Upload pipeline.py

0bd0ab1 verified about 2 months ago

No virus

55.4 kB

	# Implementation of Stable Diffusion Upscale Pipeline with Perturbed-Attention Guidance

	import inspect
	import warnings
	from typing import Any, Callable, Dict, List, Optional, Union

	import numpy as np
	import PIL.Image
	import torch
	import torch.nn.functional as F
	from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer

	from diffusers.image_processor import PipelineImageInput, VaeImageProcessor
	from diffusers.loaders import FromSingleFileMixin, LoraLoaderMixin, TextualInversionLoaderMixin
	from diffusers.models import AutoencoderKL, UNet2DConditionModel
	from diffusers.models.attention_processor import (
	Attention,
	AttnProcessor2_0,
	LoRAAttnProcessor2_0,
	LoRAXFormersAttnProcessor,
	XFormersAttnProcessor,
	)
	from diffusers.models.lora import adjust_lora_scale_text_encoder
	from diffusers.schedulers import DDPMScheduler, KarrasDiffusionSchedulers
	from diffusers.utils import USE_PEFT_BACKEND, deprecate, logging, 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 import StableDiffusionPipelineOutput


	logger = logging.get_logger(__name__) # pylint: disable=invalid-name

	class PAGIdentitySelfAttnProcessor:
	r"""
	Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0).
	"""

	def __init__(self):
	if not hasattr(F, "scaled_dot_product_attention"):
	raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")

	def __call__(
	self,
	attn: Attention,
	hidden_states: torch.FloatTensor,
	encoder_hidden_states: Optional[torch.FloatTensor] = None,
	attention_mask: Optional[torch.FloatTensor] = None,
	temb: Optional[torch.FloatTensor] = None,
	*args,
	**kwargs,
	) -> torch.FloatTensor:
	if len(args) > 0 or kwargs.get("scale", None) is not None:
	deprecation_message = "The `scale` argument is deprecated and will be ignored. Please remove it, as passing it will raise an error in the future. `scale` should directly be passed while calling the underlying pipeline component i.e., via `cross_attention_kwargs`."
	deprecate("scale", "1.0.0", deprecation_message)

	residual = hidden_states
	if attn.spatial_norm is not None:
	hidden_states = attn.spatial_norm(hidden_states, temb)

	input_ndim = hidden_states.ndim
	if input_ndim == 4:
	batch_size, channel, height, width = hidden_states.shape
	hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)

	# chunk
	hidden_states_org, hidden_states_ptb = hidden_states.chunk(2)

	# original path
	batch_size, sequence_length, _ = hidden_states_org.shape

	if attention_mask is not None:
	attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
	# scaled_dot_product_attention expects attention_mask shape to be
	# (batch, heads, source_length, target_length)
	attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])

	if attn.group_norm is not None:
	hidden_states_org = attn.group_norm(hidden_states_org.transpose(1, 2)).transpose(1, 2)

	query = attn.to_q(hidden_states_org)
	key = attn.to_k(hidden_states_org)
	value = attn.to_v(hidden_states_org)

	inner_dim = key.shape[-1]
	head_dim = inner_dim // attn.heads

	query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)

	key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
	value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)

	# the output of sdp = (batch, num_heads, seq_len, head_dim)
	# TODO: add support for attn.scale when we move to Torch 2.1
	hidden_states_org = F.scaled_dot_product_attention(
	query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
	)

	hidden_states_org = hidden_states_org.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
	hidden_states_org = hidden_states_org.to(query.dtype)

	# linear proj
	hidden_states_org = attn.to_out[0](hidden_states_org)
	# dropout
	hidden_states_org = attn.to_out[1](hidden_states_org)

	if input_ndim == 4:
	hidden_states_org = hidden_states_org.transpose(-1, -2).reshape(batch_size, channel, height, width)

	# perturbed path (identity attention)
	batch_size, sequence_length, _ = hidden_states_ptb.shape

	if attention_mask is not None:
	attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
	# scaled_dot_product_attention expects attention_mask shape to be
	# (batch, heads, source_length, target_length)
	attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])

	if attn.group_norm is not None:
	hidden_states_ptb = attn.group_norm(hidden_states_ptb.transpose(1, 2)).transpose(1, 2)

	value = attn.to_v(hidden_states_ptb)

	# hidden_states_ptb = torch.zeros(value.shape).to(value.get_device())
	hidden_states_ptb = value

	hidden_states_ptb = hidden_states_ptb.to(query.dtype)

