diffusers/pipelines/controlnet/pipeline_controlnet_inpaint_sd_xl.py

# Copyright 2023 Harutatsu Akiyama, Jinbin Bai, and 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
import torch.nn.functional as F
from transformers import CLIPTextModel, CLIPTextModelWithProjection, CLIPTokenizer

from ...image_processor import PipelineImageInput, VaeImageProcessor
from ...loaders import FromSingleFileMixin, StableDiffusionXLLoraLoaderMixin, TextualInversionLoaderMixin
from ...models import AutoencoderKL, ControlNetModel, UNet2DConditionModel
from ...models.attention_processor import (
    AttnProcessor2_0,
    LoRAAttnProcessor2_0,
    LoRAXFormersAttnProcessor,
    XFormersAttnProcessor,
)
from ...models.lora import adjust_lora_scale_text_encoder
from ...schedulers import KarrasDiffusionSchedulers
from ...utils import (
    USE_PEFT_BACKEND,
    deprecate,
    is_invisible_watermark_available,
    logging,
    replace_example_docstring,
    scale_lora_layers,
    unscale_lora_layers,
)
from ...utils.torch_utils import is_compiled_module, randn_tensor
from ..pipeline_utils import DiffusionPipeline
from ..stable_diffusion_xl.pipeline_output import StableDiffusionXLPipelineOutput
from .multicontrolnet import MultiControlNetModel


if is_invisible_watermark_available():
    from diffusers.pipelines.stable_diffusion_xl.watermark import StableDiffusionXLWatermarker


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


# 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")


EXAMPLE_DOC_STRING = """
    Examples:
        ```py
        >>> # !pip install transformers accelerate
        >>> from diffusers import StableDiffusionXLControlNetInpaintPipeline, ControlNetModel, DDIMScheduler
        >>> from diffusers.utils import load_image
        >>> import numpy as np
        >>> import torch

        >>> init_image = load_image(
        ...     "https://huggingface.co/datasets/diffusers/test-arrays/resolve/main/stable_diffusion_inpaint/boy.png"
        ... )
        >>> init_image = init_image.resize((1024, 1024))

        >>> generator = torch.Generator(device="cpu").manual_seed(1)

        >>> mask_image = load_image(
        ...     "https://huggingface.co/datasets/diffusers/test-arrays/resolve/main/stable_diffusion_inpaint/boy_mask.png"
        ... )
        >>> mask_image = mask_image.resize((1024, 1024))


        >>> def make_canny_condition(image):
        ...     image = np.array(image)
        ...     image = cv2.Canny(image, 100, 200)
        ...     image = image[:, :, None]
        ...     image = np.concatenate([image, image, image], axis=2)
        ...     image = Image.fromarray(image)
        ...     return image


        >>> control_image = make_canny_condition(init_image)

        >>> controlnet = ControlNetModel.from_pretrained(
        ...     "diffusers/controlnet-canny-sdxl-1.0", torch_dtype=torch.float16
        ... )
        >>> pipe = StableDiffusionXLControlNetInpaintPipeline.from_pretrained(
        ...     "stabilityai/stable-diffusion-xl-base-1.0", controlnet=controlnet, torch_dtype=torch.float16
        ... )

        >>> pipe.enable_model_cpu_offload()

        >>> # generate image
        >>> image = pipe(
        ...     "a handsome man with ray-ban sunglasses",
        ...     num_inference_steps=20,
        ...     generator=generator,
        ...     eta=1.0,
        ...     image=init_image,
        ...     mask_image=mask_image,
        ...     control_image=control_image,
        ... ).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 StableDiffusionXLControlNetInpaintPipeline(
    DiffusionPipeline, StableDiffusionXLLoraLoaderMixin, FromSingleFileMixin
):
    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.StableDiffusionXLLoraLoaderMixin.load_lora_weights`] for loading LoRA weights
        - [`~loaders.StableDiffusionXLLoraLoaderMixin.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 ([`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`].
    """

    model_cpu_offload_seq = "text_encoder->text_encoder_2->unet->vae"
    _optional_components = ["tokenizer", "tokenizer_2", "text_encoder", "text_encoder_2"]
    _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"]

