.Tensor]: The pair (mask, masked_image) as ``torch.Tensor`` with 4
dimensions: ``batch x channels x height x width``.
"""
# checkpoint. TOD(Yiyi) - need to clean this up later
deprecation_message = "The prepare_mask_and_masked_image method is deprecated and will be removed in a future version. Please use VaeImageProcessor.preprocess instead"
deprecate(
"prepare_mask_and_masked_image",
"0.30.0",
deprecation_message,
)
if image is None:
raise ValueError("`image` input cannot be undefined.")
if mask is None:
raise ValueError("`mask_image` input cannot be undefined.")
if isinstance(image, torch.Tensor):
if not isinstance(mask, torch.Tensor):
mask = mask_pil_to_torch(mask, height, width)
if image.ndim == 3:
image = image.unsqueeze(0)
# Batch and add channel dim for single mask
if mask.ndim == 2:
mask = mask.unsqueeze(0).unsqueeze(0)
# Batch single mask or add channel dim
if mask.ndim == 3:
# Single batched mask, no channel dim or single mask not batched but channel dim
if mask.shape[0] == 1:
mask = mask.unsqueeze(0)
# Batched masks no channel dim
else:
mask = mask.unsqueeze(1)
assert image.ndim == 4 and mask.ndim == 4, "Image and Mask must have 4 dimensions"
# assert image.shape[-2:] == mask.shape[-2:], "Image and Mask must have the same spatial dimensions"
assert image.shape[0] == mask.shape[0], "Image and Mask must have the same batch size"
# Check image is in [-1, 1]
# if image.min() < -1 or image.max() > 1:
# raise ValueError("Image should be in [-1, 1] range")
# Check mask is in [0, 1]
if mask.min() < 0 or mask.max() > 1:
raise ValueError("Mask should be in [0, 1] range")
# Binarize mask
mask[mask < 0.5] = 0
mask[mask >= 0.5] = 1
# Image as float32
image = image.to(dtype=torch.float32)
elif isinstance(mask, torch.Tensor):
raise TypeError(f"`mask` is a torch.Tensor but `image` (type: {type(image)} is not")
else:
# preprocess image
if isinstance(image, (PIL.Image.Image, np.ndarray)):
image = [image]
if isinstance(image, list) and isinstance(image[0], PIL.Image.Image):
# resize all images w.r.t passed height an width
image = [i.resize((width, height), resample=PIL.Image.LANCZOS) for i in image]
image = [np.array(i.convert("RGB"))[None, :] for i in image]
image = np.concatenate(image, axis=0)
elif isinstance(image, list) and isinstance(image[0], np.ndarray):
image = np.concatenate([i[None, :] for i in image], axis=0)
image = image.transpose(0, 3, 1, 2)
image = torch.from_numpy(image).to(dtype=torch.float32) / 127.5 - 1.0
mask = mask_pil_to_torch(mask, height, width)
mask[mask < 0.5] = 0
mask[mask >= 0.5] = 1
if image.shape[1] == 4:
# images are in latent space and thus can't
# be masked set masked_image to None
# we assume that the checkpoint is not an inpainting
# checkpoint. TOD(Yiyi) - need to clean this up later
masked_image = None
else:
masked_image = image * (mask < 0.5)
# n.b. ensure backwards compatibility as old function does not return image
if return_image:
return mask, masked_image, image
return mask, masked_image
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents
def retrieve_latents(
encoder_output: torch.Tensor, generator: Optional[torch.Generator] = None, sample_mode: str = "sample"
):
if hasattr(encoder_output, "latent_dist") and sample_mode == "sample":
return encoder_output.latent_dist.sample(generator)
elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax":
return encoder_output.latent_dist.mode()
elif hasattr(encoder_output, "latents"):
return encoder_output.latents
else:
raise AttributeError("Could not access latents of provided encoder_output")
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
def retrieve_timesteps(
scheduler,
num_inference_steps: Optional[int] = None,
device: Optional[Union[str, torch.device]] = None,
timesteps: Optional[List[int]] = None,
sigmas: Optional[List[float]] = None,
**kwargs,
):
"""
Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
Args:
scheduler (`SchedulerMixin`):
The scheduler to get timesteps from.
num_inference_steps (`int`):
The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps`
must be `None`.
device (`str` or `torch.device`, *optional*):
The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
timesteps (`List[int]`, *optional*):
Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed,
`num_inference_steps` and `sigmas` must be `None`.
sigmas (`List[float]`, *optional*):
Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed,
`num_inference_steps` and `timesteps` must be `None`.
Returns:
`Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
second element is the number of inference steps.
"""
if timesteps is not None and sigmas is not None:
raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values")
if timesteps is not None:
accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
if not accepts_timesteps:
raise ValueError(
f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
f" timestep schedules. Please check whether you are using the correct scheduler."
)
scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
timesteps = scheduler.timesteps
num_inference_steps = len(timesteps)
elif sigmas is not None:
accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
if not accept_sigmas:
raise ValueError(
f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
f" sigmas schedules. Please check whether you are using the correct scheduler."
)
scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs)
timesteps = scheduler.timesteps
num_inference_steps = len(timesteps)
else:
scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
timesteps = scheduler.timesteps
return timesteps, num_inference_steps
class StableDiffusionXLInpaintPipeline(
DiffusionPipeline,
StableDiffusionMixin,
TextualInversionLoaderMixin,
StableDiffusionXLLoraLoaderMixin,
FromSingleFileMixin,
IPAdapterMixin,
):
r"""
Pipeline for text-to-image generation using Stable Diffusion XL.
This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
The pipeline also inherits the following loading methods:
- [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
- [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files
- [`~loaders.StableDiffusionXLLoraLoaderMixin.load_lora_weights`] for loading LoRA weights
- [`~loaders.StableDiffusionXLLoraLoaderMixin.save_lora_weights`] for saving LoRA weights
- [`~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 ([`CLIPTextModel`]):
Frozen text-encoder. Stable Diffusion XL uses the text portion of
[CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically
the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant.
text_encoder_2 ([` CLIPTextModelWithProjection`]):
Second frozen text-encoder. Stable Diffusion XL uses the text and pool portion of
[CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection),
specifically the
[laion/CLIP-ViT-bigG-14-laion2B-39B-b160k](https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k)
variant.
tokenizer (`CLIPTokenizer`):
Tokenizer of class
[CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
tokenizer_2 (`CLIPTokenizer`):
Second Tokenizer of class
[CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents.
scheduler ([`SchedulerMixin`]):
A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
[`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
requires_aesthetics_score (`bool`, *optional*, defaults to `"False"`):
Whether the `unet` requires a aesthetic_score condition to be passed during inference. Also see the config
of `stabilityai/stable-diffusion-xl-refiner-1-0`.
force_zeros_for_empty_prompt (`bool`, *optional*, defaults to `"True"`):
Whether the negative prompt embeddings shall be forced to always be set to 0. Also see the config of
`stabilityai/stable-diffusion-xl-base-1-0`.
add_watermarker (`bool`, *optional*):
Whether to use the [invisible_watermark library](https://github.com/ShieldMnt/invisible-watermark/) to
watermark output images. If not defined, it will default to True if the package is installed, otherwise no
watermarker will be used.
"""
model_cpu_offload_seq = "text_encoder->text_encoder_2->image_encoder->unet->vae"
_optional_components = [
"tokenizer",
"tokenizer_2",
"text_encoder",
"text_encoder_2",
"image_encoder",
"feature_extractor",
]
_callback_tensor_inputs = [
"latents",
"prompt_embeds",
"negative_prompt_embeds",
"add_text_embeds",
"add_time_ids",
"negative_pooled_prompt_embeds",
"add_neg_time_ids",
"mask",
"masked_image_latents",
]
def __init__(
self,
vae: AutoencoderKL,
text_encoder: CLIPTextModel,
tokenizer: CLIPTokenizer,
unet: UNet2DConditionModel,
scheduler: KarrasDiffusionSchedulers,
tokenizer_2: CLIPTokenizer = None,
text_encoder_2: CLIPTextModelWithProjection = None,
image_encoder: CLIPVisionModelWithProjection = None,
feature_extractor: CLIPImageProcessor = None,
requires_aesthetics_score: bool = False,
force_zeros_for_empty_prompt: bool = True,
add_watermarker: Optional[bool] = None,
):
super().__init__()
self.register_modules(
vae=vae,
text_encoder=text_encoder,
text_encoder_2=text_encoder_2,
tokenizer=tokenizer,
tokenizer_2=tokenizer_2,
unet=unet,
image_encoder=image_encoder,
feature_extractor=feature_extractor,
scheduler=scheduler,
)
self.register_to_config(force_zeros_for_empty_prompt=force_zeros_for_empty_prompt)
self.register_to_config(requires_aesthetics_score=requires_aesthetics_score)
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.mask_processor = VaeImageProcessor(
vae_scale_factor=self.vae_scale_factor, do_normalize=False, do_binarize=True, do_convert_grayscale=True
)
add_watermarker = add_watermarker if add_watermarker is not None else is_invisible_watermark_available()
if add_watermarker:
self.watermark = StableDiffusionXLWatermarker()
else:
self.watermark = None
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image
def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None):
dtype = next(self.image_encoder.parameters()).dtype
if not isinstance(image, torch.Tensor):
image = self.feature_extractor(image, return_tensors="pt").pixel_values
image = image.to(device=device, dtype=dtype)
if output_hidden_states:
image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2]
image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0)
uncond_image_enc_hidden_states = self.image_encoder(
torch.zeros_like(image), output_hidden_states=True
).hidden_states[-2]
uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave(
num_images_per_prompt, dim=0
)
return image_enc_hidden_states, uncond_image_enc_hidden_states
else:
image_embeds = self.image_encoder(image).image_embeds
image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
uncond_image_embeds = torch.zeros_like(image_embeds)
return image_embeds, uncond_image_embeds
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_ip_adapter_image_embeds
def prepare_ip_adapter_image_embeds(
self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance
):
if ip_adapter_image_embeds is None:
if not isinstance(ip_adapter_image, list):
ip_adapter_image = [ip_adapter_image]
if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers):
raise ValueError(
f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters."
)
image_embeds = []
for single_ip_adapter_image, image_proj_layer in zip(
ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers
):
output_hidden_state = not isinstance(image_proj_layer, ImageProjection)
single_image_embeds, single_negative_image_embeds = self.encode_image(
single_ip_adapter_image, device, 1, output_hidden_state
)
single_image_embeds = torch.stack([single_image_embeds] * num_images_per_prompt, dim=0)
single_negative_image_embeds = torch.stack(
[single_negative_image_embeds] * num_images_per_prompt, dim=0
)
if do_classifier_free_guidance:
single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
single_image_embeds = single_image_embeds.to(device)
image_embeds.append(single_image_embeds)
else:
repeat_dims = [1]
image_embeds = []
for single_image_embeds in ip_adapter_image_embeds:
if do_classifier_free_guidance:
single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2)
single_image_embeds = single_image_embeds.repeat(
num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:]))
)
single_negative_image_embeds = single_negative_image_embeds.repeat(
num_images_per_prompt, *(repeat_dims * len(single_negative_image_embeds.shape[1:]))
)
single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
else:
single_image_embeds = single_image_embeds.repeat(
num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:]))
)
image_embeds.append(single_image_embeds)
return image_embeds
def encode_prompt(
self,
prompt,
device: Optional[torch.device] = None,
num_images_per_prompt: int = 1,
do_classifier_free_guidance: bool = True,
negative_prompt=None,
prompt_embeds: Optional[torch.FloatTensor] = None,
negative_prompt_embeds: Optional[torch.FloatTensor] = None,
pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
negative_pooled_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.
pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
If not provided, pooled text embeddings will be generated from `prompt` input argument.
negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
weighting. If not provided, pooled 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.