	# linear proj
	hidden_states_ptb = attn.to_out[0](hidden_states_ptb)
	# dropout
	hidden_states_ptb = attn.to_out[1](hidden_states_ptb)

	if input_ndim == 4:
	hidden_states_ptb = hidden_states_ptb.transpose(-1, -2).reshape(batch_size, channel, height, width)

	# cat
	hidden_states = torch.cat([hidden_states_org, hidden_states_ptb])

	if attn.residual_connection:
	hidden_states = hidden_states + residual

	hidden_states = hidden_states / attn.rescale_output_factor

	return hidden_states


	class PAGCFGIdentitySelfAttnProcessor:
	r"""
	Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0).
	"""

	def __init__(self):
	if not hasattr(F, "scaled_dot_product_attention"):
	raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")

	def __call__(
	self,
	attn: Attention,
	hidden_states: torch.FloatTensor,
	encoder_hidden_states: Optional[torch.FloatTensor] = None,
	attention_mask: Optional[torch.FloatTensor] = None,
	temb: Optional[torch.FloatTensor] = None,
	*args,
	**kwargs,
	) -> torch.FloatTensor:
	if len(args) > 0 or kwargs.get("scale", None) is not None:
	deprecation_message = "The `scale` argument is deprecated and will be ignored. Please remove it, as passing it will raise an error in the future. `scale` should directly be passed while calling the underlying pipeline component i.e., via `cross_attention_kwargs`."
	deprecate("scale", "1.0.0", deprecation_message)

	residual = hidden_states
	if attn.spatial_norm is not None:
	hidden_states = attn.spatial_norm(hidden_states, temb)

	input_ndim = hidden_states.ndim
	if input_ndim == 4:
	batch_size, channel, height, width = hidden_states.shape
	hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)

	# chunk
	hidden_states_uncond, hidden_states_org, hidden_states_ptb = hidden_states.chunk(3)
	hidden_states_org = torch.cat([hidden_states_uncond, hidden_states_org])

	# original path
	batch_size, sequence_length, _ = hidden_states_org.shape

	if attention_mask is not None:
	attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
	# scaled_dot_product_attention expects attention_mask shape to be
	# (batch, heads, source_length, target_length)
	attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])

	if attn.group_norm is not None:
	hidden_states_org = attn.group_norm(hidden_states_org.transpose(1, 2)).transpose(1, 2)

	query = attn.to_q(hidden_states_org)
	key = attn.to_k(hidden_states_org)
	value = attn.to_v(hidden_states_org)

	inner_dim = key.shape[-1]
	head_dim = inner_dim // attn.heads

	query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)

	key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
	value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)

	# the output of sdp = (batch, num_heads, seq_len, head_dim)
	# TODO: add support for attn.scale when we move to Torch 2.1
	hidden_states_org = F.scaled_dot_product_attention(
	query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
	)

	hidden_states_org = hidden_states_org.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
	hidden_states_org = hidden_states_org.to(query.dtype)

	# linear proj
	hidden_states_org = attn.to_out[0](hidden_states_org)
	# dropout
	hidden_states_org = attn.to_out[1](hidden_states_org)

	if input_ndim == 4:
	hidden_states_org = hidden_states_org.transpose(-1, -2).reshape(batch_size, channel, height, width)

	# perturbed path (identity attention)
	batch_size, sequence_length, _ = hidden_states_ptb.shape

	if attention_mask is not None:
	attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
	# scaled_dot_product_attention expects attention_mask shape to be
	# (batch, heads, source_length, target_length)
	attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])

	if attn.group_norm is not None:
	hidden_states_ptb = attn.group_norm(hidden_states_ptb.transpose(1, 2)).transpose(1, 2)

	value = attn.to_v(hidden_states_ptb)
	hidden_states_ptb = value
	hidden_states_ptb = hidden_states_ptb.to(query.dtype)

	# linear proj
	hidden_states_ptb = attn.to_out[0](hidden_states_ptb)
	# dropout
	hidden_states_ptb = attn.to_out[1](hidden_states_ptb)

	if input_ndim == 4:
	hidden_states_ptb = hidden_states_ptb.transpose(-1, -2).reshape(batch_size, channel, height, width)