    def __init__(
        self,
        vae: AutoencoderKL,
        text_encoder: CLIPTextModel,
        text_encoder_2: CLIPTextModelWithProjection,
        tokenizer: CLIPTokenizer,
        tokenizer_2: CLIPTokenizer,
        unet: UNet2DConditionModel,
        controlnet: ControlNetModel,
        scheduler: KarrasDiffusionSchedulers,
        requires_aesthetics_score: bool = False,
        force_zeros_for_empty_prompt: bool = True,
        add_watermarker: Optional[bool] = None,
    ):
        super().__init__()

        if isinstance(controlnet, (list, tuple)):
            controlnet = MultiControlNetModel(controlnet)

        self.register_modules(
            vae=vae,
            text_encoder=text_encoder,
            text_encoder_2=text_encoder_2,
            tokenizer=tokenizer,
            tokenizer_2=tokenizer_2,
            unet=unet,
            controlnet=controlnet,
            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
        )
        self.control_image_processor = VaeImageProcessor(
            vae_scale_factor=self.vae_scale_factor, do_convert_rgb=True, do_normalize=False
        )

        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.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 enabled, 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 for saving a large amount of memory and to allow
        processing 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 enabled, this method will go back to
        computing decoding in one step.
        """
        self.vae.disable_tiling()

    # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.encode_prompt
    def encode_prompt(
        self,
        prompt: str,
        prompt_2: Optional[str] = None,
        device: Optional[torch.device] = None,
        num_images_per_prompt: int = 1,
        do_classifier_free_guidance: bool = True,
        negative_prompt: Optional[str] = None,
        negative_prompt_2: Optional[str] = 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,
        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
            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
            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`).
            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.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.
            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.
        """
        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, StableDiffusionXLLoraLoaderMixin):
            self._lora_scale = lora_scale

            # dynamically adjust the LoRA scale
            if self.text_encoder is not None:
                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 self.text_encoder_2 is not None:
                if not USE_PEFT_BACKEND:
                    adjust_lora_scale_text_encoder(self.text_encoder_2, lora_scale)
                else:
                    scale_lora_layers(self.text_encoder_2, lora_scale)

        prompt = [prompt] if isinstance(prompt, str) else prompt

        if prompt is not None:
            batch_size = len(prompt)
        else:
            batch_size = prompt_embeds.shape[0]

        # Define tokenizers and text encoders
        tokenizers = [self.tokenizer, self.tokenizer_2] if self.tokenizer is not None else [self.tokenizer_2]
        text_encoders = (
            [self.text_encoder, self.text_encoder_2] if self.text_encoder is not None else [self.text_encoder_2]
        )

        if prompt_embeds is None:
            prompt_2 = prompt_2 or prompt
            prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2

            # textual inversion: procecss multi-vector tokens if necessary
            prompt_embeds_list = []
            prompts = [prompt, prompt_2]
            for prompt, tokenizer, text_encoder in zip(prompts, tokenizers, text_encoders):
                if isinstance(self, TextualInversionLoaderMixin):
                    prompt = self.maybe_convert_prompt(prompt, tokenizer)

                text_inputs = tokenizer(
                    prompt,
                    padding="max_length",
                    max_length=tokenizer.model_max_length,
                    truncation=True,
                    return_tensors="pt",
                )

                text_input_ids = text_inputs.input_ids
                untruncated_ids = 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 = tokenizer.batch_decode(untruncated_ids[:, 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" {tokenizer.model_max_length} tokens: {removed_text}"
                    )

                prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True)

                # We are only ALWAYS interested in the pooled output of the final text encoder
                pooled_prompt_embeds = prompt_embeds[0]
                if clip_skip is None:
                    prompt_embeds = prompt_embeds.hidden_states[-2]
                else:
                    # "2" because SDXL always indexes from the penultimate layer.
                    prompt_embeds = prompt_embeds.hidden_states[-(clip_skip + 2)]

                prompt_embeds_list.append(prompt_embeds)

            prompt_embeds = torch.concat(prompt_embeds_list, dim=-1)

        # 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 ""
            negative_prompt_2 = negative_prompt_2 or negative_prompt