"""
# from IPython import embed; embed(); exit()
device = device or self._execution_device
# 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]
# Define tokenizers and text encoders
tokenizers = [self.tokenizer]
text_encoders = [self.text_encoder]
if prompt_embeds is None:
# textual inversion: procecss multi-vector tokens if necessary
prompt_embeds_list = []
for tokenizer, text_encoder in zip(tokenizers, text_encoders):
if isinstance(self, TextualInversionLoaderMixin):
prompt = self.maybe_convert_prompt(prompt, tokenizer)
text_inputs = tokenizer(
prompt,
padding="max_length",
max_length=256,
truncation=True,
return_tensors="pt",
).to('cuda')
output = text_encoder(
input_ids=text_inputs['input_ids'] ,
attention_mask=text_inputs['attention_mask'],
position_ids=text_inputs['position_ids'],
output_hidden_states=True)
prompt_embeds = output.hidden_states[-2].permute(1, 0, 2).clone()
text_proj = output.hidden_states[-1][-1, :, :].clone()
bs_embed, seq_len, _ = prompt_embeds.shape
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)
prompt_embeds_list.append(prompt_embeds)
prompt_embeds = prompt_embeds_list[0]
# get unconditional embeddings for classifier free guidance
zero_out_negative_prompt = negative_prompt is None and self.config.force_zeros_for_empty_prompt
if do_classifier_free_guidance and negative_prompt_embeds is None and zero_out_negative_prompt:
negative_prompt_embeds = torch.zeros_like(prompt_embeds)
negative_pooled_prompt_embeds = torch.zeros_like(pooled_prompt_embeds)
elif do_classifier_free_guidance and negative_prompt_embeds is None:
# negative_prompt = negative_prompt or ""
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
negative_prompt_embeds_list = []
for tokenizer, text_encoder in zip(tokenizers, text_encoders):
# textual inversion: procecss multi-vector tokens if necessary
if isinstance(self, TextualInversionLoaderMixin):
uncond_tokens = self.maybe_convert_prompt(uncond_tokens, tokenizer)
max_length = prompt_embeds.shape[1]
uncond_input = tokenizer(
uncond_tokens,
padding="max_length",
max_length=max_length,
truncation=True,
return_tensors="pt",
).to('cuda')
output = text_encoder(
input_ids=uncond_input['input_ids'] ,
attention_mask=uncond_input['attention_mask'],
position_ids=uncond_input['position_ids'],
output_hidden_states=True)
negative_prompt_embeds = output.hidden_states[-2].permute(1, 0, 2).clone()
negative_text_proj = output.hidden_states[-1][-1, :, :].clone()
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=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
negative_prompt_embeds_list.append(negative_prompt_embeds)
negative_prompt_embeds = negative_prompt_embeds_list[0]
bs_embed = text_proj.shape[0]
text_proj = text_proj.repeat(1, num_images_per_prompt).view(
bs_embed * num_images_per_prompt, -1
)
negative_text_proj = negative_text_proj.repeat(1, num_images_per_prompt).view(
bs_embed * num_images_per_prompt, -1
)
return prompt_embeds, negative_prompt_embeds, text_proj, negative_text_proj
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
def prepare_extra_step_kwargs(self, generator, eta):
# prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
# eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
# eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
# and should be between [0, 1]
accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
extra_step_kwargs = {}
if accepts_eta:
extra_step_kwargs["eta"] = eta
# check if the scheduler accepts generator
accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
if accepts_generator:
extra_step_kwargs["generator"] = generator
return extra_step_kwargs
def check_inputs(
self,
prompt,
prompt_2,
image,
mask_image,
height,
width,
strength,
callback_steps,
output_type,
negative_prompt=None,
negative_prompt_2=None,
prompt_embeds=None,
negative_prompt_embeds=None,
ip_adapter_image=None,
ip_adapter_image_embeds=None,
callback_on_step_end_tensor_inputs=None,
padding_mask_crop=None,
):
if strength < 0 or strength > 1:
raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}")
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_2 is not None and prompt_embeds is not None:
raise ValueError(
f"Cannot forward both `prompt_2`: {prompt_2} 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)}")
elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)):
raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}")
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."
)
elif negative_prompt_2 is not None and negative_prompt_embeds is not None:
raise ValueError(
f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} 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 padding_mask_crop is not None:
if not isinstance(image, PIL.Image.Image):
raise ValueError(
f"The image should be a PIL image when inpainting mask crop, but is of type" f" {type(image)}."
)
if not isinstance(mask_image, PIL.Image.Image):
raise ValueError(
f"The mask image should be a PIL image when inpainting mask crop, but is of type"
f" {type(mask_image)}."
)
if output_type != "pil":
raise ValueError(f"The output type should be PIL when inpainting mask crop, but is" f" {output_type}.")
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."
)
if ip_adapter_image_embeds is not None:
if not isinstance(ip_adapter_image_embeds, list):
raise ValueError(
f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}"
)
elif ip_adapter_image_embeds[0].ndim not in [3, 4]:
raise ValueError(
f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D"
)
def prepare_latents(
self,
batch_size,
num_channels_latents,
height,
width,
dtype,
device,
generator,
latents=None,
image=None,
timestep=None,
is_strength_max=True,
add_noise=True,
return_noise=False,
return_image_latents=False,
):
shape = (
batch_size,
num_channels_latents,
int(height) // self.vae_scale_factor,
int(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 (image is None or timestep is None) and not is_strength_max:
raise ValueError(
"Since strength < 1. initial latents are to be initialised as a combination of Image + Noise."
"However, either the image or the noise timestep has not been provided."
)
if image.shape[1] == 4:
image_latents = image.to(device=device, dtype=dtype)
image_latents = image_latents.repeat(batch_size // image_latents.shape[0], 1, 1, 1)
elif return_image_latents or (latents is None and not is_strength_max):
image = image.to(device=device, dtype=dtype)
image_latents = self._encode_vae_image(image=image, generator=generator)
image_latents = image_latents.repeat(batch_size // image_latents.shape[0], 1, 1, 1)
if latents is None and add_noise:
noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
# if strength is 1. then initialise the latents to noise, else initial to image + noise
latents = noise if is_strength_max else self.scheduler.add_noise(image_latents, noise, timestep)
# if pure noise then scale the initial latents by the Scheduler's init sigma
latents = latents * self.scheduler.init_noise_sigma if is_strength_max else latents
elif add_noise:
noise = latents.to(device)
latents = noise * self.scheduler.init_noise_sigma
else:
noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
latents = image_latents.to(device)
outputs = (latents,)
if return_noise:
outputs += (noise,)
if return_image_latents:
outputs += (image_latents,)
return outputs
def _encode_vae_image(self, image: torch.Tensor, generator: torch.Generator):
dtype = image.dtype
if self.vae.config.force_upcast:
image = image.float()
self.vae.to(dtype=torch.float32)
if isinstance(generator, list):
image_latents = [
retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i])
for i in range(image.shape[0])
]
image_latents = torch.cat(image_latents, dim=0)
else:
image_latents = retrieve_latents(self.vae.encode(image), generator=generator)
if self.vae.config.force_upcast:
self.vae.to(dtype)
image_latents = image_latents.to(dtype)
image_latents = self.vae.config.scaling_factor * image_latents
return image_latents
def prepare_mask_latents(
self, mask, masked_image, batch_size, height, width, dtype, device, generator, do_classifier_free_guidance
):
# resize the mask to latents shape as we concatenate the mask to the latents
# we do that before converting to dtype to avoid breaking in case we're using cpu_offload
# and half precision
mask = torch.nn.functional.interpolate(
mask, size=(height // self.vae_scale_factor, width // self.vae_scale_factor)
)
mask = mask.to(device=device, dtype=dtype)
# duplicate mask and masked_image_latents for each generation per prompt, using mps friendly method
if mask.shape[0] < batch_size:
if not batch_size % mask.shape[0] == 0:
raise ValueError(
"The passed mask and the required batch size don't match. Masks are supposed to be duplicated to"
f" a total batch size of {batch_size}, but {mask.shape[0]} masks were passed. Make sure the number"
" of masks that you pass is divisible by the total requested batch size."
)
mask = mask.repeat(batch_size // mask.shape[0], 1, 1, 1)
mask = torch.cat([mask] * 2) if do_classifier_free_guidance else mask
if masked_image is not None and masked_image.shape[1] == 4:
masked_image_latents = masked_image
else:
masked_image_latents = None
if masked_image is not None:
if masked_image_latents is None:
masked_image = masked_image.to(device=device, dtype=dtype)
masked_image_latents = self._encode_vae_image(masked_image, generator=generator)
if masked_image_latents.shape[0] < batch_size:
if not batch_size % masked_image_latents.shape[0] == 0:
raise ValueError(
"The passed images and the required batch size don't match. Images are supposed to be duplicated"
f" to a total batch size of {batch_size}, but {masked_image_latents.shape[0]} images were passed."
" Make sure the number of images that you pass is divisible by the total requested batch size."
)
masked_image_latents = masked_image_latents.repeat(
batch_size // masked_image_latents.shape[0], 1, 1, 1
)
masked_image_latents = (
torch.cat([masked_image_latents] * 2) if do_classifier_free_guidance else masked_image_latents
)
# aligning device to prevent device errors when concating it with the latent model input
masked_image_latents = masked_image_latents.to(device=device, dtype=dtype)
return mask, masked_image_latents
# Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl_img2img.StableDiffusionXLImg2ImgPipeline.get_timesteps
def get_timesteps(self, num_inference_steps, strength, device, denoising_start=None):
# get the original timestep using init_timestep
if denoising_start is None:
init_timestep = min(int(num_inference_steps * strength), num_inference_steps)
t_start = max(num_inference_steps - init_timestep, 0)
else:
t_start = 0
timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :]
# Strength is irrelevant if we directly request a timestep to start at;
# that is, strength is determined by the denoising_start instead.
if denoising_start is not None:
discrete_timestep_cutoff = int(
round(
self.scheduler.config.num_train_timesteps
- (denoising_start * self.scheduler.config.num_train_timesteps)
)
)
num_inference_steps = (timesteps < discrete_timestep_cutoff).sum().item()
if self.scheduler.order == 2 and num_inference_steps % 2 == 0:
# if the scheduler is a 2nd order scheduler we might have to do +1
# because `num_inference_steps` might be even given that every timestep
# (except the highest one) is duplicated. If `num_inference_steps` is even it would
# mean that we cut the timesteps in the middle of the denoising step
# (between 1st and 2nd derivative) which leads to incorrect results. By adding 1
# we ensure that the denoising process always ends after the 2nd derivate step of the scheduler
num_inference_steps = num_inference_steps + 1
# because t_n+1 >= t_n, we slice the timesteps starting from the end
timesteps = timesteps[-num_inference_steps:]
return timesteps, num_inference_steps
return timesteps, num_inference_steps - t_start
# Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl_img2img.StableDiffusionXLImg2ImgPipeline._get_add_time_ids
def _get_add_time_ids(
self,
original_size,
crops_coords_top_left,
target_size,
aesthetic_score,
negative_aesthetic_score,
negative_original_size,
negative_crops_coords_top_left,
negative_target_size,
dtype,
text_encoder_projection_dim=None,
):
if self.config.requires_aesthetics_score:
add_time_ids = list(original_size + crops_coords_top_left + (aesthetic_score,))
add_neg_time_ids = list(
negative_original_size + negative_crops_coords_top_left + (negative_aesthetic_score,)
)
else:
add_time_ids = list(original_size + crops_coords_top_left + target_size)
add_neg_time_ids = list(negative_original_size + crops_coords_top_left + negative_target_size)
passed_add_embed_dim = (
self.unet.config.addition_time_embed_dim * len(add_time_ids) + 4096
)
expected_add_embed_dim = self.unet.add_embedding.linear_1.in_features
if (
expected_add_embed_dim > passed_add_embed_dim
and (expected_add_embed_dim - passed_add_embed_dim) == self.unet.config.addition_time_embed_dim
):
raise ValueError(
f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. Please make sure to enable `requires_aesthetics_score` with `pipe.register_to_config(requires_aesthetics_score=True)` to make sure `aesthetic_score` {aesthetic_score} and `negative_aesthetic_score` {negative_aesthetic_score} is correctly used by the model."
)
elif (
expected_add_embed_dim < passed_add_embed_dim
and (passed_add_embed_dim - expected_add_embed_dim) == self.unet.config.addition_time_embed_dim
):
raise ValueError(
f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. Please make sure to disable `requires_aesthetics_score` with `pipe.register_to_config(requires_aesthetics_score=False)` to make sure `target_size` {target_size} is correctly used by the model."