	# cat
	hidden_states = torch.cat([hidden_states_org, hidden_states_ptb])

	if attn.residual_connection:
	hidden_states = hidden_states + residual

	hidden_states = hidden_states / attn.rescale_output_factor

	return hidden_states

	def preprocess(image):
	warnings.warn(
	"The preprocess method is deprecated and will be removed in a future version. Please"
	" use VaeImageProcessor.preprocess instead",
	FutureWarning,
	)
	if isinstance(image, torch.Tensor):
	return image
	elif isinstance(image, PIL.Image.Image):
	image = [image]

	if isinstance(image[0], PIL.Image.Image):
	w, h = image[0].size
	w, h = (x - x % 64 for x in (w, h)) # resize to integer multiple of 64

	image = [np.array(i.resize((w, h)))[None, :] for i in image]
	image = np.concatenate(image, axis=0)
	image = np.array(image).astype(np.float32) / 255.0
	image = image.transpose(0, 3, 1, 2)
	image = 2.0 * image - 1.0
	image = torch.from_numpy(image)
	elif isinstance(image[0], torch.Tensor):
	image = torch.cat(image, dim=0)
	return image


	class StableDiffusionUpscalePipeline(
	DiffusionPipeline, StableDiffusionMixin, TextualInversionLoaderMixin, LoraLoaderMixin, FromSingleFileMixin
	):
	r"""
	Pipeline for text-guided image super-resolution using Stable Diffusion 2.

	This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
	implemented for all pipelines (downloading, saving, running on a particular device, etc.).

	The pipeline also inherits the following loading methods:
	- [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
	- [`~loaders.LoraLoaderMixin.load_lora_weights`] for loading LoRA weights
	- [`~loaders.LoraLoaderMixin.save_lora_weights`] for saving LoRA weights
	- [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files

	Args:
	vae ([`AutoencoderKL`]):
	Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations.
	text_encoder ([`~transformers.CLIPTextModel`]):
	Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
	tokenizer ([`~transformers.CLIPTokenizer`]):
	A `CLIPTokenizer` to tokenize text.
	unet ([`UNet2DConditionModel`]):
	A `UNet2DConditionModel` to denoise the encoded image latents.
	low_res_scheduler ([`SchedulerMixin`]):
	A scheduler used to add initial noise to the low resolution conditioning image. It must be an instance of
	[`DDPMScheduler`].
	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`].
	"""

	model_cpu_offload_seq = "text_encoder->unet->vae"
	_optional_components = ["watermarker", "safety_checker", "feature_extractor"]
	_exclude_from_cpu_offload = ["safety_checker"]

	def __init__(
	self,
	vae: AutoencoderKL,
	text_encoder: CLIPTextModel,
	tokenizer: CLIPTokenizer,
	unet: UNet2DConditionModel,
	low_res_scheduler: DDPMScheduler,
	scheduler: KarrasDiffusionSchedulers,
	safety_checker: Optional[Any] = None,
	feature_extractor: Optional[CLIPImageProcessor] = None,
	watermarker: Optional[Any] = None,
	max_noise_level: int = 350,
	):
	super().__init__()

	if hasattr(
	vae, "config"
	): # check if vae has a config attribute `scaling_factor` and if it is set to 0.08333, else set it to 0.08333 and deprecate
	is_vae_scaling_factor_set_to_0_08333 = (
	hasattr(vae.config, "scaling_factor") and vae.config.scaling_factor == 0.08333
	)
	if not is_vae_scaling_factor_set_to_0_08333:
	deprecation_message = (
	"The configuration file of the vae does not contain `scaling_factor` or it is set to"
	f" {vae.config.scaling_factor}, which seems highly unlikely. If your checkpoint is a fine-tuned"
	" version of `stabilityai/stable-diffusion-x4-upscaler` you should change 'scaling_factor' to"
	" 0.08333 Please make sure to update the config accordingly, as not doing so might lead to"
	" incorrect results in future versions. If you have downloaded this checkpoint from the Hugging"
	" Face Hub, it would be very nice if you could open a Pull Request for the `vae/config.json` file"
	)
	deprecate("wrong scaling_factor", "1.0.0", deprecation_message, standard_warn=False)
	vae.register_to_config(scaling_factor=0.08333)

	self.register_modules(
	vae=vae,
	text_encoder=text_encoder,
	tokenizer=tokenizer,
	unet=unet,
	low_res_scheduler=low_res_scheduler,
	scheduler=scheduler,
	safety_checker=safety_checker,
	watermarker=watermarker,
	feature_extractor=feature_extractor,
	)
	self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
	self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor, resample="bicubic")
	self.register_to_config(max_noise_level=max_noise_level)