            # normalize str to list
            negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt
            negative_prompt_2 = (
                batch_size * [negative_prompt_2] if isinstance(negative_prompt_2, str) else negative_prompt_2
            )

            uncond_tokens: List[str]
            if 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 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_2]

            negative_prompt_embeds_list = []
            for negative_prompt, tokenizer, text_encoder in zip(uncond_tokens, tokenizers, text_encoders):
                if isinstance(self, TextualInversionLoaderMixin):
                    negative_prompt = self.maybe_convert_prompt(negative_prompt, tokenizer)

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

                negative_prompt_embeds = text_encoder(
                    uncond_input.input_ids.to(device),
                    output_hidden_states=True,
                )
                # We are only ALWAYS interested in the pooled output of the final text encoder
                negative_pooled_prompt_embeds = negative_prompt_embeds[0]
                negative_prompt_embeds = negative_prompt_embeds.hidden_states[-2]

                negative_prompt_embeds_list.append(negative_prompt_embeds)

            negative_prompt_embeds = torch.concat(negative_prompt_embeds_list, dim=-1)

        if self.text_encoder_2 is not None:
            prompt_embeds = prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device)
        else:
            prompt_embeds = prompt_embeds.to(dtype=self.unet.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)

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

            if self.text_encoder_2 is not None:
                negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device)
            else:
                negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.unet.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)

        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
            )

        if self.text_encoder is not None:
            if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND:
                # Retrieve the original scale by scaling back the LoRA layers
                unscale_lora_layers(self.text_encoder, lora_scale)

        if self.text_encoder_2 is not None:
            if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND:
                # Retrieve the original scale by scaling back the LoRA layers
                unscale_lora_layers(self.text_encoder_2, lora_scale)

        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_image(self, image, prompt, prompt_embeds):
        image_is_pil = isinstance(image, PIL.Image.Image)
        image_is_tensor = isinstance(image, torch.Tensor)
        image_is_np = isinstance(image, np.ndarray)
        image_is_pil_list = isinstance(image, list) and isinstance(image[0], PIL.Image.Image)
        image_is_tensor_list = isinstance(image, list) and isinstance(image[0], torch.Tensor)
        image_is_np_list = isinstance(image, list) and isinstance(image[0], np.ndarray)

        if (
            not image_is_pil
            and not image_is_tensor
            and not image_is_np
            and not image_is_pil_list
            and not image_is_tensor_list
            and not image_is_np_list
        ):
            raise TypeError(
                f"image must be passed and be one of PIL image, numpy array, torch tensor, list of PIL images, list of numpy arrays or list of torch tensors, but is {type(image)}"
            )

        if image_is_pil:
            image_batch_size = 1
        else:
            image_batch_size = len(image)

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

        if image_batch_size != 1 and image_batch_size != prompt_batch_size:
            raise ValueError(
                f"If image batch size is not 1, image batch size must be same as prompt batch size. image batch size: {image_batch_size}, prompt batch size: {prompt_batch_size}"
            )

    def check_inputs(
        self,
        prompt,
        prompt_2,
        image,
        mask_image,
        strength,
        num_inference_steps,
        callback_steps,
        output_type,
        negative_prompt=None,
        negative_prompt_2=None,
        prompt_embeds=None,
        negative_prompt_embeds=None,
        pooled_prompt_embeds=None,
        negative_pooled_prompt_embeds=None,
        controlnet_conditioning_scale=1.0,
        control_guidance_start=0.0,
        control_guidance_end=1.0,
        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 num_inference_steps is None:
            raise ValueError("`num_inference_steps` cannot be None.")
        elif not isinstance(num_inference_steps, int) or num_inference_steps <= 0:
            raise ValueError(
                f"`num_inference_steps` has to be a positive integer but is {num_inference_steps} of type"
                f" {type(num_inference_steps)}."
            )

        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 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`."
            )

        # `prompt` needs more sophisticated handling when there are multiple
        # conditionings.
        if isinstance(self.controlnet, MultiControlNetModel):
            if isinstance(prompt, list):
                logger.warning(
                    f"You have {len(self.controlnet.nets)} ControlNets and you have passed {len(prompt)}"
                    " prompts. The conditionings will be fixed across the prompts."
                )