)
elif expected_add_embed_dim != passed_add_embed_dim:
raise ValueError(
f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. The model has an incorrect config. Please check `unet.config.time_embedding_type` and `text_encoder_2.config.projection_dim`."
)
add_time_ids = torch.tensor([add_time_ids], dtype=dtype)
add_neg_time_ids = torch.tensor([add_neg_time_ids], dtype=dtype)
return add_time_ids, add_neg_time_ids
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_upscale.StableDiffusionUpscalePipeline.upcast_vae
def upcast_vae(self):
dtype = self.vae.dtype
self.vae.to(dtype=torch.float32)
use_torch_2_0_or_xformers = isinstance(
self.vae.decoder.mid_block.attentions[0].processor,
(
AttnProcessor2_0,
XFormersAttnProcessor,
LoRAXFormersAttnProcessor,
LoRAAttnProcessor2_0,
),
)
# if xformers or torch_2_0 is used attention block does not need
# to be in float32 which can save lots of memory
if use_torch_2_0_or_xformers:
self.vae.post_quant_conv.to(dtype)
self.vae.decoder.conv_in.to(dtype)
self.vae.decoder.mid_block.to(dtype)
# Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding
def get_guidance_scale_embedding(
self, w: torch.Tensor, embedding_dim: int = 512, dtype: torch.dtype = torch.float32
) -> torch.Tensor:
"""
See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298
Args:
w (`torch.Tensor`):
Generate embedding vectors with a specified guidance scale to subsequently enrich timestep embeddings.
embedding_dim (`int`, *optional*, defaults to 512):
Dimension of the embeddings to generate.
dtype (`torch.dtype`, *optional*, defaults to `torch.float32`):
Data type of the generated embeddings.
Returns:
`torch.Tensor`: Embedding vectors with shape `(len(w), embedding_dim)`.
"""
assert len(w.shape) == 1
w = w * 1000.0
half_dim = embedding_dim // 2
emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1)
emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb)
emb = w.to(dtype)[:, None] * emb[None, :]
emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1)
if embedding_dim % 2 == 1: # zero pad
emb = torch.nn.functional.pad(emb, (0, 1))
assert emb.shape == (w.shape[0], embedding_dim)
return emb
@property
def guidance_scale(self):
return self._guidance_scale
@property
def guidance_rescale(self):
return self._guidance_rescale
@property
def clip_skip(self):
return self._clip_skip
# 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.
@property
def do_classifier_free_guidance(self):
return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None
@property
def cross_attention_kwargs(self):
return self._cross_attention_kwargs
@property
def denoising_end(self):
return self._denoising_end
@property
def denoising_start(self):
return self._denoising_start
@property
def num_timesteps(self):
return self._num_timesteps
@property
def interrupt(self):
return self._interrupt
@torch.no_grad()
@replace_example_docstring(EXAMPLE_DOC_STRING)
def __call__(
self,
prompt: Union[str, List[str]] = None,
prompt_2: Optional[Union[str, List[str]]] = None,
image: PipelineImageInput = None,
mask_image: PipelineImageInput = None,
masked_image_latents: torch.Tensor = None,
height: Optional[int] = None,
width: Optional[int] = None,
padding_mask_crop: Optional[int] = None,
strength: float = 0.9999,
num_inference_steps: int = 50,
timesteps: List[int] = None,
sigmas: List[float] = None,
denoising_start: Optional[float] = None,
denoising_end: Optional[float] = None,
guidance_scale: float = 7.5,
negative_prompt: Optional[Union[str, List[str]]] = None,
negative_prompt_2: 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,
pooled_prompt_embeds: Optional[torch.Tensor] = None,
negative_pooled_prompt_embeds: Optional[torch.Tensor] = None,
ip_adapter_image: Optional[PipelineImageInput] = None,
ip_adapter_image_embeds: Optional[List[torch.Tensor]] = None,
output_type: Optional[str] = "pil",
return_dict: bool = True,
cross_attention_kwargs: Optional[Dict[str, Any]] = None,
guidance_rescale: float = 0.0,
original_size: Tuple[int, int] = None,
crops_coords_top_left: Tuple[int, int] = (0, 0),
target_size: Tuple[int, int] = None,
negative_original_size: Optional[Tuple[int, int]] = None,
negative_crops_coords_top_left: Tuple[int, int] = (0, 0),
negative_target_size: Optional[Tuple[int, int]] = None,
aesthetic_score: float = 6.0,
negative_aesthetic_score: float = 2.5,
clip_skip: Optional[int] = None,
callback_on_step_end: Optional[
Union[Callable[[int, int, Dict], None], PipelineCallback, MultiPipelineCallbacks]
] = None,
callback_on_step_end_tensor_inputs: List[str] = ["latents"],
**kwargs,
):
r"""
Function invoked when calling the pipeline for generation.
Args:
prompt (`str` or `List[str]`, *optional*):
The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
instead.
prompt_2 (`str` or `List[str]`, *optional*):
The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
used in both text-encoders
image (`PIL.Image.Image`):
`Image`, or tensor representing an image batch which will be inpainted, *i.e.* parts of the image will
be masked out with `mask_image` and repainted according to `prompt`.
mask_image (`PIL.Image.Image`):
`Image`, or tensor representing an image batch, to mask `image`. White pixels in the mask will be
repainted, while black pixels will be preserved. If `mask_image` is a PIL image, it will be converted
to a single channel (luminance) before use. If it's a tensor, it should contain one color channel (L)
instead of 3, so the expected shape would be `(B, H, W, 1)`.
height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
The height in pixels of the generated image. This is set to 1024 by default for the best results.
Anything below 512 pixels won't work well for
[stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0)
and checkpoints that are not specifically fine-tuned on low resolutions.
width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
The width in pixels of the generated image. This is set to 1024 by default for the best results.
Anything below 512 pixels won't work well for
[stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0)
and checkpoints that are not specifically fine-tuned on low resolutions.
padding_mask_crop (`int`, *optional*, defaults to `None`):
The size of margin in the crop to be applied to the image and masking. If `None`, no crop is applied to
image and mask_image. If `padding_mask_crop` is not `None`, it will first find a rectangular region
with the same aspect ration of the image and contains all masked area, and then expand that area based
on `padding_mask_crop`. The image and mask_image will then be cropped based on the expanded area before
resizing to the original image size for inpainting. This is useful when the masked area is small while
the image is large and contain information irrelevant for inpainting, such as background.
strength (`float`, *optional*, defaults to 0.9999):
Conceptually, indicates how much to transform the masked portion of the reference `image`. Must be
between 0 and 1. `image` will be used as a starting point, adding more noise to it the larger the
`strength`. The number of denoising steps depends on the amount of noise initially added. When
`strength` is 1, added noise will be maximum and the denoising process will run for the full number of
iterations specified in `num_inference_steps`. A value of 1, therefore, essentially ignores the masked
portion of the reference `image`. Note that in the case of `denoising_start` being declared as an
integer, the value of `strength` will be ignored.
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.
timesteps (`List[int]`, *optional*):
Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument
in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is
passed will be used. Must be in descending order.
sigmas (`List[float]`, *optional*):
Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in
their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed
will be used.
denoising_start (`float`, *optional*):
When specified, indicates the fraction (between 0.0 and 1.0) of the total denoising process to be
bypassed before it is initiated. Consequently, the initial part of the denoising process is skipped and
it is assumed that the passed `image` is a partly denoised image. Note that when this is specified,
strength will be ignored. The `denoising_start` parameter is particularly beneficial when this pipeline
is integrated into a "Mixture of Denoisers" multi-pipeline setup, as detailed in [**Refining the Image
Output**](https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion/stable_diffusion_xl#refining-the-image-output).
denoising_end (`float`, *optional*):
When specified, determines the fraction (between 0.0 and 1.0) of the total denoising process to be
completed before it is intentionally prematurely terminated. As a result, the returned sample will
still retain a substantial amount of noise (ca. final 20% of timesteps still needed) and should be
denoised by a successor pipeline that has `denoising_start` set to 0.8 so that it only denoises the
final 20% of the scheduler. The denoising_end parameter should ideally be utilized when this pipeline
forms a part of a "Mixture of Denoisers" multi-pipeline setup, as elaborated in [**Refining the Image
Output**](https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion/stable_diffusion_xl#refining-the-image-output).
guidance_scale (`float`, *optional*, defaults to 7.5):
Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
`guidance_scale` is defined as `w` of equation 2. of [Imagen
Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
usually at the expense of lower image quality.
negative_prompt (`str` or `List[str]`, *optional*):
The prompt or prompts not to guide the image generation. 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`).
negative_prompt_2 (`str` or `List[str]`, *optional*):
The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and
`text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders
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.
pooled_prompt_embeds (`torch.Tensor`, *optional*):
Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
If not provided, pooled text embeddings will be generated from `prompt` input argument.
negative_pooled_prompt_embeds (`torch.Tensor`, *optional*):
Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt`
input argument.
ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters.
ip_adapter_image_embeds (`List[torch.Tensor]`, *optional*):
Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of
IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should
contain the negative image embedding if `do_classifier_free_guidance` is set to `True`. If not
provided, embeddings are computed from the `ip_adapter_image` input argument.
num_images_per_prompt (`int`, *optional*, defaults to 1):
The number of images to generate per prompt.
eta (`float`, *optional*, defaults to 0.0):
Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
[`schedulers.DDIMScheduler`], will be ignored for others.
generator (`torch.Generator`, *optional*):
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 ge generated by sampling using the supplied random `generator`.
output_type (`str`, *optional*, defaults to `"pil"`):
The output format of the generate image. Choose between
[PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
return_dict (`bool`, *optional*, defaults to `True`):
Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
plain tuple.
cross_attention_kwargs (`dict`, *optional*):
A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
`self.processor` in
[diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
If `original_size` is not the same as `target_size` the image will appear to be down- or upsampled.
`original_size` defaults to `(height, width)` if not specified. Part of SDXL's micro-conditioning as
explained in section 2.2 of
[https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)):
`crops_coords_top_left` can be used to generate an image that appears to be "cropped" from the position
`crops_coords_top_left` downwards. Favorable, well-centered images are usually achieved by setting
`crops_coords_top_left` to (0, 0). Part of SDXL's micro-conditioning as explained in section 2.2 of
[https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
For most cases, `target_size` should be set to the desired height and width of the generated image. If
not specified it will default to `(height, width)`. Part of SDXL's micro-conditioning as explained in
section 2.2 of [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
negative_original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
To negatively condition the generation process based on a specific image resolution. Part of SDXL's
micro-conditioning as explained in section 2.2 of
[https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more
information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208.
negative_crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)):
To negatively condition the generation process based on a specific crop coordinates. Part of SDXL's
micro-conditioning as explained in section 2.2 of
[https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more
information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208.
negative_target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
To negatively condition the generation process based on a target image resolution. It should be as same
as the `target_size` for most cases. Part of SDXL's micro-conditioning as explained in section 2.2 of
[https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more
information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208.
aesthetic_score (`float`, *optional*, defaults to 6.0):
Used to simulate an aesthetic score of the generated image by influencing the positive text condition.
Part of SDXL's micro-conditioning as explained in section 2.2 of
[https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
negative_aesthetic_score (`float`, *optional*, defaults to 2.5):
Part of SDXL's micro-conditioning as explained in section 2.2 of
[https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). Can be used to
simulate an aesthetic score of the generated image by influencing the negative text condition.
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.
callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*):
A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of
each denoising step during the inference. with the following arguments: `callback_on_step_end(self:
DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a
list of all tensors as specified by `callback_on_step_end_tensor_inputs`.
callback_on_step_end_tensor_inputs (`List`, *optional*):
The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
`._callback_tensor_inputs` attribute of your pipeline class.
Examples:
Returns:
[`~pipelines.stable_diffusion.StableDiffusionXLPipelineOutput`] or `tuple`:
[`~pipelines.stable_diffusion.StableDiffusionXLPipelineOutput`] if `return_dict` is True, otherwise a
`tuple. `tuple. When returning a tuple, the first element is a list with the generated images.