	def run_safety_checker(self, image, device, dtype):
	if self.safety_checker is not None:
	feature_extractor_input = self.image_processor.postprocess(image, output_type="pil")
	safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device)
	image, nsfw_detected, watermark_detected = self.safety_checker(
	images=image,
	clip_input=safety_checker_input.pixel_values.to(dtype=dtype),
	)
	else:
	nsfw_detected = None
	watermark_detected = None

	if hasattr(self, "unet_offload_hook") and self.unet_offload_hook is not None:
	self.unet_offload_hook.offload()

	return image, nsfw_detected, watermark_detected

	# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt
	def _encode_prompt(
	self,
	prompt,
	device,
	num_images_per_prompt,
	do_classifier_free_guidance,
	negative_prompt=None,
	prompt_embeds: Optional[torch.FloatTensor] = None,
	negative_prompt_embeds: Optional[torch.FloatTensor] = None,
	lora_scale: Optional[float] = None,
	**kwargs,
	):
	deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple."
	deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False)

	prompt_embeds_tuple = self.encode_prompt(
	prompt=prompt,
	device=device,
	num_images_per_prompt=num_images_per_prompt,
	do_classifier_free_guidance=do_classifier_free_guidance,
	negative_prompt=negative_prompt,
	prompt_embeds=prompt_embeds,
	negative_prompt_embeds=negative_prompt_embeds,
	lora_scale=lora_scale,
	**kwargs,
	)

	# concatenate for backwards comp
	prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]])

	return prompt_embeds

	# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt
	def encode_prompt(
	self,
	prompt,
	device,
	num_images_per_prompt,
	do_classifier_free_guidance,
	negative_prompt=None,
	prompt_embeds: Optional[torch.FloatTensor] = None,
	negative_prompt_embeds: Optional[torch.FloatTensor] = None,
	lora_scale: Optional[float] = None,
	clip_skip: Optional[int] = None,
	):
	r"""
	Encodes the prompt into text encoder hidden states.

	Args:
	prompt (`str` or `List[str]`, optional):
	prompt to be encoded
	device: (`torch.device`):
	torch device
	num_images_per_prompt (`int`):
	number of images that should be generated per prompt
	do_classifier_free_guidance (`bool`):
	whether to use classifier free guidance or not
	negative_prompt (`str` or `List[str]`, optional):
	The prompt or prompts not to guide the image generation. If not defined, one has to pass
	`negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
	less than `1`).
	prompt_embeds (`torch.FloatTensor`, optional):
	Pre-generated text embeddings. Can be used to easily tweak text inputs, e.g. prompt weighting. If not
	provided, text embeddings will be generated from `prompt` input argument.
	negative_prompt_embeds (`torch.FloatTensor`, optional):
	Pre-generated negative text embeddings. Can be used to easily tweak text inputs, e.g. prompt
	weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
	argument.
	lora_scale (`float`, optional):
	A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
	clip_skip (`int`, optional):
	Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
	the output of the pre-final layer will be used for computing the prompt embeddings.
	"""
	# set lora scale so that monkey patched LoRA
	# function of text encoder can correctly access it
	if lora_scale is not None and isinstance(self, LoraLoaderMixin):
	self._lora_scale = lora_scale

	# dynamically adjust the LoRA scale
	if not USE_PEFT_BACKEND:
	adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
	else:
	scale_lora_layers(self.text_encoder, lora_scale)

	if prompt is not None and isinstance(prompt, str):
	batch_size = 1
	elif prompt is not None and isinstance(prompt, list):
	batch_size = len(prompt)
	else:
	batch_size = prompt_embeds.shape[0]

	if prompt_embeds is None:
	# textual inversion: process multi-vector tokens if necessary
	if isinstance(self, TextualInversionLoaderMixin):
	prompt = self.maybe_convert_prompt(prompt, self.tokenizer)

	text_inputs = self.tokenizer(
	prompt,
	padding="max_length",
	max_length=self.tokenizer.model_max_length,
	truncation=True,
	return_tensors="pt",
	)
	text_input_ids = text_inputs.input_ids
	untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids

	if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
	text_input_ids, untruncated_ids
	):
	removed_text = self.tokenizer.batch_decode(
	untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
	)
	logger.warning(
	"The following part of your input was truncated because CLIP can only handle sequences up to"
	f" {self.tokenizer.model_max_length} tokens: {removed_text}"
	)

	if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
	attention_mask = text_inputs.attention_mask.to(device)
	else:
	attention_mask = None

