stable_diffusion_xl_reference.py

# Based on stable_diffusion_reference.py

from typing import Any, Callable, Dict, List, Optional, Tuple, Union

import numpy as np
import PIL.Image
import torch

from diffusers import StableDiffusionXLPipeline
from diffusers.models.attention import BasicTransformerBlock
from diffusers.models.unets.unet_2d_blocks import (
    CrossAttnDownBlock2D,
    CrossAttnUpBlock2D,
    DownBlock2D,
    UpBlock2D,
)
from diffusers.pipelines.stable_diffusion_xl import StableDiffusionXLPipelineOutput
from diffusers.utils import PIL_INTERPOLATION, logging
from diffusers.utils.torch_utils import randn_tensor


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

EXAMPLE_DOC_STRING = """
    Examples:
        ```py
        >>> import torch
        >>> from diffusers import UniPCMultistepScheduler
        >>> from diffusers.utils import load_image

        >>> input_image = load_image("https://hf.co/datasets/huggingface/documentation-images/resolve/main/diffusers/input_image_vermeer.png")

        >>> pipe = StableDiffusionXLReferencePipeline.from_pretrained(
            "stabilityai/stable-diffusion-xl-base-1.0",
            torch_dtype=torch.float16,
            use_safetensors=True,
            variant="fp16").to('cuda:0')

        >>> pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config)
        >>> result_img = pipe(ref_image=input_image,
                        prompt="1girl",
                        num_inference_steps=20,
                        reference_attn=True,
                        reference_adain=True).images[0]

        >>> result_img.show()
        ```
"""


def torch_dfs(model: torch.nn.Module):
    result = [model]
    for child in model.children():
        result += torch_dfs(child)
    return result


# 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 StableDiffusionXLReferencePipeline(StableDiffusionXLPipeline):
    def _default_height_width(self, height, width, image):
        # NOTE: It is possible that a list of images have different
        # dimensions for each image, so just checking the first image
        # is not _exactly_ correct, but it is simple.
        while isinstance(image, list):
            image = image[0]

        if height is None:
            if isinstance(image, PIL.Image.Image):
                height = image.height
            elif isinstance(image, torch.Tensor):
                height = image.shape[2]

            height = (height // 8) * 8  # round down to nearest multiple of 8

        if width is None:
            if isinstance(image, PIL.Image.Image):
                width = image.width
            elif isinstance(image, torch.Tensor):
                width = image.shape[3]

            width = (width // 8) * 8

        return height, width

    def prepare_image(
        self,
        image,
        width,
        height,
        batch_size,
        num_images_per_prompt,
        device,
        dtype,
        do_classifier_free_guidance=False,
        guess_mode=False,
    ):
        if not isinstance(image, torch.Tensor):
            if isinstance(image, PIL.Image.Image):
                image = [image]

            if isinstance(image[0], PIL.Image.Image):
                images = []

                for image_ in image:
                    image_ = image_.convert("RGB")
                    image_ = image_.resize((width, height), resample=PIL_INTERPOLATION["lanczos"])
                    image_ = np.array(image_)
                    image_ = image_[None, :]
                    images.append(image_)

                image = images

                image = np.concatenate(image, axis=0)
                image = np.array(image).astype(np.float32) / 255.0
                image = (image - 0.5) / 0.5
                image = image.transpose(0, 3, 1, 2)
                image = torch.from_numpy(image)

            elif isinstance(image[0], torch.Tensor):
                image = torch.stack(image, dim=0)

        image_batch_size = image.shape[0]

        if image_batch_size == 1:
            repeat_by = batch_size
        else:
            repeat_by = num_images_per_prompt

        image = image.repeat_interleave(repeat_by, dim=0)

        image = image.to(device=device, dtype=dtype)

        if do_classifier_free_guidance and not guess_mode:
            image = torch.cat([image] * 2)

        return image

    def prepare_ref_latents(self, refimage, batch_size, dtype, device, generator, do_classifier_free_guidance):
        refimage = refimage.to(device=device)
        if self.vae.dtype == torch.float16 and self.vae.config.force_upcast:
            self.upcast_vae()
            refimage = refimage.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
        if refimage.dtype != self.vae.dtype:
            refimage = refimage.to(dtype=self.vae.dtype)
        # encode the mask image into latents space so we can concatenate it to the latents
        if isinstance(generator, list):
            ref_image_latents = [
                self.vae.encode(refimage[i : i + 1]).latent_dist.sample(generator=generator[i])
                for i in range(batch_size)
            ]
            ref_image_latents = torch.cat(ref_image_latents, dim=0)
        else:
            ref_image_latents = self.vae.encode(refimage).latent_dist.sample(generator=generator)
        ref_image_latents = self.vae.config.scaling_factor * ref_image_latents