"""
callback = kwargs.pop("callback", None)
callback_steps = kwargs.pop("callback_steps", None)
if callback is not None:
deprecate(
"callback",
"1.0.0",
"Passing `callback` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`",
)
if callback_steps is not None:
deprecate(
"callback_steps",
"1.0.0",
"Passing `callback_steps` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`",
)
if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)):
callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs
# 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
self.check_inputs(
prompt,
prompt_2,
image,
mask_image,
height,
width,
strength,
callback_steps,
output_type,
negative_prompt,
negative_prompt_2,
prompt_embeds,
negative_prompt_embeds,
ip_adapter_image,
ip_adapter_image_embeds,
callback_on_step_end_tensor_inputs,
padding_mask_crop,
)
self._guidance_scale = guidance_scale
self._guidance_rescale = guidance_rescale
self._clip_skip = clip_skip
self._cross_attention_kwargs = cross_attention_kwargs
self._denoising_end = denoising_end
self._denoising_start = denoising_start
self._interrupt = False
# 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
# 3. Encode input prompt
text_encoder_lora_scale = (
self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None
)
(
prompt_embeds,
negative_prompt_embeds,
pooled_prompt_embeds,
negative_pooled_prompt_embeds,
) = self.encode_prompt(
prompt=prompt,
device=device,
num_images_per_prompt=num_images_per_prompt,
do_classifier_free_guidance=self.do_classifier_free_guidance,
negative_prompt=negative_prompt,
prompt_embeds=prompt_embeds,
negative_prompt_embeds=negative_prompt_embeds,
pooled_prompt_embeds=pooled_prompt_embeds,
negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
lora_scale=text_encoder_lora_scale,
)
# 4. set timesteps
def denoising_value_valid(dnv):
return isinstance(dnv, float) and 0 < dnv < 1
timesteps, num_inference_steps = retrieve_timesteps(
self.scheduler, num_inference_steps, device, timesteps, sigmas
)
timesteps, num_inference_steps = self.get_timesteps(
num_inference_steps,
strength,
device,
denoising_start=self.denoising_start if denoising_value_valid(self.denoising_start) else None,
)
# check that number of inference steps is not < 1 - as this doesn't make sense
if num_inference_steps < 1:
raise ValueError(
f"After adjusting the num_inference_steps by strength parameter: {strength}, the number of pipeline"
f"steps is {num_inference_steps} which is < 1 and not appropriate for this pipeline."
)
# at which timestep to set the initial noise (n.b. 50% if strength is 0.5)
latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt)
# create a boolean to check if the strength is set to 1. if so then initialise the latents with pure noise
is_strength_max = strength == 1.0
# 5. Preprocess mask and image
if padding_mask_crop is not None:
crops_coords = self.mask_processor.get_crop_region(mask_image, width, height, pad=padding_mask_crop)
resize_mode = "fill"
else:
crops_coords = None
resize_mode = "default"
original_image = image
init_image = self.image_processor.preprocess(
image, height=height, width=width, crops_coords=crops_coords, resize_mode=resize_mode
)
init_image = init_image.to(dtype=torch.float32)
mask = self.mask_processor.preprocess(
mask_image, height=height, width=width, resize_mode=resize_mode, crops_coords=crops_coords
)
if masked_image_latents is not None:
masked_image = masked_image_latents
elif init_image.shape[1] == 4:
# if images are in latent space, we can't mask it
masked_image = None
else:
masked_image = init_image * (mask < 0.5)
# 6. Prepare latent variables
num_channels_latents = self.vae.config.latent_channels
num_channels_unet = self.unet.config.in_channels
return_image_latents = num_channels_unet == 4
add_noise = True if self.denoising_start is None else False
latents_outputs = self.prepare_latents(
batch_size * num_images_per_prompt,
num_channels_latents,
height,
width,
prompt_embeds.dtype,
device,
generator,
latents,
image=init_image,
timestep=latent_timestep,
is_strength_max=is_strength_max,
add_noise=add_noise,
return_noise=True,
return_image_latents=return_image_latents,
)
if return_image_latents:
latents, noise, image_latents = latents_outputs
else:
latents, noise = latents_outputs
# 7. Prepare mask latent variables
mask, masked_image_latents = self.prepare_mask_latents(
mask,
masked_image,
batch_size * num_images_per_prompt,
height,
width,
prompt_embeds.dtype,
device,
generator,
self.do_classifier_free_guidance,
)
# 8. Check that sizes of mask, masked image and latents match
if num_channels_unet == 9:
# default case for runwayml/stable-diffusion-inpainting
num_channels_mask = mask.shape[1]
num_channels_masked_image = masked_image_latents.shape[1]
if num_channels_latents + num_channels_mask + num_channels_masked_image != self.unet.config.in_channels:
raise ValueError(
f"Incorrect configuration settings! The config of `pipeline.unet`: {self.unet.config} expects"
f" {self.unet.config.in_channels} but received `num_channels_latents`: {num_channels_latents} +"
f" `num_channels_mask`: {num_channels_mask} + `num_channels_masked_image`: {num_channels_masked_image}"
f" = {num_channels_latents+num_channels_masked_image+num_channels_mask}. Please verify the config of"
" `pipeline.unet` or your `mask_image` or `image` input."
)
elif num_channels_unet != 4:
raise ValueError(
f"The unet {self.unet.__class__} should have either 4 or 9 input channels, not {self.unet.config.in_channels}."
)
# 8.1 Prepare extra step kwargs.
extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
# 9. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
height, width = latents.shape[-2:]
height = height * self.vae_scale_factor
width = width * self.vae_scale_factor
original_size = original_size or (height, width)
target_size = target_size or (height, width)
# 10. Prepare added time ids & embeddings
if negative_original_size is None:
negative_original_size = original_size
if negative_target_size is None:
negative_target_size = target_size
add_text_embeds = pooled_prompt_embeds
if self.text_encoder_2 is None:
text_encoder_projection_dim = int(pooled_prompt_embeds.shape[-1])
else:
text_encoder_projection_dim = self.text_encoder_2.config.projection_dim
add_time_ids, add_neg_time_ids = self._get_add_time_ids(
original_size,
crops_coords_top_left,
target_size,
aesthetic_score,
negative_aesthetic_score,
negative_original_size,
negative_crops_coords_top_left,
negative_target_size,
dtype=prompt_embeds.dtype,
text_encoder_projection_dim=text_encoder_projection_dim,
)
add_time_ids = add_time_ids.repeat(batch_size * num_images_per_prompt, 1)
if self.do_classifier_free_guidance:
prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0)
add_neg_time_ids = add_neg_time_ids.repeat(batch_size * num_images_per_prompt, 1)
add_time_ids = torch.cat([add_neg_time_ids, add_time_ids], dim=0)
prompt_embeds = prompt_embeds.to(device)
add_text_embeds = add_text_embeds.to(device)
add_time_ids = add_time_ids.to(device)
if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
image_embeds = self.prepare_ip_adapter_image_embeds(
ip_adapter_image,
ip_adapter_image_embeds,
device,
batch_size * num_images_per_prompt,
self.do_classifier_free_guidance,
)
# 11. Denoising loop
num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
if (
self.denoising_end is not None
and self.denoising_start is not None
and denoising_value_valid(self.denoising_end)
and denoising_value_valid(self.denoising_start)
and self.denoising_start >= self.denoising_end
):
raise ValueError(
f"`denoising_start`: {self.denoising_start} cannot be larger than or equal to `denoising_end`: "
+ f" {self.denoising_end} when using type float."
)
elif self.denoising_end is not None and denoising_value_valid(self.denoising_end):
discrete_timestep_cutoff = int(
round(
self.scheduler.config.num_train_timesteps
- (self.denoising_end * self.scheduler.config.num_train_timesteps)
)
)
num_inference_steps = len(list(filter(lambda ts: ts >= discrete_timestep_cutoff, timesteps)))
timesteps = timesteps[:num_inference_steps]
# 11.1 Optionally get Guidance Scale Embedding
timestep_cond = None
if self.unet.config.time_cond_proj_dim is not None:
guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt)
timestep_cond = self.get_guidance_scale_embedding(
guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim
).to(device=device, dtype=latents.dtype)
self._num_timesteps = len(timesteps)
with self.progress_bar(total=num_inference_steps) as progress_bar:
for i, t in enumerate(timesteps):
if self.interrupt:
continue
# expand the latents if we are doing classifier free guidance
latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
# concat latents, mask, masked_image_latents in the channel dimension
latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
if num_channels_unet == 9:
latent_model_input = torch.cat([latent_model_input, mask, masked_image_latents], dim=1)
# predict the noise residual
added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids}
if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
added_cond_kwargs["image_embeds"] = image_embeds
noise_pred = self.unet(
latent_model_input,
t,
encoder_hidden_states=prompt_embeds,
timestep_cond=timestep_cond,
cross_attention_kwargs=self.cross_attention_kwargs,
added_cond_kwargs=added_cond_kwargs,
return_dict=False,
)[0]
# perform guidance
if self.do_classifier_free_guidance:
noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond)
if self.do_classifier_free_guidance and self.guidance_rescale > 0.0:
# Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf
noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=self.guidance_rescale)
# compute the previous noisy sample x_t -> x_t-1
latents_dtype = latents.dtype
latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
if latents.dtype != latents_dtype:
if torch.backends.mps.is_available():
# some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272
latents = latents.to(latents_dtype)
if num_channels_unet == 4:
init_latents_proper = image_latents
if self.do_classifier_free_guidance:
init_mask, _ = mask.chunk(2)
else:
init_mask = mask
if i < len(timesteps) - 1:
noise_timestep = timesteps[i + 1]
init_latents_proper = self.scheduler.add_noise(
init_latents_proper, noise, torch.tensor([noise_timestep])
)
latents = (1 - init_mask) * init_latents_proper + init_mask * latents
if callback_on_step_end is not None:
callback_kwargs = {}
for k in callback_on_step_end_tensor_inputs:
callback_kwargs[k] = locals()[k]
callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
latents = callback_outputs.pop("latents", latents)
prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
add_text_embeds = callback_outputs.pop("add_text_embeds", add_text_embeds)
negative_pooled_prompt_embeds = callback_outputs.pop(
"negative_pooled_prompt_embeds", negative_pooled_prompt_embeds
)
add_time_ids = callback_outputs.pop("add_time_ids", add_time_ids)
add_neg_time_ids = callback_outputs.pop("add_neg_time_ids", add_neg_time_ids)
mask = callback_outputs.pop("mask", mask)
masked_image_latents = callback_outputs.pop("masked_image_latents", masked_image_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)
if XLA_AVAILABLE:
xm.mark_step()
if not output_type == "latent":
# make sure the VAE is in float32 mode, as it overflows in float16
needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
if needs_upcasting:
self.upcast_vae()
latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
elif latents.dtype != self.vae.dtype:
if torch.backends.mps.is_available():
# some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272
self.vae = self.vae.to(latents.dtype)
# unscale/denormalize the latents
# denormalize with the mean and std if available and not None
has_latents_mean = hasattr(self.vae.config, "latents_mean") and self.vae.config.latents_mean is not None
has_latents_std = hasattr(self.vae.config, "latents_std") and self.vae.config.latents_std is not None
if has_latents_mean and has_latents_std:
latents_mean = (
torch.tensor(self.vae.config.latents_mean).view(1, 4, 1, 1).to(latents.device, latents.dtype)
)
latents_std = (
torch.tensor(self.vae.config.latents_std).view(1, 4, 1, 1).to(latents.device, latents.dtype)
)
latents = latents * latents_std / self.vae.config.scaling_factor + latents_mean
else:
latents = latents / self.vae.config.scaling_factor
image = self.vae.decode(latents, return_dict=False)[0]
# cast back to fp16 if needed
if needs_upcasting:
self.vae.to(dtype=torch.float16)
else:
return StableDiffusionXLPipelineOutput(images=latents)
# apply watermark if available
if self.watermark is not None:
image = self.watermark.apply_watermark(image)
image = self.image_processor.postprocess(image, output_type=output_type)
if padding_mask_crop is not None:
image = [self.image_processor.apply_overlay(mask_image, original_image, i, crops_coords) for i in image]
# Offload all models
self.maybe_free_model_hooks()
if not return_dict:
return (image,)
return StableDiffusionXLPipelineOutput(images=image)
Kolors-FaceID/kolors/pipelines/pipeline_stable_diffusion_xl_chatglm_256_ipadapter.py 0 → 100644
View file @ 0a1fba1c
# 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 sys
import os
sys.path.append(os.path.abspath(os.path.join(os.path.dirname(__file__), '..')))