	if clip_skip is None:
	prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask)
	prompt_embeds = prompt_embeds[0]
	else:
	prompt_embeds = self.text_encoder(
	text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True
	)
	# Access the `hidden_states` first, that contains a tuple of
	# all the hidden states from the encoder layers. Then index into
	# the tuple to access the hidden states from the desired layer.
	prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)]
	# We also need to apply the final LayerNorm here to not mess with the
	# representations. The `last_hidden_states` that we typically use for
	# obtaining the final prompt representations passes through the LayerNorm
	# layer.
	prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds)

	if self.text_encoder is not None:
	prompt_embeds_dtype = self.text_encoder.dtype
	elif self.unet is not None:
	prompt_embeds_dtype = self.unet.dtype
	else:
	prompt_embeds_dtype = prompt_embeds.dtype

	prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)

	bs_embed, seq_len, _ = prompt_embeds.shape
	# duplicate text embeddings for each generation per prompt, using mps friendly method
	prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
	prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)

	# get unconditional embeddings for classifier free guidance
	if do_classifier_free_guidance and negative_prompt_embeds is None:
	uncond_tokens: List[str]
	if negative_prompt is None:
	uncond_tokens = [""] * batch_size
	elif prompt is not None and type(prompt) is not type(negative_prompt):
	raise TypeError(
	f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
	f" {type(prompt)}."
	)
	elif isinstance(negative_prompt, str):
	uncond_tokens = [negative_prompt]
	elif batch_size != len(negative_prompt):
	raise ValueError(
	f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
	f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
	" the batch size of `prompt`."
	)
	else:
	uncond_tokens = negative_prompt

	# textual inversion: process multi-vector tokens if necessary
	if isinstance(self, TextualInversionLoaderMixin):
	uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer)

	max_length = prompt_embeds.shape[1]
	uncond_input = self.tokenizer(
	uncond_tokens,
	padding="max_length",
	max_length=max_length,
	truncation=True,
	return_tensors="pt",
	)

	if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
	attention_mask = uncond_input.attention_mask.to(device)
	else:
	attention_mask = None

	negative_prompt_embeds = self.text_encoder(
	uncond_input.input_ids.to(device),
	attention_mask=attention_mask,
	)
	negative_prompt_embeds = negative_prompt_embeds[0]

	if do_classifier_free_guidance:
	# duplicate unconditional embeddings for each generation per prompt, using mps friendly method
	seq_len = negative_prompt_embeds.shape[1]

	negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)

	negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
	negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)

	if isinstance(self, LoraLoaderMixin) and USE_PEFT_BACKEND:
	# Retrieve the original scale by scaling back the LoRA layers
	unscale_lora_layers(self.text_encoder, lora_scale)

	return prompt_embeds, negative_prompt_embeds

	# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
	def prepare_extra_step_kwargs(self, generator, eta):
	# prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
	# eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
	# eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
	# and should be between [0, 1]

	accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
	extra_step_kwargs = {}
	if accepts_eta:
	extra_step_kwargs["eta"] = eta

	# check if the scheduler accepts generator
	accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
	if accepts_generator:
	extra_step_kwargs["generator"] = generator
	return extra_step_kwargs

	# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents
	def decode_latents(self, latents):
	deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead"
	deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)

	latents = 1 / self.vae.config.scaling_factor * latents
	image = self.vae.decode(latents, return_dict=False)[0]
	image = (image / 2 + 0.5).clamp(0, 1)
	# we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
	image = image.cpu().permute(0, 2, 3, 1).float().numpy()
	return image

	def check_inputs(
	self,
	prompt,
	image,
	noise_level,
	callback_steps,
	negative_prompt=None,
	prompt_embeds=None,
	negative_prompt_embeds=None,
	):
	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 (
	not isinstance(image, torch.Tensor)
	and not isinstance(image, PIL.Image.Image)
	and not isinstance(image, np.ndarray)
	and not isinstance(image, list)
	):
	raise ValueError(
	f"`image` has to be of type `torch.Tensor`, `np.ndarray`, `PIL.Image.Image` or `list` but is {type(image)}"
	)

	# verify batch size of prompt and image are same if image is a list or tensor or numpy array
	if isinstance(image, list) or isinstance(image, torch.Tensor) or isinstance(image, np.ndarray):
	if prompt is not None and isinstance(prompt, str):
	batch_size = 1
	elif prompt is not None and isinstance(prompt, list):
	batch_size = len(prompt)
	else:
	batch_size = prompt_embeds.shape[0]

	if isinstance(image, list):
	image_batch_size = len(image)
	else:
	image_batch_size = image.shape[0]
	if batch_size != image_batch_size:
	raise ValueError(
	f"`prompt` has batch size {batch_size} and `image` has batch size {image_batch_size}."
	" Please make sure that passed `prompt` matches the batch size of `image`."
	)