        # duplicate mask and ref_image_latents for each generation per prompt, using mps friendly method
        if ref_image_latents.shape[0] < batch_size:
            if not batch_size % ref_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 {ref_image_latents.shape[0]} images were passed."
                    " Make sure the number of images that you pass is divisible by the total requested batch size."
                )
            ref_image_latents = ref_image_latents.repeat(batch_size // ref_image_latents.shape[0], 1, 1, 1)

        ref_image_latents = torch.cat([ref_image_latents] * 2) if do_classifier_free_guidance else ref_image_latents

        # aligning device to prevent device errors when concating it with the latent model input
        ref_image_latents = ref_image_latents.to(device=device, dtype=dtype)
        return ref_image_latents

    @torch.no_grad()
    def __call__(
        self,
        prompt: Union[str, List[str]] = None,
        prompt_2: Optional[Union[str, List[str]]] = None,
        ref_image: Union[torch.FloatTensor, PIL.Image.Image] = 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,
        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.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,
        attention_auto_machine_weight: float = 1.0,
        gn_auto_machine_weight: float = 1.0,
        style_fidelity: float = 0.5,
        reference_attn: bool = True,
        reference_adain: bool = True,
    ):
        assert reference_attn or reference_adain, "`reference_attn` or `reference_adain` must be True."

        # 0. Default height and width to unet
        # height, width = self._default_height_width(height, width, ref_image)

        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,
            prompt_2,
            height,
            width,
            callback_steps,
            negative_prompt,
            negative_prompt_2,
            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=prompt,
            prompt_2=prompt_2,
            device=device,
            num_images_per_prompt=num_images_per_prompt,
            do_classifier_free_guidance=do_classifier_free_guidance,
            negative_prompt=negative_prompt,
            negative_prompt_2=negative_prompt_2,
            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. Preprocess reference image
        ref_image = self.prepare_image(
            image=ref_image,
            width=width,
            height=height,
            batch_size=batch_size * num_images_per_prompt,
            num_images_per_prompt=num_images_per_prompt,
            device=device,
            dtype=prompt_embeds.dtype,
        )

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

        timesteps = self.scheduler.timesteps

        # 6. Prepare latent variables
        num_channels_latents = self.unet.config.in_channels
        latents = self.prepare_latents(
            batch_size * num_images_per_prompt,
            num_channels_latents,
            height,
            width,
            prompt_embeds.dtype,
            device,
            generator,
            latents,
        )
        # 7. Prepare reference latent variables
        ref_image_latents = self.prepare_ref_latents(
            ref_image,
            batch_size * num_images_per_prompt,
            prompt_embeds.dtype,
            device,
            generator,
            do_classifier_free_guidance,
        )

        # 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. Modify self attebtion and group norm
        MODE = "write"
        uc_mask = (
            torch.Tensor([1] * batch_size * num_images_per_prompt + [0] * batch_size * num_images_per_prompt)
            .type_as(ref_image_latents)
            .bool()
        )

        def hacked_basic_transformer_inner_forward(
            self,
            hidden_states: torch.FloatTensor,
            attention_mask: Optional[torch.FloatTensor] = None,
            encoder_hidden_states: Optional[torch.FloatTensor] = None,
            encoder_attention_mask: Optional[torch.FloatTensor] = None,
            timestep: Optional[torch.LongTensor] = None,
            cross_attention_kwargs: Dict[str, Any] = None,
            class_labels: Optional[torch.LongTensor] = None,
        ):
            if self.use_ada_layer_norm:
                norm_hidden_states = self.norm1(hidden_states, timestep)
            elif self.use_ada_layer_norm_zero:
                norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.norm1(
                    hidden_states, timestep, class_labels, hidden_dtype=hidden_states.dtype
                )
            else:
                norm_hidden_states = self.norm1(hidden_states)