        # Check `image`
        is_compiled = hasattr(F, "scaled_dot_product_attention") and isinstance(
            self.controlnet, torch._dynamo.eval_frame.OptimizedModule
        )
        if (
            isinstance(self.controlnet, ControlNetModel)
            or is_compiled
            and isinstance(self.controlnet._orig_mod, ControlNetModel)
        ):
            self.check_image(image, prompt, prompt_embeds)
        elif (
            isinstance(self.controlnet, MultiControlNetModel)
            or is_compiled
            and isinstance(self.controlnet._orig_mod, MultiControlNetModel)
        ):
            if not isinstance(image, list):
                raise TypeError("For multiple controlnets: `image` must be type `list`")

            # When `image` is a nested list:
            # (e.g. [[canny_image_1, pose_image_1], [canny_image_2, pose_image_2]])
            elif any(isinstance(i, list) for i in image):
                raise ValueError("A single batch of multiple conditionings are supported at the moment.")
            elif len(image) != len(self.controlnet.nets):
                raise ValueError(
                    f"For multiple controlnets: `image` must have the same length as the number of controlnets, but got {len(image)} images and {len(self.controlnet.nets)} ControlNets."
                )

            for image_ in image:
                self.check_image(image_, prompt, prompt_embeds)
        else:
            assert False

        # Check `controlnet_conditioning_scale`
        if (
            isinstance(self.controlnet, ControlNetModel)
            or is_compiled
            and isinstance(self.controlnet._orig_mod, ControlNetModel)
        ):
            if not isinstance(controlnet_conditioning_scale, float):
                raise TypeError("For single controlnet: `controlnet_conditioning_scale` must be type `float`.")
        elif (
            isinstance(self.controlnet, MultiControlNetModel)
            or is_compiled
            and isinstance(self.controlnet._orig_mod, MultiControlNetModel)
        ):
            if isinstance(controlnet_conditioning_scale, list):
                if any(isinstance(i, list) for i in controlnet_conditioning_scale):
                    raise ValueError("A single batch of multiple conditionings are supported at the moment.")
            elif isinstance(controlnet_conditioning_scale, list) and len(controlnet_conditioning_scale) != len(
                self.controlnet.nets
            ):
                raise ValueError(
                    "For multiple controlnets: When `controlnet_conditioning_scale` is specified as `list`, it must have"
                    " the same length as the number of controlnets"
                )
        else:
            assert False

        if not isinstance(control_guidance_start, (tuple, list)):
            control_guidance_start = [control_guidance_start]

        if not isinstance(control_guidance_end, (tuple, list)):
            control_guidance_end = [control_guidance_end]

        if len(control_guidance_start) != len(control_guidance_end):
            raise ValueError(
                f"`control_guidance_start` has {len(control_guidance_start)} elements, but `control_guidance_end` has {len(control_guidance_end)} elements. Make sure to provide the same number of elements to each list."
            )

        if isinstance(self.controlnet, MultiControlNetModel):
            if len(control_guidance_start) != len(self.controlnet.nets):
                raise ValueError(
                    f"`control_guidance_start`: {control_guidance_start} has {len(control_guidance_start)} elements but there are {len(self.controlnet.nets)} controlnets available. Make sure to provide {len(self.controlnet.nets)}."
                )

        for start, end in zip(control_guidance_start, control_guidance_end):
            if start >= end:
                raise ValueError(
                    f"control guidance start: {start} cannot be larger or equal to control guidance end: {end}."
                )
            if start < 0.0:
                raise ValueError(f"control guidance start: {start} can't be smaller than 0.")
            if end > 1.0:
                raise ValueError(f"control guidance end: {end} can't be larger than 1.0.")