	# check noise level
	if noise_level > self.config.max_noise_level:
	raise ValueError(f"`noise_level` has to be <= {self.config.max_noise_level} but is {noise_level}")

	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)}."
	)

	def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
	shape = (batch_size, num_channels_latents, height, width)
	if latents is None:
	latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
	else:
	if latents.shape != shape:
	raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}")
	latents = latents.to(device)

	# scale the initial noise by the standard deviation required by the scheduler
	latents = latents * self.scheduler.init_noise_sigma
	return latents

	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)

	@property
	def guidance_scale(self):
	return self._guidance_scale

	@property
	def do_classifier_free_guidance(self):
	return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None

	@property
	def pag_scale(self):
	return self._pag_scale

	@property
	def do_perturbed_attention_guidance(self):
	return self._pag_scale > 0

	@property
	def pag_adaptive_scaling(self):
	return self._pag_adaptive_scaling

	@property
	def do_pag_adaptive_scaling(self):
	return self._pag_adaptive_scaling > 0

	@property
	def pag_applied_layers_index(self):
	return self._pag_applied_layers_index

	@torch.no_grad()
	def __call__(
	self,
	prompt: Union[str, List[str]] = None,
	image: PipelineImageInput = None,
	num_inference_steps: int = 75,
	guidance_scale: float = 9.0,
	pag_scale: float = 0.0,
	pag_adaptive_scaling: float = 0.0,
	pag_applied_layers_index: List[str] = ["d4"], # ['d4', 'd5', 'm0']
	noise_level: int = 20,
	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,
	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,
	clip_skip: int = None,
	):
	r"""
	The call function to the pipeline for generation.

	Args:
	prompt (`str` or `List[str]`, optional):
	The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`.
	image (`torch.FloatTensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.FloatTensor]`, `List[PIL.Image.Image]`, or `List[np.ndarray]`):
	`Image` or tensor representing an image batch to be upscaled.
	num_inference_steps (`int`, optional, defaults to 50):
	The number of denoising steps. More denoising steps usually lead to a higher quality image at the
	expense of slower inference.
	guidance_scale (`float`, optional, defaults to 7.5):
	A higher guidance scale value encourages the model to generate images closely linked to the text
	`prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
	negative_prompt (`str` or `List[str]`, optional):
	The prompt or prompts to guide what to not include in image generation. If not defined, you need to
	pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 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) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
	to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
	generator (`torch.Generator` or `List[torch.Generator]`, optional):
	A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
	generation deterministic.
	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 is 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 (prompt weighting). If not
	provided, text embeddings are generated from the `prompt` input argument.
	negative_prompt_embeds (`torch.FloatTensor`, optional):
	Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
	not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
	output_type (`str`, optional, defaults to `"pil"`):
	The output format of the generated image. Choose between `PIL.Image` or `np.array`.
	return_dict (`bool`, optional, defaults to `True`):
	Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
	plain tuple.
	callback (`Callable`, optional):
	A function that calls every `callback_steps` steps during inference. The function is called with the
	following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
	callback_steps (`int`, optional, defaults to 1):
	The frequency at which the `callback` function is called. If not specified, the callback is called at
	every step.
	cross_attention_kwargs (`dict`, optional):
	A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in
	[`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
	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.
	Examples:
	```py
	>>> import requests
	>>> from PIL import Image
	>>> from io import BytesIO
	>>> from diffusers import StableDiffusionUpscalePipeline
	>>> import torch

	>>> # load model and scheduler
	>>> model_id = "stabilityai/stable-diffusion-x4-upscaler"
	>>> pipeline = StableDiffusionUpscalePipeline.from_pretrained(
	... model_id, revision="fp16", torch_dtype=torch.float16
	... )
	>>> pipeline = pipeline.to("cuda")

	>>> # let's download an image
	>>> url = "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/sd2-upscale/low_res_cat.png"
	>>> response = requests.get(url)
	>>> low_res_img = Image.open(BytesIO(response.content)).convert("RGB")
	>>> low_res_img = low_res_img.resize((128, 128))
	>>> prompt = "a white cat"

	>>> upscaled_image = pipeline(prompt=prompt, image=low_res_img).images[0]
	>>> upscaled_image.save("upsampled_cat.png")
	```