            # 1. Self-Attention
            cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {}
            if self.only_cross_attention:
                attn_output = self.attn1(
                    norm_hidden_states,
                    encoder_hidden_states=encoder_hidden_states if self.only_cross_attention else None,
                    attention_mask=attention_mask,
                    **cross_attention_kwargs,
                )
            else:
                if MODE == "write":
                    self.bank.append(norm_hidden_states.detach().clone())
                    attn_output = self.attn1(
                        norm_hidden_states,
                        encoder_hidden_states=encoder_hidden_states if self.only_cross_attention else None,
                        attention_mask=attention_mask,
                        **cross_attention_kwargs,
                    )
                if MODE == "read":
                    if attention_auto_machine_weight > self.attn_weight:
                        attn_output_uc = self.attn1(
                            norm_hidden_states,
                            encoder_hidden_states=torch.cat([norm_hidden_states] + self.bank, dim=1),
                            # attention_mask=attention_mask,
                            **cross_attention_kwargs,
                        )
                        attn_output_c = attn_output_uc.clone()
                        if do_classifier_free_guidance and style_fidelity > 0:
                            attn_output_c[uc_mask] = self.attn1(
                                norm_hidden_states[uc_mask],
                                encoder_hidden_states=norm_hidden_states[uc_mask],
                                **cross_attention_kwargs,
                            )
                        attn_output = style_fidelity * attn_output_c + (1.0 - style_fidelity) * attn_output_uc
                        self.bank.clear()
                    else:
                        attn_output = self.attn1(
                            norm_hidden_states,
                            encoder_hidden_states=encoder_hidden_states if self.only_cross_attention else None,
                            attention_mask=attention_mask,
                            **cross_attention_kwargs,
                        )
            if self.use_ada_layer_norm_zero:
                attn_output = gate_msa.unsqueeze(1) * attn_output
            hidden_states = attn_output + hidden_states

            if self.attn2 is not None:
                norm_hidden_states = (
                    self.norm2(hidden_states, timestep) if self.use_ada_layer_norm else self.norm2(hidden_states)
                )

                # 2. Cross-Attention
                attn_output = self.attn2(
                    norm_hidden_states,
                    encoder_hidden_states=encoder_hidden_states,
                    attention_mask=encoder_attention_mask,
                    **cross_attention_kwargs,
                )
                hidden_states = attn_output + hidden_states

            # 3. Feed-forward
            norm_hidden_states = self.norm3(hidden_states)

            if self.use_ada_layer_norm_zero:
                norm_hidden_states = norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None]

            ff_output = self.ff(norm_hidden_states)

            if self.use_ada_layer_norm_zero:
                ff_output = gate_mlp.unsqueeze(1) * ff_output

            hidden_states = ff_output + hidden_states

            return hidden_states

        def hacked_mid_forward(self, *args, **kwargs):
            eps = 1e-6
            x = self.original_forward(*args, **kwargs)
            if MODE == "write":
                if gn_auto_machine_weight >= self.gn_weight:
                    var, mean = torch.var_mean(x, dim=(2, 3), keepdim=True, correction=0)
                    self.mean_bank.append(mean)
                    self.var_bank.append(var)
            if MODE == "read":
                if len(self.mean_bank) > 0 and len(self.var_bank) > 0:
                    var, mean = torch.var_mean(x, dim=(2, 3), keepdim=True, correction=0)
                    std = torch.maximum(var, torch.zeros_like(var) + eps) ** 0.5
                    mean_acc = sum(self.mean_bank) / float(len(self.mean_bank))
                    var_acc = sum(self.var_bank) / float(len(self.var_bank))
                    std_acc = torch.maximum(var_acc, torch.zeros_like(var_acc) + eps) ** 0.5
                    x_uc = (((x - mean) / std) * std_acc) + mean_acc
                    x_c = x_uc.clone()
                    if do_classifier_free_guidance and style_fidelity > 0:
                        x_c[uc_mask] = x[uc_mask]
                    x = style_fidelity * x_c + (1.0 - style_fidelity) * x_uc
                self.mean_bank = []
                self.var_bank = []
            return x

        def hack_CrossAttnDownBlock2D_forward(
            self,
            hidden_states: torch.FloatTensor,
            temb: Optional[torch.FloatTensor] = None,
            encoder_hidden_states: Optional[torch.FloatTensor] = None,
            attention_mask: Optional[torch.FloatTensor] = None,
            cross_attention_kwargs: Optional[Dict[str, Any]] = None,
            encoder_attention_mask: Optional[torch.FloatTensor] = None,
        ):
            eps = 1e-6