    def prepare_control_image(
        self,
        image,
        width,
        height,
        batch_size,
        num_images_per_prompt,
        device,
        dtype,
        crops_coords,
        resize_mode,
        do_classifier_free_guidance=False,
        guess_mode=False,
    ):
        image = self.control_image_processor.preprocess(
            image, height=height, width=width, crops_coords=crops_coords, resize_mode=resize_mode
        ).to(dtype=torch.float32)
        image_batch_size = image.shape[0]

        if image_batch_size == 1:
            repeat_by = batch_size
        else:
            # image batch size is the same as prompt batch size
            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_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, 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 (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 return_image_latents or (latents is None and not is_strength_max):
            image = image.to(device=device, dtype=dtype)

            if image.shape[1] == 4:
                image_latents = image
            else:
                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

        masked_image_latents = None
        if masked_image is not 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 devirative) 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

    def _get_add_time_ids(
        self,
        original_size,
        crops_coords_top_left,
        target_size,
        aesthetic_score,
        negative_aesthetic_score,
        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(original_size + 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(original_size + crops_coords_top_left + target_size)

        passed_add_embed_dim = (
            self.unet.config.addition_time_embed_dim * len(add_time_ids) + text_encoder_projection_dim
        )
        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.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_freeu
    def enable_freeu(self, s1: float, s2: float, b1: float, b2: float):
        r"""Enables the FreeU mechanism as in https://arxiv.org/abs/2309.11497.

        The suffixes after the scaling factors represent the stages where they are being applied.

        Please refer to the [official repository](https://github.com/ChenyangSi/FreeU) for combinations of the values
        that are known to work well for different pipelines such as Stable Diffusion v1, v2, and Stable Diffusion XL.

        Args:
            s1 (`float`):
                Scaling factor for stage 1 to attenuate the contributions of the skip features. This is done to
                mitigate "oversmoothing effect" in the enhanced denoising process.
            s2 (`float`):
                Scaling factor for stage 2 to attenuate the contributions of the skip features. This is done to
                mitigate "oversmoothing effect" in the enhanced denoising process.
            b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features.
            b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features.
        """
        if not hasattr(self, "unet"):
            raise ValueError("The pipeline must have `unet` for using FreeU.")
        self.unet.enable_freeu(s1=s1, s2=s2, b1=b1, b2=b2)

    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_freeu
    def disable_freeu(self):
        """Disables the FreeU mechanism if enabled."""
        self.unet.disable_freeu()

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

    @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

    @property
    def cross_attention_kwargs(self):
        return self._cross_attention_kwargs

    @property
    def num_timesteps(self):
        return self._num_timesteps

    @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,
        control_image: Union[
            PipelineImageInput,
            List[PipelineImageInput],
        ] = None,
        height: Optional[int] = None,
        width: Optional[int] = None,
        padding_mask_crop: Optional[int] = None,
        strength: float = 0.9999,
        num_inference_steps: int = 50,
        denoising_start: Optional[float] = None,
        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,
        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
        controlnet_conditioning_scale: Union[float, List[float]] = 1.0,
        guess_mode: bool = False,
        control_guidance_start: Union[float, List[float]] = 0.0,
        control_guidance_end: Union[float, List[float]] = 1.0,
        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,
        aesthetic_score: float = 6.0,
        negative_aesthetic_score: float = 2.5,
        clip_skip: Optional[int] = None,
        callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = 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.
            width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
                The width in pixels of the generated image.
            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 inreleant for inpainging, 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.
            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.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.
            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.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`.
            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 `(width, height)` 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 `(width, height)`. 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).
            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`, *optional*):
                A function that calls at the end of each denoising steps during the inference. The function is called
                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 pipeine 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 using `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 using `callback_on_step_end`",
            )

        controlnet = self.controlnet._orig_mod if is_compiled_module(self.controlnet) else self.controlnet

        # align format for control guidance
        if not isinstance(control_guidance_start, list) and isinstance(control_guidance_end, list):
            control_guidance_start = len(control_guidance_end) * [control_guidance_start]
        elif not isinstance(control_guidance_end, list) and isinstance(control_guidance_start, list):
            control_guidance_end = len(control_guidance_start) * [control_guidance_end]
        elif not isinstance(control_guidance_start, list) and not isinstance(control_guidance_end, list):
            mult = len(controlnet.nets) if isinstance(controlnet, MultiControlNetModel) else 1
            control_guidance_start, control_guidance_end = (
                mult * [control_guidance_start],
                mult * [control_guidance_end],
            )