from kolors.models.modeling_chatglm import ChatGLMModel
from kolors.models.tokenization_chatglm import ChatGLMTokenizer
import inspect
from typing import Any, Callable, Dict, List, Optional, Tuple, Union
import torch
from transformers import (
CLIPImageProcessor,
CLIPTextModel,
CLIPTextModelWithProjection,
CLIPTokenizer,
CLIPVisionModelWithProjection,
)
from transformers import XLMRobertaModel, ChineseCLIPTextModel
from diffusers.image_processor import VaeImageProcessor,PipelineImageInput
from diffusers.loaders import (
FromSingleFileMixin,
IPAdapterMixin,
LoraLoaderMixin,
TextualInversionLoaderMixin
)
from diffusers.models import AutoencoderKL, UNet2DConditionModel,ImageProjection
from diffusers.models.attention_processor import (
AttnProcessor2_0,
LoRAAttnProcessor2_0,
LoRAXFormersAttnProcessor,
XFormersAttnProcessor,
)
from diffusers.schedulers import KarrasDiffusionSchedulers
from diffusers.utils import (
is_accelerate_available,
is_accelerate_version,
logging,
replace_example_docstring,
)
try:
from diffusers.utils import randn_tensor
except:
from diffusers.utils.torch_utils import randn_tensor
from diffusers.pipelines.pipeline_utils import DiffusionPipeline
from diffusers.pipelines.stable_diffusion_xl import StableDiffusionXLPipelineOutput
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
EXAMPLE_DOC_STRING = """
Examples:
```py
>>> import torch
>>> from diffusers import StableDiffusionXLPipeline
>>> pipe = StableDiffusionXLPipeline.from_pretrained(
... "stabilityai/stable-diffusion-xl-base-0.9", torch_dtype=torch.float16
... )
>>> pipe = pipe.to("cuda")
>>> prompt = "a photo of an astronaut riding a horse on mars"
>>> image = pipe(prompt).images[0]
```
"""
# 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://arxiv.org/pdf/2305.08891.pdf). 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 StableDiffusionXLPipeline(DiffusionPipeline, FromSingleFileMixin, LoraLoaderMixin, IPAdapterMixin,):
r"""
Pipeline for text-to-image generation using Stable Diffusion XL.
This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
In addition the pipeline inherits the following loading methods:
- *Textual-Inversion*: [`loaders.TextualInversionLoaderMixin.load_textual_inversion`]
- *LoRA*: [`loaders.LoraLoaderMixin.load_lora_weights`]
- *Ckpt*: [`loaders.FromSingleFileMixin.from_single_file`]
as well as the following saving methods:
- *LoRA*: [`loaders.LoraLoaderMixin.save_lora_weights`]
Args:
vae ([`AutoencoderKL`]):
Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
text_encoder ([`CLIPTextModel`]):
Frozen text-encoder. Stable Diffusion XL uses the text portion of
[CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically
the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant.
tokenizer (`CLIPTokenizer`):
Tokenizer of class
[CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents.
scheduler ([`SchedulerMixin`]):
A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
[`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
"""
def __init__(
self,
vae: AutoencoderKL,
text_encoder: ChatGLMModel,
tokenizer: ChatGLMTokenizer,
unet: UNet2DConditionModel,
scheduler: KarrasDiffusionSchedulers,
image_encoder: CLIPVisionModelWithProjection = None,
feature_extractor: CLIPImageProcessor = None,
force_zeros_for_empty_prompt: bool = True,
):
super().__init__()
self.register_modules(
vae=vae,
text_encoder=text_encoder,
tokenizer=tokenizer,
unet=unet,
scheduler=scheduler,
image_encoder=image_encoder,
feature_extractor=feature_extractor,
)
self.register_to_config(force_zeros_for_empty_prompt=force_zeros_for_empty_prompt)
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.default_sample_size = self.unet.config.sample_size
# self.watermark = StableDiffusionXLWatermarker()
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_slicing
def enable_vae_slicing(self):
r"""
Enable sliced VAE decoding.
When this option is enabled, the VAE will split the input tensor in slices to compute decoding in several
steps. This is useful to save some memory and allow larger batch sizes.
"""
self.vae.enable_slicing()
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_slicing
def disable_vae_slicing(self):
r"""
Disable sliced VAE decoding. If `enable_vae_slicing` was previously invoked, this method will go back to
computing decoding in one step.
"""
self.vae.disable_slicing()
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_tiling
def enable_vae_tiling(self):
r"""
Enable tiled VAE decoding.
When this option is enabled, the VAE will split the input tensor into tiles to compute decoding and encoding in
several steps. This is useful to save a large amount of memory and to allow the processing of larger images.
"""
self.vae.enable_tiling()
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_tiling
def disable_vae_tiling(self):
r"""
Disable tiled VAE decoding. If `enable_vae_tiling` was previously invoked, this method will go back to
computing decoding in one step.
"""
self.vae.disable_tiling()
def enable_sequential_cpu_offload(self, gpu_id=0):
r"""
Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet,
text_encoder, vae and safety checker have their state dicts saved to CPU and then are moved to a
`torch.device('meta') and loaded to GPU only when their specific submodule has its `forward` method called.
Note that offloading happens on a submodule basis. Memory savings are higher than with
`enable_model_cpu_offload`, but performance is lower.
"""
if is_accelerate_available() and is_accelerate_version(">=", "0.14.0"):
from accelerate import cpu_offload
else:
raise ImportError("`enable_sequential_cpu_offload` requires `accelerate v0.14.0` 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)
for cpu_offloaded_model in [self.unet, self.text_encoder, self.vae]:
cpu_offload(cpu_offloaded_model, device)
def enable_model_cpu_offload(self, gpu_id=0):
r"""
Offloads all models to CPU using accelerate, reducing memory usage with a low impact on performance. Compared
to `enable_sequential_cpu_offload`, this method 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 with
`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)
model_sequence = (
[self.text_encoder, self.image_encoder]
)
model_sequence.extend([self.unet, self.vae])
hook = None
for cpu_offloaded_model in model_sequence:
_, hook = cpu_offload_with_hook(cpu_offloaded_model, device, prev_module_hook=hook)
# We'll offload the last model manually.
self.final_offload_hook = hook
@property
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._execution_device
def _execution_device(self):
r"""
Returns the device on which the pipeline's models will be executed. After calling
`pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module
hooks.
"""
if not hasattr(self.unet, "_hf_hook"):
return self.device
for module in self.unet.modules():
if (
hasattr(module, "_hf_hook")
and hasattr(module._hf_hook, "execution_device")
and module._hf_hook.execution_device is not None
):
return torch.device(module._hf_hook.execution_device)
return self.device
def encode_prompt(
self,
prompt,
device: Optional[torch.device] = None,
num_images_per_prompt: int = 1,
do_classifier_free_guidance: bool = True,
negative_prompt=None,
prompt_embeds: Optional[torch.FloatTensor] = None,
negative_prompt_embeds: Optional[torch.FloatTensor] = None,
pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
negative_pooled_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.
pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
If not provided, pooled text embeddings will be generated from `prompt` input argument.
negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
weighting. If not provided, pooled 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.
"""
# from IPython import embed; embed(); exit()
device = device or self._execution_device
# 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]
# Define tokenizers and text encoders
tokenizers = [self.tokenizer]
text_encoders = [self.text_encoder]
if prompt_embeds is None:
# textual inversion: procecss multi-vector tokens if necessary
prompt_embeds_list = []
for tokenizer, text_encoder in zip(tokenizers, text_encoders):
if isinstance(self, TextualInversionLoaderMixin):
prompt = self.maybe_convert_prompt(prompt, tokenizer)
text_inputs = tokenizer(
prompt,
padding="max_length",
max_length=256,
truncation=True,
return_tensors="pt",
).to('cuda')
output = text_encoder(
input_ids=text_inputs['input_ids'] ,
attention_mask=text_inputs['attention_mask'],
position_ids=text_inputs['position_ids'],
output_hidden_states=True)
prompt_embeds = output.hidden_states[-2].permute(1, 0, 2).clone()
pooled_prompt_embeds = output.hidden_states[-1][-1, :, :].clone() # [batch_size, 4096]
bs_embed, seq_len, _ = prompt_embeds.shape
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)
prompt_embeds_list.append(prompt_embeds)
# prompt_embeds = torch.concat(prompt_embeds_list, dim=-1)
prompt_embeds = prompt_embeds_list[0]
# get unconditional embeddings for classifier free guidance
zero_out_negative_prompt = negative_prompt is None and self.config.force_zeros_for_empty_prompt
if do_classifier_free_guidance and negative_prompt_embeds is None and zero_out_negative_prompt:
negative_prompt_embeds = torch.zeros_like(prompt_embeds)
negative_pooled_prompt_embeds = torch.zeros_like(pooled_prompt_embeds)
elif do_classifier_free_guidance and negative_prompt_embeds is None:
# negative_prompt = negative_prompt or ""
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
negative_prompt_embeds_list = []
for tokenizer, text_encoder in zip(tokenizers, text_encoders):
# textual inversion: procecss multi-vector tokens if necessary
if isinstance(self, TextualInversionLoaderMixin):
uncond_tokens = self.maybe_convert_prompt(uncond_tokens, tokenizer)
max_length = prompt_embeds.shape[1]
uncond_input = tokenizer(
uncond_tokens,
padding="max_length",
max_length=max_length,
truncation=True,
return_tensors="pt",
).to('cuda')
output = text_encoder(
input_ids=uncond_input['input_ids'] ,
attention_mask=uncond_input['attention_mask'],
position_ids=uncond_input['position_ids'],
output_hidden_states=True)
negative_prompt_embeds = output.hidden_states[-2].permute(1, 0, 2).clone()
negative_pooled_prompt_embeds = output.hidden_states[-1][-1, :, :].clone() # [batch_size, 4096]
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=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
negative_prompt_embeds_list.append(negative_prompt_embeds)
# negative_prompt_embeds = torch.concat(negative_prompt_embeds_list, dim=-1)
negative_prompt_embeds = negative_prompt_embeds_list[0]
bs_embed = pooled_prompt_embeds.shape[0]
pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt).view(
bs_embed * num_images_per_prompt, -1
)
if do_classifier_free_guidance:
negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.repeat(1, num_images_per_prompt).view(
bs_embed * num_images_per_prompt, -1
)
return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image
def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None):
dtype = next(self.image_encoder.parameters()).dtype
if not isinstance(image, torch.Tensor):
image = self.feature_extractor(image, return_tensors="pt").pixel_values
image = image.to(device=device, dtype=dtype)
if output_hidden_states:
image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2]
image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0)
uncond_image_enc_hidden_states = self.image_encoder(
torch.zeros_like(image), output_hidden_states=True
).hidden_states[-2]
uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave(
num_images_per_prompt, dim=0
)
return image_enc_hidden_states, uncond_image_enc_hidden_states
else:
image_embeds = self.image_encoder(image).image_embeds
image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
uncond_image_embeds = torch.zeros_like(image_embeds)
return image_embeds, uncond_image_embeds
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_ip_adapter_image_embeds
def prepare_ip_adapter_image_embeds(
self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance
):
image_embeds = []
if do_classifier_free_guidance:
negative_image_embeds = []
if ip_adapter_image_embeds is None:
if not isinstance(ip_adapter_image, list):
ip_adapter_image = [ip_adapter_image]
if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers):
raise ValueError(
f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters."