	Returns:
	[`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
	If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned,
	otherwise a `tuple` is returned where the first element is a list with the generated images and the
	second element is a list of `bool`s indicating whether the corresponding generated image contains
	"not-safe-for-work" (nsfw) content.
	"""

	# 1. Check inputs
	self.check_inputs(
	prompt,
	image,
	noise_level,
	callback_steps,
	negative_prompt,
	prompt_embeds,
	negative_prompt_embeds,
	)

	self._guidance_scale = guidance_scale

	self._pag_scale = pag_scale
	self._pag_adaptive_scaling = pag_adaptive_scaling
	self._pag_applied_layers_index = pag_applied_layers_index

	if image is None:
	raise ValueError("`image` input cannot be undefined.")

	# 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.

	# 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 = self.encode_prompt(
	prompt,
	device,
	num_images_per_prompt,
	self.do_classifier_free_guidance,
	negative_prompt,
	prompt_embeds=prompt_embeds,
	negative_prompt_embeds=negative_prompt_embeds,
	lora_scale=text_encoder_lora_scale,
	clip_skip=clip_skip,
	)
	# 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

	# cfg
	if self.do_classifier_free_guidance and not self.do_perturbed_attention_guidance:
	prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
	# pag
	elif not self.do_classifier_free_guidance and self.do_perturbed_attention_guidance:
	prompt_embeds = torch.cat([prompt_embeds, prompt_embeds])
	# both
	elif self.do_classifier_free_guidance and self.do_perturbed_attention_guidance:
	prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds, prompt_embeds])

	# 4. Preprocess image
	image = self.image_processor.preprocess(image)
	image = image.to(dtype=prompt_embeds.dtype, device=device)

	# 5. set timesteps
	self.scheduler.set_timesteps(num_inference_steps, device=device)
	timesteps = self.scheduler.timesteps

	# 5. Add noise to image
	noise_level = torch.tensor([noise_level], dtype=torch.long, device=device)
	noise = randn_tensor(image.shape, generator=generator, device=device, dtype=prompt_embeds.dtype)
	image = self.low_res_scheduler.add_noise(image, noise, noise_level)

	image = torch.cat([image] * num_images_per_prompt)
	noise_level = torch.cat([noise_level] * image.shape[0])

	# 6. Prepare latent variables
	height, width = image.shape[2:]
	num_channels_latents = self.vae.config.latent_channels
	latents = self.prepare_latents(
	batch_size * num_images_per_prompt,
	num_channels_latents,
	height,
	width,
	prompt_embeds.dtype,
	device,
	generator,
	latents,
	)

	# 7. Check that sizes of image and latents match
	num_channels_image = image.shape[1]
	if num_channels_latents + num_channels_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_image`: {num_channels_image} "
	f" = {num_channels_latents+num_channels_image}. Please verify the config of"
	" `pipeline.unet` or your `image` input."
	)

	# 8. 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)

	# 9. Denoising loop
	if self.do_perturbed_attention_guidance:
	down_layers = []
	mid_layers = []
	up_layers = []
	for name, module in self.unet.named_modules():
	if "attn1" in name and "to" not in name:
	layer_type = name.split(".")[0].split("_")[0]
	if layer_type == "down":
	down_layers.append(module)
	elif layer_type == "mid":
	mid_layers.append(module)
	elif layer_type == "up":
	up_layers.append(module)
	else:
	raise ValueError(f"Invalid layer type: {layer_type}")

	# change attention layer in UNet if use PAG
	if self.do_perturbed_attention_guidance:
	if self.do_classifier_free_guidance:
	replace_processor = PAGCFGIdentitySelfAttnProcessor()
	else:
	replace_processor = PAGIdentitySelfAttnProcessor()

	drop_layers = self.pag_applied_layers_index
	for drop_layer in drop_layers:
	try:
	if drop_layer[0] == "d":
	down_layers[int(drop_layer[1])].processor = replace_processor
	elif drop_layer[0] == "m":
	mid_layers[int(drop_layer[1])].processor = replace_processor
	elif drop_layer[0] == "u":
	up_layers[int(drop_layer[1])].processor = replace_processor
	else:
	raise ValueError(f"Invalid layer type: {drop_layer[0]}")
	except IndexError:
	raise ValueError(
	f"Invalid layer index: {drop_layer}. Available layers: {len(down_layers)} down layers, {len(mid_layers)} mid layers, {len(up_layers)} up layers."
	)

	num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
	with self.progress_bar(total=num_inference_steps) as progress_bar:
	for i, t in enumerate(timesteps):