        # # 0.0 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

        # 0.1 align format for control guidance
        if not isinstance(control_guidance_start, list) and isinstance(control_guidance_end, list):
            control_guidance_start = len(control_guidance_end) * [control_guidance_start]
        elif not isinstance(control_guidance_end, list) and isinstance(control_guidance_start, list):
            control_guidance_end = len(control_guidance_start) * [control_guidance_end]
        elif not isinstance(control_guidance_start, list) and not isinstance(control_guidance_end, list):
            mult = len(controlnet.nets) if isinstance(controlnet, MultiControlNetModel) else 1
            control_guidance_start, control_guidance_end = (
                mult * [control_guidance_start],
                mult * [control_guidance_end],
            )

        # 1. Check inputs
        self.check_inputs(
            prompt,
            prompt_2,
            control_image,
            mask_image,
            strength,
            num_inference_steps,
            callback_steps,
            output_type,
            negative_prompt,
            negative_prompt_2,
            prompt_embeds,
            negative_prompt_embeds,
            pooled_prompt_embeds,
            negative_pooled_prompt_embeds,
            controlnet_conditioning_scale,
            control_guidance_start,
            control_guidance_end,
            callback_on_step_end_tensor_inputs,
            padding_mask_crop,
        )

        self._guidance_scale = guidance_scale
        self._clip_skip = clip_skip
        self._cross_attention_kwargs = cross_attention_kwargs

        # 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

        if isinstance(controlnet, MultiControlNetModel) and isinstance(controlnet_conditioning_scale, float):
            controlnet_conditioning_scale = [controlnet_conditioning_scale] * len(controlnet.nets)

        # 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,
            prompt_2=prompt_2,
            device=device,
            num_images_per_prompt=num_images_per_prompt,
            do_classifier_free_guidance=self.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,
            clip_skip=self.clip_skip,
        )

        # 4. set timesteps
        def denoising_value_valid(dnv):
            return isinstance(denoising_end, float) and 0 < dnv < 1

        self.scheduler.set_timesteps(num_inference_steps, device=device)
        timesteps, num_inference_steps = self.get_timesteps(
            num_inference_steps, strength, device, denoising_start=denoising_start if denoising_value_valid 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
        self._num_timesteps = len(timesteps)

        # 5. Preprocess mask and image - resizes image and mask w.r.t height and width
        # 5.1 Prepare init image
        if padding_mask_crop is not None:
            height, width = self.image_processor.get_default_height_width(image, height, width)
            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)

        # 5.2 Prepare control images
        if isinstance(controlnet, ControlNetModel):
            control_image = self.prepare_control_image(
                image=control_image,
                width=width,
                height=height,
                batch_size=batch_size * num_images_per_prompt,
                num_images_per_prompt=num_images_per_prompt,
                device=device,
                dtype=controlnet.dtype,
                crops_coords=crops_coords,
                resize_mode=resize_mode,
                do_classifier_free_guidance=self.do_classifier_free_guidance,
                guess_mode=guess_mode,
            )
        elif isinstance(controlnet, MultiControlNetModel):
            control_images = []

            for control_image_ in control_image:
                control_image_ = self.prepare_control_image(
                    image=control_image_,
                    width=width,
                    height=height,
                    batch_size=batch_size * num_images_per_prompt,
                    num_images_per_prompt=num_images_per_prompt,
                    device=device,
                    dtype=controlnet.dtype,
                    crops_coords=crops_coords,
                    resize_mode=resize_mode,
                    do_classifier_free_guidance=self.do_classifier_free_guidance,
                    guess_mode=guess_mode,
                )

                control_images.append(control_image_)

            control_image = control_images
        else:
            raise ValueError(f"{controlnet.__class__} is not supported.")