)
for single_ip_adapter_image, image_proj_layer in zip(
ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers
):
output_hidden_state = not isinstance(image_proj_layer, ImageProjection)
single_image_embeds, single_negative_image_embeds = self.encode_image(
single_ip_adapter_image, device, 1, output_hidden_state
)
image_embeds.append(single_image_embeds[None, :])
if do_classifier_free_guidance:
negative_image_embeds.append(single_negative_image_embeds[None, :])
else:
for single_image_embeds in ip_adapter_image_embeds:
if do_classifier_free_guidance:
single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2)
negative_image_embeds.append(single_negative_image_embeds)
image_embeds.append(single_image_embeds)
ip_adapter_image_embeds = []
for i, single_image_embeds in enumerate(image_embeds):
single_image_embeds = torch.cat([single_image_embeds] * num_images_per_prompt, dim=0)
if do_classifier_free_guidance:
single_negative_image_embeds = torch.cat([negative_image_embeds[i]] * num_images_per_prompt, dim=0)
single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds], dim=0)
single_image_embeds = single_image_embeds.to(device=device)
ip_adapter_image_embeds.append(single_image_embeds)
return ip_adapter_image_embeds
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
def prepare_extra_step_kwargs(self, generator, eta):
# prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
# eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
# eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
# and should be between [0, 1]
accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
extra_step_kwargs = {}
if accepts_eta:
extra_step_kwargs["eta"] = eta
# check if the scheduler accepts generator
accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
if accepts_generator:
extra_step_kwargs["generator"] = generator
return extra_step_kwargs
def check_inputs(
self,
prompt,
height,
width,
callback_steps,
negative_prompt=None,
prompt_embeds=None,
negative_prompt_embeds=None,
pooled_prompt_embeds=None,
negative_pooled_prompt_embeds=None,
):
if height % 8 != 0 or width % 8 != 0:
raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
if (callback_steps is None) or (
callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
):
raise ValueError(
f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
f" {type(callback_steps)}."
)
if prompt is not None and prompt_embeds is not None:
raise ValueError(
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 prompt_embeds is not None and pooled_prompt_embeds is None:
raise ValueError(
"If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`."
)
if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None:
raise ValueError(
"If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`."
)
# 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
def _get_add_time_ids(self, original_size, crops_coords_top_left, target_size, dtype):
add_time_ids = list(original_size + crops_coords_top_left + target_size)
passed_add_embed_dim = (
self.unet.config.addition_time_embed_dim * len(add_time_ids) + 4096
)
expected_add_embed_dim = self.unet.add_embedding.linear_1.in_features
if expected_add_embed_dim != passed_add_embed_dim:
raise ValueError(
f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. The model has an incorrect config. Please check `unet.config.time_embedding_type` and `text_encoder_2.config.projection_dim`."
)
add_time_ids = torch.tensor([add_time_ids], dtype=dtype)
return add_time_ids
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_upscale.StableDiffusionUpscalePipeline.upcast_vae
def upcast_vae(self):
dtype = self.vae.dtype
self.vae.to(dtype=torch.float32)
use_torch_2_0_or_xformers = isinstance(
self.vae.decoder.mid_block.attentions[0].processor,
(
AttnProcessor2_0,
XFormersAttnProcessor,
LoRAXFormersAttnProcessor,
LoRAAttnProcessor2_0,
),
)
# if xformers or torch_2_0 is used attention block does not need
# to be in float32 which can save lots of memory
if use_torch_2_0_or_xformers:
self.vae.post_quant_conv.to(dtype)
self.vae.decoder.conv_in.to(dtype)
self.vae.decoder.mid_block.to(dtype)
@torch.no_grad()
@replace_example_docstring(EXAMPLE_DOC_STRING)
def __call__(
self,
prompt: Union[str, List[str]] = None,
height: Optional[int] = None,
width: Optional[int] = None,
num_inference_steps: int = 50,
denoising_end: Optional[float] = None,
guidance_scale: float = 5.0,
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.FloatTensor] = None,
prompt_embeds: Optional[torch.FloatTensor] = None,
negative_prompt_embeds: Optional[torch.FloatTensor] = None,
pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
ip_adapter_image: Optional[PipelineImageInput] = None,
ip_adapter_image_embeds: Optional[List[torch.Tensor]] = 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,
guidance_rescale: float = 0.0,
original_size: Optional[Tuple[int, int]] = None,
crops_coords_top_left: Tuple[int, int] = (0, 0),
target_size: Optional[Tuple[int, int]] = None,
use_dynamic_threshold: Optional[bool] = False,
):
r"""
Function invoked when calling the pipeline for generation.
Args:
prompt (`str` or `List[str]`, *optional*):
The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
instead.
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.
denoising_end (`float`, *optional*):
When specified, determines the fraction (between 0.0 and 1.0) of the total denoising process to be
completed before it is intentionally prematurely terminated. For instance, if denoising_end is set to
0.7 and `num_inference_steps` is fixed at 50, the process will execute only 35 (i.e., 0.7 * 50)
Denoisers" multi-pipeline setup, as elaborated in [**Refining the Image
Output**](https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion/stable_diffusion_xl#refining-the-image-output)
guidance_scale (`float`, *optional*, defaults to 7.5):
`guidance_scale` is defined as `w` of equation 2. of [Imagen
Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
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`).
num_images_per_prompt (`int`, *optional*, defaults to 1):
The number of images to generate per prompt.
eta (`float`, *optional*, defaults to 0.0):
Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
[`schedulers.DDIMScheduler`], will be ignored for others.
generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
to make generation deterministic.
latents (`torch.FloatTensor`, *optional*):
Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
tensor will ge generated by sampling using the supplied random `generator`.
prompt_embeds (`torch.FloatTensor`, *optional*):
Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
provided, text embeddings will be generated from `prompt` input argument.
negative_prompt_embeds (`torch.FloatTensor`, *optional*):
Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
argument.
pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
If not provided, pooled text embeddings will be generated from `prompt` input argument.
negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
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.StableDiffusionXLPipelineOutput`] instead of a
callback (`Callable`, *optional*):
A function that will be called every `callback_steps` steps during inference. The function will be
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 under
`self.processor` in
[diffusers.cross_attention](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/cross_attention.py).
guidance_rescale (`float`, *optional*, defaults to 0.7):
Guidance rescale factor proposed by [Common Diffusion Noise Schedules and Sample Steps are
Flawed](https://arxiv.org/pdf/2305.08891.pdf) `guidance_scale` is defined as `φ` in equation 16. of
[Common Diffusion Noise Schedules and Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf).
Guidance rescale factor should fix overexposure when using zero terminal SNR.
original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
TODO
crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)):
TODO
target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
TODO
Examples:
Returns:
[`~pipelines.stable_diffusion.StableDiffusionXLPipelineOutput`] or `tuple`:
[`~pipelines.stable_diffusion.StableDiffusionXLPipelineOutput`] 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`.
"""
# 0. Default height and width to unet
height = height or self.default_sample_size * self.vae_scale_factor
width = width or self.default_sample_size * self.vae_scale_factor
original_size = original_size or (height, width)
target_size = target_size or (height, width)
# 1. Check inputs. Raise error if not correct
self.check_inputs(
prompt,
height,
width,
callback_steps,
negative_prompt,
prompt_embeds,
negative_prompt_embeds,
pooled_prompt_embeds,
negative_pooled_prompt_embeds,
)
# 2. Define call parameters
if prompt is not None and isinstance(prompt, str):
batch_size = 1
elif prompt is not None and isinstance(prompt, list):
batch_size = len(prompt)
else:
batch_size = prompt_embeds.shape[0]
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,
negative_prompt_embeds,
pooled_prompt_embeds,
negative_pooled_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,
pooled_prompt_embeds=pooled_prompt_embeds,
negative_pooled_prompt_embeds=negative_pooled_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. Prepare added time ids & embeddings
add_text_embeds = pooled_prompt_embeds
add_time_ids = self._get_add_time_ids(
original_size, crops_coords_top_left, target_size, dtype=prompt_embeds.dtype
)
if do_classifier_free_guidance:
prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0)
add_time_ids = torch.cat([add_time_ids, add_time_ids], dim=0)
prompt_embeds = prompt_embeds.to(device)
add_text_embeds = add_text_embeds.to(device)
add_time_ids = add_time_ids.to(device).repeat(batch_size * num_images_per_prompt, 1)
if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
image_embeds = self.prepare_ip_adapter_image_embeds(
ip_adapter_image,
ip_adapter_image_embeds,
device,
batch_size * num_images_per_prompt,
do_classifier_free_guidance,
)
# 8. Denoising loop
num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
# 7.1 Apply denoising_end
if denoising_end is not None:
num_inference_steps = int(round(denoising_end * num_inference_steps))
timesteps = timesteps[: num_warmup_steps + self.scheduler.order * num_inference_steps]
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
added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids}
if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
added_cond_kwargs["image_embeds"] = image_embeds
# import pdb; pdb.set_trace()
noise_pred = self.unet(
latent_model_input,
t,
encoder_hidden_states=prompt_embeds,
cross_attention_kwargs=cross_attention_kwargs,
added_cond_kwargs=added_cond_kwargs,
return_dict=False,
)[0]
# 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 use_dynamic_threshold:
DynamicThresh = DynThresh(maxSteps=num_inference_steps, experiment_mode=0)
noise_pred = DynamicThresh.dynthresh(noise_pred_text,
noise_pred_uncond,
guidance_scale,
None)
if do_classifier_free_guidance and guidance_rescale > 0.0:
# Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf
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, return_dict=False)[0]
# 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)
# make sureo the VAE is in float32 mode, as it overflows in float16
# torch.cuda.empty_cache()
if self.vae.dtype == torch.float16 and self.vae.config.force_upcast:
self.upcast_vae()
latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
if not output_type == "latent":
latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
else:
image = latents
return StableDiffusionXLPipelineOutput(images=image)
# image = self.watermark.apply_watermark(image)
image = self.image_processor.postprocess(image, output_type=output_type)
# 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,)
return StableDiffusionXLPipelineOutput(images=image)
if __name__ == "__main__":
pass
Kolors-FaceID/kolors/pipelines/pipeline_stable_diffusion_xl_chatglm_256_ipadapter_FaceID.py 0 → 100644
View file @ 0a1fba1c
# 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 sys
import os
sys.path.append(os.path.abspath(os.path.join(os.path.dirname(__file__), '..')))
from kolors.models.modeling_chatglm import ChatGLMModel
from kolors.models.tokenization_chatglm import ChatGLMTokenizer
import inspect
from typing import Any, Callable, Dict, List, Optional, Tuple, Union
import torch
from transformers import CLIPTextModel, CLIPTextModelWithProjection, CLIPTokenizer
from transformers import XLMRobertaModel, ChineseCLIPTextModel
from transformers import CLIPImageProcessor, CLIPVisionModelWithProjection
from diffusers.image_processor import VaeImageProcessor, PipelineImageInput
from diffusers.loaders import FromSingleFileMixin, LoraLoaderMixin, TextualInversionLoaderMixin
from diffusers.models import AutoencoderKL, UNet2DConditionModel
from diffusers.models.attention_processor import (
AttnProcessor2_0,
LoRAAttnProcessor2_0,
LoRAXFormersAttnProcessor,
XFormersAttnProcessor,
)
from diffusers.schedulers import KarrasDiffusionSchedulers
from diffusers.utils import (
is_accelerate_available,
is_accelerate_version,
logging,
replace_example_docstring,
)
try:
from diffusers.utils import randn_tensor
except:
from diffusers.utils.torch_utils import randn_tensor
from diffusers.pipelines.pipeline_utils import DiffusionPipeline
from diffusers.pipelines.stable_diffusion_xl import StableDiffusionXLPipelineOutput
from kolors.models.ipa_faceid_plus.ipa_faceid_plus import ProjPlusModel
from kolors.models.ipa_faceid_plus.attention_processor import IPAttnProcessor2_0 as IPAttnProcessor, AttnProcessor2_0 as AttnProcessor
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
EXAMPLE_DOC_STRING = """
Examples:
```py
>>> import torch
>>> from diffusers import StableDiffusionXLPipeline
>>> pipe = StableDiffusionXLPipeline.from_pretrained(
... "stabilityai/stable-diffusion-xl-base-0.9", torch_dtype=torch.float16
... )
>>> pipe = pipe.to("cuda")
>>> prompt = "a photo of an astronaut riding a horse on mars"
>>> image = pipe(prompt).images[0]
```
"""
# 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://arxiv.org/pdf/2305.08891.pdf). 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 StableDiffusionXLPipeline(DiffusionPipeline, FromSingleFileMixin, LoraLoaderMixin):
r"""
Pipeline for text-to-image generation using Stable Diffusion XL.