	# cfg
	if self.do_classifier_free_guidance and not self.do_perturbed_attention_guidance:
	latent_model_input = torch.cat([latents] * 2)
	image_input = torch.cat([image] * 2)
	# pag
	elif not self.do_classifier_free_guidance and self.do_perturbed_attention_guidance:
	latent_model_input = torch.cat([latents] * 2)
	image_input = torch.cat([image] * 2)
	# both
	elif self.do_classifier_free_guidance and self.do_perturbed_attention_guidance:
	latent_model_input = torch.cat([latents] * 3)
	image_input = torch.cat([image] * 3)
	# no
	else:
	latent_model_input = latents
	image_input = image

	# concat latents, mask, masked_image_latents in the channel dimension
	latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
	latent_model_input = torch.cat([latent_model_input, image_input], dim=1)

	# predict the noise residual
	noise_pred = self.unet(
	latent_model_input,
	t,
	encoder_hidden_states=prompt_embeds,
	cross_attention_kwargs=cross_attention_kwargs,
	class_labels=noise_level,
	return_dict=False,
	)[0]

	# perform guidance
	# cfg
	if self.do_classifier_free_guidance and not self.do_perturbed_attention_guidance:
	noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)

	delta = noise_pred_text - noise_pred_uncond
	noise_pred = noise_pred_uncond + self.guidance_scale * delta

	# pag
	elif not self.do_classifier_free_guidance and self.do_perturbed_attention_guidance:
	noise_pred_original, noise_pred_perturb = noise_pred.chunk(2)

	signal_scale = self.pag_scale
	if self.do_pag_adaptive_scaling:
	signal_scale = self.pag_scale - self.pag_adaptive_scaling * (1000 - t)
	if signal_scale < 0:
	signal_scale = 0

	noise_pred = noise_pred_original + signal_scale * (noise_pred_original - noise_pred_perturb)

	# both
	elif self.do_classifier_free_guidance and self.do_perturbed_attention_guidance:
	noise_pred_uncond, noise_pred_text, noise_pred_text_perturb = noise_pred.chunk(3)

	signal_scale = self.pag_scale
	if self.do_pag_adaptive_scaling:
	signal_scale = self.pag_scale - self.pag_adaptive_scaling * (1000 - t)
	if signal_scale < 0:
	signal_scale = 0

	noise_pred = (
	noise_pred_text
	+ (self.guidance_scale - 1.0) * (noise_pred_text - noise_pred_uncond)
	+ signal_scale * (noise_pred_text - noise_pred_text_perturb)
	)

	# 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:
	step_idx = i // getattr(self.scheduler, "order", 1)
	callback(step_idx, t, latents)

	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()

	# Ensure latents are always the same type as the VAE
	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]

	# cast back to fp16 if needed
	if needs_upcasting:
	self.vae.to(dtype=torch.float16)

	image, has_nsfw_concept, _ = self.run_safety_checker(image, device, prompt_embeds.dtype)
	else:
	image = latents
	has_nsfw_concept = None

	if has_nsfw_concept is None:
	do_denormalize = [True] * image.shape[0]
	else:
	do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept]

	image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize)

	# 11. Apply watermark
	if output_type == "pil" and self.watermarker is not None:
	image = self.watermarker.apply_watermark(image)

	# Offload all models
	self.maybe_free_model_hooks()

	# change attention layer in UNet if use PAG
	if self.do_perturbed_attention_guidance:
	drop_layers = self.pag_applied_layers_index
	for drop_layer in drop_layers:
	try:
	if drop_layer[0] == "d":
	down_layers[int(drop_layer[1])].processor = AttnProcessor2_0()
	elif drop_layer[0] == "m":
	mid_layers[int(drop_layer[1])].processor = AttnProcessor2_0()
	elif drop_layer[0] == "u":
	up_layers[int(drop_layer[1])].processor = AttnProcessor2_0()
	else:
	raise ValueError(f"Invalid layer type: {drop_layer[0]}")
	except IndexError:
	raise ValueError(
	f"Invalid layer index: {drop_layer}. Available layers: {len(down_layers)} down layers, {len(mid_layers)} mid layers, {len(up_layers)} up layers."
	)

	if not return_dict:
	return (image, has_nsfw_concept)

	return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)