            # TODO(Patrick, William) - attention mask is not used
            output_states = ()

            for i, (resnet, attn) in enumerate(zip(self.resnets, self.attentions)):
                hidden_states = resnet(hidden_states, temb)
                hidden_states = attn(
                    hidden_states,
                    encoder_hidden_states=encoder_hidden_states,
                    cross_attention_kwargs=cross_attention_kwargs,
                    attention_mask=attention_mask,
                    encoder_attention_mask=encoder_attention_mask,
                    return_dict=False,
                )[0]
                if MODE == "write":
                    if gn_auto_machine_weight >= self.gn_weight:
                        var, mean = torch.var_mean(hidden_states, dim=(2, 3), keepdim=True, correction=0)
                        self.mean_bank.append([mean])
                        self.var_bank.append([var])
                if MODE == "read":
                    if len(self.mean_bank) > 0 and len(self.var_bank) > 0:
                        var, mean = torch.var_mean(hidden_states, dim=(2, 3), keepdim=True, correction=0)
                        std = torch.maximum(var, torch.zeros_like(var) + eps) ** 0.5
                        mean_acc = sum(self.mean_bank[i]) / float(len(self.mean_bank[i]))
                        var_acc = sum(self.var_bank[i]) / float(len(self.var_bank[i]))
                        std_acc = torch.maximum(var_acc, torch.zeros_like(var_acc) + eps) ** 0.5
                        hidden_states_uc = (((hidden_states - mean) / std) * std_acc) + mean_acc
                        hidden_states_c = hidden_states_uc.clone()
                        if do_classifier_free_guidance and style_fidelity > 0:
                            hidden_states_c[uc_mask] = hidden_states[uc_mask]
                        hidden_states = style_fidelity * hidden_states_c + (1.0 - style_fidelity) * hidden_states_uc

                output_states = output_states + (hidden_states,)

            if MODE == "read":
                self.mean_bank = []
                self.var_bank = []

            if self.downsamplers is not None:
                for downsampler in self.downsamplers:
                    hidden_states = downsampler(hidden_states)

                output_states = output_states + (hidden_states,)

            return hidden_states, output_states

        def hacked_DownBlock2D_forward(self, hidden_states, temb=None, *args, **kwargs):
            eps = 1e-6

            output_states = ()

            for i, resnet in enumerate(self.resnets):
                hidden_states = resnet(hidden_states, temb)

                if MODE == "write":
                    if gn_auto_machine_weight >= self.gn_weight:
                        var, mean = torch.var_mean(hidden_states, dim=(2, 3), keepdim=True, correction=0)
                        self.mean_bank.append([mean])
                        self.var_bank.append([var])
                if MODE == "read":
                    if len(self.mean_bank) > 0 and len(self.var_bank) > 0:
                        var, mean = torch.var_mean(hidden_states, dim=(2, 3), keepdim=True, correction=0)
                        std = torch.maximum(var, torch.zeros_like(var) + eps) ** 0.5
                        mean_acc = sum(self.mean_bank[i]) / float(len(self.mean_bank[i]))
                        var_acc = sum(self.var_bank[i]) / float(len(self.var_bank[i]))
                        std_acc = torch.maximum(var_acc, torch.zeros_like(var_acc) + eps) ** 0.5
                        hidden_states_uc = (((hidden_states - mean) / std) * std_acc) + mean_acc
                        hidden_states_c = hidden_states_uc.clone()
                        if do_classifier_free_guidance and style_fidelity > 0:
                            hidden_states_c[uc_mask] = hidden_states[uc_mask]
                        hidden_states = style_fidelity * hidden_states_c + (1.0 - style_fidelity) * hidden_states_uc

                output_states = output_states + (hidden_states,)

            if MODE == "read":
                self.mean_bank = []
                self.var_bank = []

            if self.downsamplers is not None:
                for downsampler in self.downsamplers:
                    hidden_states = downsampler(hidden_states)

                output_states = output_states + (hidden_states,)

            return hidden_states, output_states

        def hacked_CrossAttnUpBlock2D_forward(
            self,
            hidden_states: torch.FloatTensor,
            res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
            temb: Optional[torch.FloatTensor] = None,
            encoder_hidden_states: Optional[torch.FloatTensor] = None,
            cross_attention_kwargs: Optional[Dict[str, Any]] = None,
            upsample_size: Optional[int] = None,
            attention_mask: Optional[torch.FloatTensor] = None,
            encoder_attention_mask: Optional[torch.FloatTensor] = None,
        ):
            eps = 1e-6
            # TODO(Patrick, William) - attention mask is not used
            for i, (resnet, attn) in enumerate(zip(self.resnets, self.attentions)):
                # pop res hidden states
                res_hidden_states = res_hidden_states_tuple[-1]
                res_hidden_states_tuple = res_hidden_states_tuple[:-1]
                hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
                hidden_states = resnet(hidden_states, temb)
                hidden_states = attn(
                    hidden_states,
                    encoder_hidden_states=encoder_hidden_states,
                    cross_attention_kwargs=cross_attention_kwargs,
                    attention_mask=attention_mask,
                    encoder_attention_mask=encoder_attention_mask,
                    return_dict=False,
                )[0]

                if MODE == "write":
                    if gn_auto_machine_weight >= self.gn_weight:
                        var, mean = torch.var_mean(hidden_states, dim=(2, 3), keepdim=True, correction=0)
                        self.mean_bank.append([mean])
                        self.var_bank.append([var])
                if MODE == "read":
                    if len(self.mean_bank) > 0 and len(self.var_bank) > 0:
                        var, mean = torch.var_mean(hidden_states, dim=(2, 3), keepdim=True, correction=0)
                        std = torch.maximum(var, torch.zeros_like(var) + eps) ** 0.5
                        mean_acc = sum(self.mean_bank[i]) / float(len(self.mean_bank[i]))
                        var_acc = sum(self.var_bank[i]) / float(len(self.var_bank[i]))
                        std_acc = torch.maximum(var_acc, torch.zeros_like(var_acc) + eps) ** 0.5
                        hidden_states_uc = (((hidden_states - mean) / std) * std_acc) + mean_acc
                        hidden_states_c = hidden_states_uc.clone()
                        if do_classifier_free_guidance and style_fidelity > 0:
                            hidden_states_c[uc_mask] = hidden_states[uc_mask]
                        hidden_states = style_fidelity * hidden_states_c + (1.0 - style_fidelity) * hidden_states_uc

            if MODE == "read":
                self.mean_bank = []
                self.var_bank = []

            if self.upsamplers is not None:
                for upsampler in self.upsamplers:
                    hidden_states = upsampler(hidden_states, upsample_size)

            return hidden_states

        def hacked_UpBlock2D_forward(
            self, hidden_states, res_hidden_states_tuple, temb=None, upsample_size=None, **kwargs
        ):
            eps = 1e-6
            for i, resnet in enumerate(self.resnets):
                # pop res hidden states
                res_hidden_states = res_hidden_states_tuple[-1]
                res_hidden_states_tuple = res_hidden_states_tuple[:-1]
                hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
                hidden_states = resnet(hidden_states, temb)

                if MODE == "write":
                    if gn_auto_machine_weight >= self.gn_weight:
                        var, mean = torch.var_mean(hidden_states, dim=(2, 3), keepdim=True, correction=0)
                        self.mean_bank.append([mean])
                        self.var_bank.append([var])
                if MODE == "read":
                    if len(self.mean_bank) > 0 and len(self.var_bank) > 0:
                        var, mean = torch.var_mean(hidden_states, dim=(2, 3), keepdim=True, correction=0)
                        std = torch.maximum(var, torch.zeros_like(var) + eps) ** 0.5
                        mean_acc = sum(self.mean_bank[i]) / float(len(self.mean_bank[i]))
                        var_acc = sum(self.var_bank[i]) / float(len(self.var_bank[i]))
                        std_acc = torch.maximum(var_acc, torch.zeros_like(var_acc) + eps) ** 0.5
                        hidden_states_uc = (((hidden_states - mean) / std) * std_acc) + mean_acc
                        hidden_states_c = hidden_states_uc.clone()
                        if do_classifier_free_guidance and style_fidelity > 0:
                            hidden_states_c[uc_mask] = hidden_states[uc_mask]
                        hidden_states = style_fidelity * hidden_states_c + (1.0 - style_fidelity) * hidden_states_uc