This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
In addition the pipeline inherits the following loading methods:
- *Textual-Inversion*: [`loaders.TextualInversionLoaderMixin.load_textual_inversion`]
- *LoRA*: [`loaders.LoraLoaderMixin.load_lora_weights`]
- *Ckpt*: [`loaders.FromSingleFileMixin.from_single_file`]
as well as the following saving methods:
- *LoRA*: [`loaders.LoraLoaderMixin.save_lora_weights`]
Args:
vae ([`AutoencoderKL`]):
Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
text_encoder ([`CLIPTextModel`]):
Frozen text-encoder. Stable Diffusion XL uses the text portion of
[CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically
the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant.
tokenizer (`CLIPTokenizer`):
Tokenizer of class
[CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents.
scheduler ([`SchedulerMixin`]):
A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
[`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
"""
def __init__(
self,
vae: AutoencoderKL,
text_encoder: ChatGLMModel,
tokenizer: ChatGLMTokenizer,
unet: UNet2DConditionModel,
scheduler: KarrasDiffusionSchedulers,
force_zeros_for_empty_prompt: bool = True,
face_clip_encoder: CLIPVisionModelWithProjection = None,
face_clip_processor: CLIPImageProcessor = None,
):
super().__init__()
#### image project with Q-former for FaceID-Plus
if face_clip_encoder is not None:
self.image_proj_model = self.init_ip_adapter_proj_layer(
clip_embeddings_dim = face_clip_encoder.config.hidden_size,
num_tokens = 6
)
else:
raise NotImplemented("face clip encoder is not provided...")
self.image_proj_model = None
self.register_modules(
vae=vae,
text_encoder=text_encoder,
tokenizer=tokenizer,
unet=unet,
scheduler=scheduler,
face_clip_encoder = face_clip_encoder,
face_clip_processor = face_clip_processor,
# image_proj_model = image_proj_model,
)
self.register_to_config(force_zeros_for_empty_prompt=force_zeros_for_empty_prompt)
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.default_sample_size = self.unet.config.sample_size
# self.watermark = StableDiffusionXLWatermarker()
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_slicing
def enable_vae_slicing(self):
r"""
Enable sliced VAE decoding.
When this option is enabled, the VAE will split the input tensor in slices to compute decoding in several
steps. This is useful to save some memory and allow larger batch sizes.
"""
self.vae.enable_slicing()
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_slicing
def disable_vae_slicing(self):
r"""
Disable sliced VAE decoding. If `enable_vae_slicing` was previously invoked, this method will go back to
computing decoding in one step.
"""
self.vae.disable_slicing()
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_tiling
def enable_vae_tiling(self):
r"""
Enable tiled VAE decoding.
When this option is enabled, the VAE will split the input tensor into tiles to compute decoding and encoding in
several steps. This is useful to save a large amount of memory and to allow the processing of larger images.
"""
self.vae.enable_tiling()
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_tiling
def disable_vae_tiling(self):
r"""
Disable tiled VAE decoding. If `enable_vae_tiling` was previously invoked, this method will go back to
computing decoding in one step.
"""
self.vae.disable_tiling()
def enable_sequential_cpu_offload(self, gpu_id=0):
r"""
Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet,
text_encoder, vae and safety checker have their state dicts saved to CPU and then are moved to a
`torch.device('meta') and loaded to GPU only when their specific submodule has its `forward` method called.
Note that offloading happens on a submodule basis. Memory savings are higher than with
`enable_model_cpu_offload`, but performance is lower.
"""
if is_accelerate_available() and is_accelerate_version(">=", "0.14.0"):
from accelerate import cpu_offload
else:
raise ImportError("`enable_sequential_cpu_offload` requires `accelerate v0.14.0` 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)
for cpu_offloaded_model in [self.unet, self.text_encoder, self.vae]:
cpu_offload(cpu_offloaded_model, device)
def enable_model_cpu_offload(self, gpu_id=0):
r"""
Offloads all models to CPU using accelerate, reducing memory usage with a low impact on performance. Compared
to `enable_sequential_cpu_offload`, this method 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 with
`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)
model_sequence = (
[self.text_encoder, self.face_clip_encoder]
)
model_sequence.extend([self.unet, self.vae])
hook = None
for cpu_offloaded_model in model_sequence:
_, hook = cpu_offload_with_hook(cpu_offloaded_model, device, prev_module_hook=hook)
# We'll offload the last model manually.
self.final_offload_hook = hook
@property
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._execution_device
def _execution_device(self):
r"""
Returns the device on which the pipeline's models will be executed. After calling
`pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module
hooks.
"""
if not hasattr(self.unet, "_hf_hook"):
return self.device
for module in self.unet.modules():
if (
hasattr(module, "_hf_hook")
and hasattr(module._hf_hook, "execution_device")
and module._hf_hook.execution_device is not None
):
return torch.device(module._hf_hook.execution_device)
return self.device
def encode_prompt(
self,
prompt,
device: Optional[torch.device] = None,
num_images_per_prompt: int = 1,
do_classifier_free_guidance: bool = True,
negative_prompt=None,
prompt_embeds: Optional[torch.FloatTensor] = None,
negative_prompt_embeds: Optional[torch.FloatTensor] = None,
pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
negative_pooled_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.
pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
If not provided, pooled text embeddings will be generated from `prompt` input argument.
negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
weighting. If not provided, pooled 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.
"""
# from IPython import embed; embed(); exit()
device = device or self._execution_device
# 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]
# Define tokenizers and text encoders
tokenizers = [self.tokenizer]
text_encoders = [self.text_encoder]
if prompt_embeds is None:
# textual inversion: procecss multi-vector tokens if necessary
prompt_embeds_list = []
for tokenizer, text_encoder in zip(tokenizers, text_encoders):
if isinstance(self, TextualInversionLoaderMixin):
prompt = self.maybe_convert_prompt(prompt, tokenizer)
text_inputs = tokenizer(
prompt,
padding="max_length",
max_length=256,
truncation=True,
return_tensors="pt",
).to('cuda')
output = text_encoder(
input_ids=text_inputs['input_ids'] ,
attention_mask=text_inputs['attention_mask'],
position_ids=text_inputs['position_ids'],
output_hidden_states=True)
prompt_embeds = output.hidden_states[-2].permute(1, 0, 2).clone()
pooled_prompt_embeds = output.hidden_states[-1][-1, :, :].clone() # [batch_size, 4096]
bs_embed, seq_len, _ = prompt_embeds.shape
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)
prompt_embeds_list.append(prompt_embeds)
# prompt_embeds = torch.concat(prompt_embeds_list, dim=-1)
prompt_embeds = prompt_embeds_list[0]
# get unconditional embeddings for classifier free guidance
zero_out_negative_prompt = negative_prompt is None and self.config.force_zeros_for_empty_prompt
if do_classifier_free_guidance and negative_prompt_embeds is None and zero_out_negative_prompt:
negative_prompt_embeds = torch.zeros_like(prompt_embeds)
negative_pooled_prompt_embeds = torch.zeros_like(pooled_prompt_embeds)
elif do_classifier_free_guidance and negative_prompt_embeds is None:
# negative_prompt = negative_prompt or ""
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
negative_prompt_embeds_list = []
for tokenizer, text_encoder in zip(tokenizers, text_encoders):
# textual inversion: procecss multi-vector tokens if necessary
if isinstance(self, TextualInversionLoaderMixin):
uncond_tokens = self.maybe_convert_prompt(uncond_tokens, tokenizer)
max_length = prompt_embeds.shape[1]
uncond_input = tokenizer(
uncond_tokens,
padding="max_length",
max_length=max_length,
truncation=True,
return_tensors="pt",
).to('cuda')
output = text_encoder(
input_ids=uncond_input['input_ids'] ,
attention_mask=uncond_input['attention_mask'],
position_ids=uncond_input['position_ids'],
output_hidden_states=True)
negative_prompt_embeds = output.hidden_states[-2].permute(1, 0, 2).clone()
negative_pooled_prompt_embeds = output.hidden_states[-1][-1, :, :].clone() # [batch_size, 4096]
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=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
negative_prompt_embeds_list.append(negative_prompt_embeds)
# negative_prompt_embeds = torch.concat(negative_prompt_embeds_list, dim=-1)
negative_prompt_embeds = negative_prompt_embeds_list[0]
bs_embed = pooled_prompt_embeds.shape[0]
pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt).view(
bs_embed * num_images_per_prompt, -1
)
if do_classifier_free_guidance:
negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.repeat(1, num_images_per_prompt).view(
bs_embed * num_images_per_prompt, -1
)
return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
def prepare_extra_step_kwargs(self, generator, eta):
# prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
# eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
# eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
# and should be between [0, 1]
accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
extra_step_kwargs = {}
if accepts_eta:
extra_step_kwargs["eta"] = eta
# check if the scheduler accepts generator
accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
if accepts_generator:
extra_step_kwargs["generator"] = generator
return extra_step_kwargs
def check_inputs(
self,
prompt,
height,
width,
callback_steps,
negative_prompt=None,
prompt_embeds=None,
negative_prompt_embeds=None,
pooled_prompt_embeds=None,
negative_pooled_prompt_embeds=None,
):
if height % 8 != 0 or width % 8 != 0:
raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
if (callback_steps is None) or (
callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
):
raise ValueError(
f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
f" {type(callback_steps)}."
)
if prompt is not None and prompt_embeds is not None:
raise ValueError(
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 prompt_embeds is not None and pooled_prompt_embeds is None:
raise ValueError(
"If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`."
)
if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None:
raise ValueError(
"If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`."
)
# 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
def _get_add_time_ids(self, original_size, crops_coords_top_left, target_size, dtype):
add_time_ids = list(original_size + crops_coords_top_left + target_size)
passed_add_embed_dim = (
self.unet.config.addition_time_embed_dim * len(add_time_ids) + 4096
)
expected_add_embed_dim = self.unet.add_embedding.linear_1.in_features
if expected_add_embed_dim != passed_add_embed_dim:
raise ValueError(
f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. The model has an incorrect config. Please check `unet.config.time_embedding_type` and `text_encoder_2.config.projection_dim`."