        # 5.3 Prepare mask
        mask = self.mask_processor.preprocess(
            mask_image, height=height, width=width, resize_mode=resize_mode, crops_coords=crops_coords
        )

        masked_image = init_image * (mask < 0.5)
        _, _, height, width = init_image.shape

        # 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 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)

        # 8.2 Create tensor stating which controlnets to keep
        controlnet_keep = []
        for i in range(len(timesteps)):
            keeps = [
                1.0 - float(i / len(timesteps) < s or (i + 1) / len(timesteps) > e)
                for s, e in zip(control_guidance_start, control_guidance_end)
            ]
            if isinstance(self.controlnet, MultiControlNetModel):
                controlnet_keep.append(keeps)
            else:
                controlnet_keep.append(keeps[0])

        # 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
        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,
            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)

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

        if (
            denoising_end is not None
            and denoising_start is not None
            and denoising_value_valid(denoising_end)
            and denoising_value_valid(denoising_start)
            and denoising_start >= denoising_end
        ):
            raise ValueError(
                f"`denoising_start`: {denoising_start} cannot be larger than or equal to `denoising_end`: "
                + f" {denoising_end} when using type float."
            )
        elif denoising_end is not None and denoising_value_valid(denoising_end):
            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 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)

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

                # controlnet(s) inference
                if guess_mode and self.do_classifier_free_guidance:
                    # Infer ControlNet only for the conditional batch.
                    control_model_input = latents
                    control_model_input = self.scheduler.scale_model_input(control_model_input, t)
                    controlnet_prompt_embeds = prompt_embeds.chunk(2)[1]
                    controlnet_added_cond_kwargs = {
                        "text_embeds": add_text_embeds.chunk(2)[1],
                        "time_ids": add_time_ids.chunk(2)[1],
                    }
                else:
                    control_model_input = latent_model_input
                    controlnet_prompt_embeds = prompt_embeds
                    controlnet_added_cond_kwargs = added_cond_kwargs

                if isinstance(controlnet_keep[i], list):
                    cond_scale = [c * s for c, s in zip(controlnet_conditioning_scale, controlnet_keep[i])]
                else:
                    controlnet_cond_scale = controlnet_conditioning_scale
                    if isinstance(controlnet_cond_scale, list):
                        controlnet_cond_scale = controlnet_cond_scale[0]
                    cond_scale = controlnet_cond_scale * controlnet_keep[i]

                # # Resize control_image to match the size of the input to the controlnet
                # if control_image.shape[-2:] != control_model_input.shape[-2:]:
                #     control_image = F.interpolate(control_image, size=control_model_input.shape[-2:], mode="bilinear", align_corners=False)

                down_block_res_samples, mid_block_res_sample = self.controlnet(
                    control_model_input,
                    t,
                    encoder_hidden_states=controlnet_prompt_embeds,
                    controlnet_cond=control_image,
                    conditioning_scale=cond_scale,
                    guess_mode=guess_mode,
                    added_cond_kwargs=controlnet_added_cond_kwargs,
                    return_dict=False,
                )

                if guess_mode and self.do_classifier_free_guidance:
                    # Infered ControlNet only for the conditional batch.
                    # To apply the output of ControlNet to both the unconditional and conditional batches,
                    # add 0 to the unconditional batch to keep it unchanged.
                    down_block_res_samples = [torch.cat([torch.zeros_like(d), d]) for d in down_block_res_samples]
                    mid_block_res_sample = torch.cat([torch.zeros_like(mid_block_res_sample), mid_block_res_sample])

                if num_channels_unet == 9:
                    latent_model_input = torch.cat([latent_model_input, mask, masked_image_latents], dim=1)

                # predict the noise residual
                noise_pred = self.unet(
                    latent_model_input,
                    t,
                    encoder_hidden_states=prompt_embeds,
                    cross_attention_kwargs=self.cross_attention_kwargs,
                    down_block_additional_residuals=down_block_res_samples,
                    mid_block_additional_residual=mid_block_res_sample,
                    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 + guidance_scale * (noise_pred_text - noise_pred_uncond)

                if self.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]

                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)

                # 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)

        # make sure the VAE is in float32 mode, as it overflows in float16
        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 we do sequential model offloading, let's offload unet and controlnet
        # manually for max memory savings
        if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
            self.unet.to("cpu")
            self.controlnet.to("cpu")
            torch.cuda.empty_cache()

        if not output_type == "latent":
            image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
        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)