            if MODE == "read":
                self.mean_bank = []
                self.var_bank = []

            if self.upsamplers is not None:
                for upsampler in self.upsamplers:
                    hidden_states = upsampler(hidden_states, upsample_size)

            return hidden_states

        if reference_attn:
            attn_modules = [module for module in torch_dfs(self.unet) if isinstance(module, BasicTransformerBlock)]
            attn_modules = sorted(attn_modules, key=lambda x: -x.norm1.normalized_shape[0])

            for i, module in enumerate(attn_modules):
                module._original_inner_forward = module.forward
                module.forward = hacked_basic_transformer_inner_forward.__get__(module, BasicTransformerBlock)
                module.bank = []
                module.attn_weight = float(i) / float(len(attn_modules))

        if reference_adain:
            gn_modules = [self.unet.mid_block]
            self.unet.mid_block.gn_weight = 0

            down_blocks = self.unet.down_blocks
            for w, module in enumerate(down_blocks):
                module.gn_weight = 1.0 - float(w) / float(len(down_blocks))
                gn_modules.append(module)

            up_blocks = self.unet.up_blocks
            for w, module in enumerate(up_blocks):
                module.gn_weight = float(w) / float(len(up_blocks))
                gn_modules.append(module)

            for i, module in enumerate(gn_modules):
                if getattr(module, "original_forward", None) is None:
                    module.original_forward = module.forward
                if i == 0:
                    # mid_block
                    module.forward = hacked_mid_forward.__get__(module, torch.nn.Module)
                elif isinstance(module, CrossAttnDownBlock2D):
                    module.forward = hack_CrossAttnDownBlock2D_forward.__get__(module, CrossAttnDownBlock2D)
                elif isinstance(module, DownBlock2D):
                    module.forward = hacked_DownBlock2D_forward.__get__(module, DownBlock2D)
                elif isinstance(module, CrossAttnUpBlock2D):
                    module.forward = hacked_CrossAttnUpBlock2D_forward.__get__(module, CrossAttnUpBlock2D)
                elif isinstance(module, UpBlock2D):
                    module.forward = hacked_UpBlock2D_forward.__get__(module, UpBlock2D)
                module.mean_bank = []
                module.var_bank = []
                module.gn_weight *= 2

        # 10. Prepare added time ids & embeddings
        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 = self._get_add_time_ids(
            original_size,
            crops_coords_top_left,
            target_size,
            dtype=prompt_embeds.dtype,
            text_encoder_projection_dim=text_encoder_projection_dim,
        )

        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)

        # 11. Denoising loop
        num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)

        # 10.1 Apply denoising_end
        if denoising_end is not None and isinstance(denoising_end, float) and denoising_end > 0 and denoising_end < 1:
            discrete_timestep_cutoff = int(
                round(
                    self.scheduler.config.num_train_timesteps
                    - (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]

        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)

                added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids}

                # ref only part
                noise = randn_tensor(
                    ref_image_latents.shape, generator=generator, device=device, dtype=ref_image_latents.dtype
                )
                ref_xt = self.scheduler.add_noise(
                    ref_image_latents,
                    noise,
                    t.reshape(
                        1,
                    ),
                )
                ref_xt = self.scheduler.scale_model_input(ref_xt, t)

                MODE = "write"

                self.unet(
                    ref_xt,
                    t,
                    encoder_hidden_states=prompt_embeds,
                    cross_attention_kwargs=cross_attention_kwargs,
                    added_cond_kwargs=added_cond_kwargs,
                    return_dict=False,
                )

                # predict the noise residual
                MODE = "read"
                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 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:
                        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()
                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)
        else:
            image = latents
            return StableDiffusionXLPipelineOutput(images=image)

        # 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)

        # 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)