)
add_time_ids = torch.tensor([add_time_ids], dtype=dtype)
return add_time_ids
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_upscale.StableDiffusionUpscalePipeline.upcast_vae
def upcast_vae(self):
dtype = self.vae.dtype
self.vae.to(dtype=torch.float32)
use_torch_2_0_or_xformers = isinstance(
self.vae.decoder.mid_block.attentions[0].processor,
(
AttnProcessor2_0,
XFormersAttnProcessor,
LoRAXFormersAttnProcessor,
LoRAAttnProcessor2_0,
),
)
# if xformers or torch_2_0 is used attention block does not need
# to be in float32 which can save lots of memory
if use_torch_2_0_or_xformers:
self.vae.post_quant_conv.to(dtype)
self.vae.decoder.conv_in.to(dtype)
self.vae.decoder.mid_block.to(dtype)
#### set ip adapter module
def set_ip_adapter(self, device, num_tokens = 6):
unet = self.unet
attn_procs = {}
for name in unet.attn_processors.keys():
cross_attention_dim = None if name.endswith("attn1.processor") else unet.config.cross_attention_dim
if name.startswith("mid_block"):
hidden_size = unet.config.block_out_channels[-1]
elif name.startswith("up_blocks"):
block_id = int(name[len("up_blocks.")])
hidden_size = list(reversed(unet.config.block_out_channels))[block_id]
elif name.startswith("down_blocks"):
block_id = int(name[len("down_blocks.")])
hidden_size = unet.config.block_out_channels[block_id]
if cross_attention_dim is None:
attn_procs[name] = AttnProcessor()
else:
attn_procs[name] = IPAttnProcessor(
hidden_size = hidden_size,
cross_attention_dim = cross_attention_dim,
scale = 1.0,
num_tokens = num_tokens
).to(device, dtype = unet.dtype)
unet.set_attn_processor(attn_procs)
def init_ip_adapter_proj_layer(self, clip_embeddings_dim, num_tokens):
image_proj_model = ProjPlusModel(
cross_attention_dim = 4096,
id_embeddings_dim = 512,
clip_embeddings_dim = clip_embeddings_dim,
num_tokens = num_tokens
)
return image_proj_model
#### load ip adapter model weight
def load_ip_adapter_faceid_plus(self, ip_faceid_model_path, device):
params = torch.load(ip_faceid_model_path, 'cpu')
self.image_proj_model.load_state_dict(params["image_proj"])
self.image_proj_model.to(device, dtype = torch.float16)
self.set_ip_adapter(num_tokens = 6, device = device)
ip_layers = torch.nn.ModuleList(self.unet.attn_processors.values())
ip_layers.load_state_dict(params["adapter_modules"])
#### get image embeddings ####
def get_clip_feat(self, face_crop_image, device):
face_clip_images = self.face_clip_processor(images = face_crop_image, return_tensors = "pt").pixel_values
face_clip_images = face_clip_images.to(device, dtype = torch.float16)
with torch.no_grad():
face_clip_embeddings = self.face_clip_encoder(
face_clip_images,
output_hidden_states = True
).hidden_states[-2]
return face_clip_embeddings
def get_fused_face_embedds(self, face_insightface_embeds, face_crop_image, num_images_per_prompt, device):
with torch.inference_mode():
face_clip_embeds = self.get_clip_feat(face_crop_image, device)
face_clip_embeds = face_clip_embeds.to(device = device, dtype = torch.float16)
image_prompt_embeds = self.image_proj_model(face_insightface_embeds, face_clip_embeds)
uncond_image_prompt_embeds = self.image_proj_model(torch.zeros_like(face_insightface_embeds), torch.zeros_like(face_clip_embeds))
bs_embed, seq_len, _ = image_prompt_embeds.shape
image_prompt_embeds = image_prompt_embeds.repeat(1, num_images_per_prompt, 1)
image_prompt_embeds = image_prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
uncond_image_prompt_embeds = uncond_image_prompt_embeds.repeat(1, num_images_per_prompt, 1)
uncond_image_prompt_embeds = uncond_image_prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
return (image_prompt_embeds, uncond_image_prompt_embeds)
def set_face_fidelity_scale(self, scale):
for attn_processor in self.unet.attn_processors.values():
if isinstance(attn_processor, IPAttnProcessor):
attn_processor.scale = scale
################################
@torch.no_grad()
@replace_example_docstring(EXAMPLE_DOC_STRING)
def __call__(
self,
prompt: Union[str, List[str]] = None,
height: Optional[int] = None,
width: Optional[int] = None,
num_inference_steps: int = 50,
denoising_end: Optional[float] = None,
guidance_scale: float = 5.0,
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.FloatTensor] = None,
prompt_embeds: Optional[torch.FloatTensor] = None,
negative_prompt_embeds: Optional[torch.FloatTensor] = None,
pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
negative_pooled_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,
guidance_rescale: float = 0.0,
original_size: Optional[Tuple[int, int]] = None,
crops_coords_top_left: Tuple[int, int] = (0, 0),
target_size: Optional[Tuple[int, int]] = None,
use_dynamic_threshold: Optional[bool] = False,
face_crop_image: Optional[PipelineImageInput] = None,
face_insightface_embeds: Optional[torch.FloatTensor] = None,
):
r"""
Function invoked when calling the pipeline for generation.
Args:
prompt (`str` or `List[str]`, *optional*):
The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
instead.
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.
denoising_end (`float`, *optional*):
When specified, determines the fraction (between 0.0 and 1.0) of the total denoising process to be
completed before it is intentionally prematurely terminated. For instance, if denoising_end is set to
0.7 and `num_inference_steps` is fixed at 50, the process will execute only 35 (i.e., 0.7 * 50)
Denoisers" multi-pipeline setup, as elaborated in [**Refining the Image
Output**](https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion/stable_diffusion_xl#refining-the-image-output)
guidance_scale (`float`, *optional*, defaults to 7.5):
`guidance_scale` is defined as `w` of equation 2. of [Imagen
Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
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`).
num_images_per_prompt (`int`, *optional*, defaults to 1):
The number of images to generate per prompt.
eta (`float`, *optional*, defaults to 0.0):
Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
[`schedulers.DDIMScheduler`], will be ignored for others.
generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
to make generation deterministic.
latents (`torch.FloatTensor`, *optional*):
Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
tensor will ge generated by sampling using the supplied random `generator`.
prompt_embeds (`torch.FloatTensor`, *optional*):
Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
provided, text embeddings will be generated from `prompt` input argument.
negative_prompt_embeds (`torch.FloatTensor`, *optional*):
Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
argument.
pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
If not provided, pooled text embeddings will be generated from `prompt` input argument.
negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
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.StableDiffusionXLPipelineOutput`] instead of a
callback (`Callable`, *optional*):
A function that will be called every `callback_steps` steps during inference. The function will be
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 under
`self.processor` in
[diffusers.cross_attention](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/cross_attention.py).
guidance_rescale (`float`, *optional*, defaults to 0.7):
Guidance rescale factor proposed by [Common Diffusion Noise Schedules and Sample Steps are
Flawed](https://arxiv.org/pdf/2305.08891.pdf) `guidance_scale` is defined as `φ` in equation 16. of
[Common Diffusion Noise Schedules and Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf).
Guidance rescale factor should fix overexposure when using zero terminal SNR.
original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
TODO
crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)):
TODO
target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
TODO
Examples:
Returns:
[`~pipelines.stable_diffusion.StableDiffusionXLPipelineOutput`] or `tuple`:
[`~pipelines.stable_diffusion.StableDiffusionXLPipelineOutput`] 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`.
"""
# 0. Default height and width to unet
height = height or self.default_sample_size * self.vae_scale_factor
width = width or self.default_sample_size * self.vae_scale_factor
original_size = original_size or (height, width)
target_size = target_size or (height, width)
# 1. Check inputs. Raise error if not correct
self.check_inputs(
prompt,
height,
width,
callback_steps,
negative_prompt,
prompt_embeds,
negative_prompt_embeds,
pooled_prompt_embeds,
negative_pooled_prompt_embeds,
)
# 2. Define call parameters
if prompt is not None and isinstance(prompt, str):
batch_size = 1
elif prompt is not None and isinstance(prompt, list):
batch_size = len(prompt)
else:
batch_size = prompt_embeds.shape[0]
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,
negative_prompt_embeds,
pooled_prompt_embeds,
negative_pooled_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,
pooled_prompt_embeds=pooled_prompt_embeds,
negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
lora_scale=text_encoder_lora_scale,
)
##### prepare fused face embeds
if face_crop_image is not None and face_insightface_embeds is not None:
image_prompt_embeds, uncond_image_prompt_embeds = self.get_fused_face_embedds(
face_insightface_embeds = face_insightface_embeds,
face_crop_image = face_crop_image,
num_images_per_prompt = num_images_per_prompt,
device = device
)
prompt_embeds = torch.cat([prompt_embeds, image_prompt_embeds], dim=1)
negative_prompt_embeds = torch.cat([negative_prompt_embeds, uncond_image_prompt_embeds], dim=1)
# 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. Prepare added time ids & embeddings
add_text_embeds = pooled_prompt_embeds
add_time_ids = self._get_add_time_ids(
original_size, crops_coords_top_left, target_size, dtype=prompt_embeds.dtype
)
if do_classifier_free_guidance:
prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0)
add_time_ids = torch.cat([add_time_ids, add_time_ids], dim=0)
prompt_embeds = prompt_embeds.to(device)
add_text_embeds = add_text_embeds.to(device)
add_time_ids = add_time_ids.to(device).repeat(batch_size * num_images_per_prompt, 1)
# 8. Denoising loop
num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
# 7.1 Apply denoising_end
if denoising_end is not None:
num_inference_steps = int(round(denoising_end * num_inference_steps))
timesteps = timesteps[: num_warmup_steps + self.scheduler.order * num_inference_steps]
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
added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids}
noise_pred = self.unet(
latent_model_input,
t,
encoder_hidden_states=prompt_embeds,
cross_attention_kwargs=cross_attention_kwargs,
added_cond_kwargs=added_cond_kwargs,
return_dict=False,
)[0]
# 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 use_dynamic_threshold:
DynamicThresh = DynThresh(maxSteps=num_inference_steps, experiment_mode=0)
noise_pred = DynamicThresh.dynthresh(noise_pred_text,
noise_pred_uncond,
guidance_scale,
None)
if do_classifier_free_guidance and guidance_rescale > 0.0:
# Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf
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, return_dict=False)[0]
# 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)
# make sureo the VAE is in float32 mode, as it overflows in float16
# torch.cuda.empty_cache()
if self.vae.dtype == torch.float16 and self.vae.config.force_upcast:
self.upcast_vae()
latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
if not output_type == "latent":
latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
else:
image = latents
return StableDiffusionXLPipelineOutput(images=image)
# image = self.watermark.apply_watermark(image)
image = self.image_processor.postprocess(image, output_type=output_type)
# 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,)
return StableDiffusionXLPipelineOutput(images=image)
if __name__ == "__main__":
pass
Kolors-FaceID/requirements.txt 0 → 100644
View file @ 0a1fba1c
accelerate
diffusers
invisible_watermark
torch
transformers
sentencepiece
torchvision
timm
einops
scipy
insightface
opencv-python
gradio==3.40.0
\ No newline at end of file
hf_down.py 0 → 100644
View file @ 0a1fba1c
# pip install huggingface-cli
import os
os.environ['HF_ENDPOINT'] = 'https://hf-mirror.com'
model_list = [
"Kwai-Kolors/Kolors",
"Kwai-Kolors/Kolors-IP-Adapter-FaceID-Plus"
]
for model_path in model_list:
os.system(
f"huggingface-cli download --resume-download {model_path} --local-dir ./{model_path} --local-dir-use-symlinks False")
start.sh 0 → 100644
View file @ 0a1fba1c
#!/bin/bash
cd /root/Kolors-FaceID
python app.py
启动器.ipynb 0 → 100644
View file @ 0a1fba1c
{
"cells": [
{
"cell_type": "markdown",
"id": "e5c5a211-2ccd-4341-af10-ac546484b91f",
"metadata": {
"tags": []
},
"source": [
"## 项目介绍\n",
"- 原项目地址：https://huggingface.co/spaces/Kwai-Kolors/Kolors-FaceID\n",
"- Kolors-FaceID：基于Kolors的个人写真生成器，仅需一张照片、中文提示词即可生成\n",
"## 使用说明\n",
"- 启动需要加载模型，需要2分钟左右的时间\n",
"- 启动和重启 Notebook 点上方工具栏中的「重启并运行所有单元格」。出现如下内容就算成功了：\n",
" - `Running on local URL: http://0.0.0.0:7860`\n",
" - `Running on public URL: https://xxxxxxxxxxxxxxx.gradio.live`\n",
"- 通过以下方式开启页面：\n",
" - 控制台打开「自定义服务」了，访问自定义服务端口号设置为7860\n",
" - 直接打开显示的公开链接`public URL`\n",
"\n"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "53a96614-e2d2-4710-a82b-0d5ca9cb9872",
"metadata": {
"tags": [],
"is_executing": true
},
"outputs": [],
"source": [
"# 启动\n",
"!sh start.sh"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "9e81ae9d-3a34-43a0-943a-ff5e9d6ce961",
"metadata": {},
"outputs": [],
"source": []
}
],
"metadata": {
"kernelspec": {
"display_name": "Python 3 (ipykernel)",
"language": "python",
"name": "python3"
},
"language_info": {
"codemirror_mode": {
"name": "ipython",
"version": 3
},
"file_extension": ".py",
"mimetype": "text/x-python",
"name": "python",
"nbconvert_exporter": "python",
"pygments_lexer": "ipython3",
"version": "3.10.13"
}
},
"nbformat": 4,
"nbformat_minor": 5
}
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment