sdxl_decensor_pipeline.py

from typing import Callable

import torch.nn.functional as F
from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl_inpaint import *

from .attention_processor import AttnProcessorFA2


class SDXLDecensorPipeline(StableDiffusionXLInpaintPipeline):
    vae: AutoencoderKL
    text_encoder: CLIPTextModel
    text_encoder_2: CLIPTextModelWithProjection
    tokenizer: CLIPTokenizer
    tokenizer_2: CLIPTokenizer
    unet: UNet2DConditionModel

    def __init__(
        self,
        vae: AutoencoderKL,
        text_encoder: CLIPTextModel,
        text_encoder_2: CLIPTextModelWithProjection,
        tokenizer: CLIPTokenizer,
        tokenizer_2: CLIPTokenizer,
        unet: UNet2DConditionModel,
        scheduler: KarrasDiffusionSchedulers,
        image_encoder: CLIPVisionModelWithProjection = None,
        feature_extractor: CLIPImageProcessor = None,
        requires_aesthetics_score: bool = False,
        force_zeros_for_empty_prompt: bool = True,
        add_watermarker: Optional[bool] = None,
        use_flash_attention: bool = False,
        suppress_tokenizer_warning: bool = False
    ) -> None:
        super(StableDiffusionXLInpaintPipeline).__init__()

        self.register_modules(
            vae=vae,
            text_encoder=text_encoder,
            text_encoder_2=text_encoder_2,
            tokenizer=tokenizer,
            tokenizer_2=tokenizer_2,
            unet=unet,
            image_encoder=image_encoder,
            feature_extractor=feature_extractor,
            scheduler=scheduler,
        )
        self.register_to_config(force_zeros_for_empty_prompt=force_zeros_for_empty_prompt)
        self.register_to_config(requires_aesthetics_score=requires_aesthetics_score)
        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8
        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=False,
            do_convert_grayscale=True
        )

        add_watermarker = add_watermarker if add_watermarker is not None else is_invisible_watermark_available()

        if add_watermarker:
            self.watermark = StableDiffusionXLWatermarker()
        else:
            self.watermark = None

        if use_flash_attention:
            self.unet.set_attn_processor(AttnProcessorFA2())
        
        self.set_progress_bar_config(leave=False)

        if suppress_tokenizer_warning:
            self.tokenizer.deprecation_warnings['sequence-length-is-longer-than-the-specified-maximum'] = True
            self.tokenizer_2.deprecation_warnings['sequence-length-is-longer-than-the-specified-maximum'] = True

    def check_inputs(
        self,
        prompt,
        prompt_2,
        image,
        mask_image,
        height,
        width,
        strength,
        callback_steps,
        output_type,
        negative_prompt=None,
        negative_prompt_2=None,
        prompt_embeds=None,
        negative_prompt_embeds=None,
        ip_adapter_image=None,
        ip_adapter_image_embeds=None,
        callback_on_step_end_tensor_inputs=None,
        padding_mask_crop=None,
    ):
        if strength < 0 or strength > 1:
            raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}")

        if height % 8 != 0 or width % 8 != 0:
            raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")

        if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0):
            raise ValueError(
                f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
                f" {type(callback_steps)}."
            )

        if callback_on_step_end_tensor_inputs is not None and not all(
            k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
        ):
            raise ValueError(
                f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
            )

        if prompt is not None and prompt_embeds is not None:
            raise ValueError(
                f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
                " only forward one of the two."
            )
        elif prompt_2 is not None and prompt_embeds is not None:
            raise ValueError(
                f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
                " only forward one of the two."
            )
        elif prompt is None and prompt_embeds is None:
            raise ValueError(
                "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
            )
        elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
        elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)):
            raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}")

        if negative_prompt is not None and negative_prompt_embeds is not None:
            raise ValueError(
                f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
            )
        elif negative_prompt_2 is not None and negative_prompt_embeds is not None:
            raise ValueError(
                f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:"
                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
            )

        if prompt_embeds is not None and negative_prompt_embeds is not None:
            if prompt_embeds.shape != negative_prompt_embeds.shape:
                raise ValueError(
                    "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
                    f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
                    f" {negative_prompt_embeds.shape}."
                )
        if padding_mask_crop is not None:
            if not isinstance(image, PIL.Image.Image):
                raise ValueError(
                    f"The image should be a PIL image when inpainting mask crop, but is of type {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 {output_type}.")

        if ip_adapter_image is not None and ip_adapter_image_embeds is not None:
            raise ValueError(
                "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined."
            )

        if ip_adapter_image_embeds is not None:
            if not isinstance(ip_adapter_image_embeds, list):
                raise ValueError(
                    f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}"
                )
            elif ip_adapter_image_embeds[0].ndim not in [3, 4]:
                raise ValueError(
                    f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D"
                )

    def prepare_mask_latents(
        self,
        mask,
        masked_image,
        batch_size,
        height,
        width,
        dtype,
        device,
        generator,
        do_classifier_free_guidance,
        dilate_latent_mask: int | None
    ):
        # 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),
            mode='nearest-exact'
        )
        
        if dilate_latent_mask is not None:
            kernel_size = dilate_latent_mask
            kernel_size += 1 - kernel_size % 2
            mask = F.max_pool2d(mask, kernel_size, stride=1, padding=kernel_size // 2)
        
        mask = mask.to(device=device, dtype=dtype)

        # duplicate mask and masked_image_latents for each generation per prompt, using mps friendly method
        if mask.shape[0] < batch_size:
            if not batch_size % mask.shape[0] == 0:
                raise ValueError(
                    "The passed mask and the required batch size don't match. Masks are supposed to be duplicated to"
                    f" a total batch size of {batch_size}, but {mask.shape[0]} masks were passed. Make sure the number"
                    " of masks that you pass is divisible by the total requested batch size."
                )
            mask = mask.repeat(batch_size // mask.shape[0], 1, 1, 1)

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

        if masked_image is not None and masked_image.shape[1] == 4:
            masked_image_latents = masked_image
        else:
            masked_image_latents = None

        if masked_image is not None:
            if masked_image_latents is None:
                masked_image = masked_image.to(device=device, dtype=dtype)
                masked_image_latents = self._encode_vae_image(masked_image, generator=generator)

            if masked_image_latents.shape[0] < batch_size:
                if not batch_size % masked_image_latents.shape[0] == 0:
                    raise ValueError(
                        "The passed images and the required batch size don't match. Images are supposed to be duplicated"
                        f" to a total batch size of {batch_size}, but {masked_image_latents.shape[0]} images were passed."
                        " Make sure the number of images that you pass is divisible by the total requested batch size."
                    )
                masked_image_latents = masked_image_latents.repeat(
                    batch_size // masked_image_latents.shape[0], 1, 1, 1
                )

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

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

        return mask, masked_image_latents

    @torch.no_grad()
    def __call__(
        self,
        prompt: Union[str, List[str]] = None,
        prompt_2: Optional[Union[str, List[str]]] = None,
        image: PipelineImageInput = None,
        mask_image: PipelineImageInput = None,
        masked_image_latents: torch.Tensor = None,
        height: Optional[int] = None,
        width: Optional[int] = None,
        padding_mask_crop: Optional[int] = None,
        strength: float = 0.9999,
        num_inference_steps: int = 50,
        timesteps: List[int] = None,
        sigmas: List[float] = None,
        denoising_start: Optional[float] = None,
        denoising_end: Optional[float] = None,
        guidance_scale: float = 7.5,
        negative_prompt: Optional[Union[str, List[str]]] = None,
        negative_prompt_2: Optional[Union[str, List[str]]] = None,
        num_images_per_prompt: Optional[int] = 1,
        eta: float = 0.0,
        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
        latents: Optional[torch.Tensor] = None,
        prompt_embeds: Optional[torch.Tensor] = None,
        negative_prompt_embeds: Optional[torch.Tensor] = None,
        pooled_prompt_embeds: Optional[torch.Tensor] = None,
        negative_pooled_prompt_embeds: Optional[torch.Tensor] = None,
        ip_adapter_image: Optional[PipelineImageInput] = None,
        ip_adapter_image_embeds: Optional[List[torch.Tensor]] = None,
        output_type: Optional[str] = "pil",
        return_dict: bool = True,
        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
        guidance_rescale: float = 0.0,
        original_size: Tuple[int, int] = None,
        crops_coords_top_left: Tuple[int, int] = (0, 0),
        target_size: Tuple[int, int] = None,
        negative_original_size: Optional[Tuple[int, int]] = None,
        negative_crops_coords_top_left: Tuple[int, int] = (0, 0),
        negative_target_size: Optional[Tuple[int, int]] = None,
        aesthetic_score: float = 6.0,
        negative_aesthetic_score: float = 2.5,
        clip_skip: Optional[int] = None,
        callback_on_step_end: Optional[
            Union[
                Callable[[int, int, Dict], None],
                PipelineCallback,
                MultiPipelineCallbacks
            ]
        ] = None,
        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
        ensure_image_consistency: bool = False,
        dilate_latent_mask: int | None = None,
        **kwargs
    ):
        r"""
        Function invoked when calling the pipeline for generation.

        Args:
            prompt (`str` or `List[str]`, *optional*):
                The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
                instead.
            prompt_2 (`str` or `List[str]`, *optional*):
                The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
                used in both text-encoders
            image (`PIL.Image.Image`):
                `Image`, or tensor representing an image batch which will be inpainted, *i.e.* parts of the image will
                be masked out with `mask_image` and repainted according to `prompt`.
            mask_image (`PIL.Image.Image`):
                `Image`, or tensor representing an image batch, to mask `image`. White pixels in the mask will be
                repainted, while black pixels will be preserved. If `mask_image` is a PIL image, it will be converted
                to a single channel (luminance) before use. If it's a tensor, it should contain one color channel (L)
                instead of 3, so the expected shape would be `(B, H, W, 1)`.
            height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
                The height in pixels of the generated image. This is set to 1024 by default for the best results.
                Anything below 512 pixels won't work well for
                [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0)
                and checkpoints that are not specifically fine-tuned on low resolutions.
            width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
                The width in pixels of the generated image. This is set to 1024 by default for the best results.
                Anything below 512 pixels won't work well for
                [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0)
                and checkpoints that are not specifically fine-tuned on low resolutions.
            padding_mask_crop (`int`, *optional*, defaults to `None`):
                The size of margin in the crop to be applied to the image and masking. If `None`, no crop is applied to
                image and mask_image. If `padding_mask_crop` is not `None`, it will first find a rectangular region
                with the same aspect ration of the image and contains all masked area, and then expand that area based
                on `padding_mask_crop`. The image and mask_image will then be cropped based on the expanded area before
                resizing to the original image size for inpainting. This is useful when the masked area is small while
                the image is large and contain information irrelevant for inpainting, such as background.
            strength (`float`, *optional*, defaults to 0.9999):
                Conceptually, indicates how much to transform the masked portion of the reference `image`. Must be
                between 0 and 1. `image` will be used as a starting point, adding more noise to it the larger the
                `strength`. The number of denoising steps depends on the amount of noise initially added. When
                `strength` is 1, added noise will be maximum and the denoising process will run for the full number of
                iterations specified in `num_inference_steps`. A value of 1, therefore, essentially ignores the masked
                portion of the reference `image`. Note that in the case of `denoising_start` being declared as an
                integer, the value of `strength` will be ignored.
            num_inference_steps (`int`, *optional*, defaults to 50):
                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
                expense of slower inference.
            timesteps (`List[int]`, *optional*):
                Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument
                in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is
                passed will be used. Must be in descending order.
            sigmas (`List[float]`, *optional*):
                Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in
                their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed
                will be used.
            denoising_start (`float`, *optional*):
                When specified, indicates the fraction (between 0.0 and 1.0) of the total denoising process to be
                bypassed before it is initiated. Consequently, the initial part of the denoising process is skipped and
                it is assumed that the passed `image` is a partly denoised image. Note that when this is specified,
                strength will be ignored. The `denoising_start` parameter is particularly beneficial when this pipeline
                is integrated into a "Mixture of Denoisers" multi-pipeline setup, as detailed in [**Refining the Image
                Output**](https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion/stable_diffusion_xl#refining-the-image-output).
            denoising_end (`float`, *optional*):
                When specified, determines the fraction (between 0.0 and 1.0) of the total denoising process to be
                completed before it is intentionally prematurely terminated. As a result, the returned sample will
                still retain a substantial amount of noise (ca. final 20% of timesteps still needed) and should be
                denoised by a successor pipeline that has `denoising_start` set to 0.8 so that it only denoises the
                final 20% of the scheduler. The denoising_end parameter should ideally be utilized when this pipeline
                forms a part of a "Mixture of Denoisers" multi-pipeline setup, as elaborated in [**Refining the Image
                Output**](https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion/stable_diffusion_xl#refining-the-image-output).
            guidance_scale (`float`, *optional*, defaults to 7.5):
                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
                `guidance_scale` is defined as `w` of equation 2. of [Imagen
                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
                usually at the expense of lower image quality.
            negative_prompt (`str` or `List[str]`, *optional*):
                The prompt or prompts not to guide the image generation. If not defined, one has to pass
                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
                less than `1`).
            negative_prompt_2 (`str` or `List[str]`, *optional*):
                The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and
                `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders
            prompt_embeds (`torch.Tensor`, *optional*):
                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
                provided, text embeddings will be generated from `prompt` input argument.
            negative_prompt_embeds (`torch.Tensor`, *optional*):
                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
                argument.
            pooled_prompt_embeds (`torch.Tensor`, *optional*):
                Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
                If not provided, pooled text embeddings will be generated from `prompt` input argument.
            negative_pooled_prompt_embeds (`torch.Tensor`, *optional*):
                Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
                weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt`
                input argument.
            ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters.
            ip_adapter_image_embeds (`List[torch.Tensor]`, *optional*):
                Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of
                IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should
                contain the negative image embedding if `do_classifier_free_guidance` is set to `True`. If not
                provided, embeddings are computed from the `ip_adapter_image` input argument.
            num_images_per_prompt (`int`, *optional*, defaults to 1):
                The number of images to generate per prompt.
            eta (`float`, *optional*, defaults to 0.0):
                Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
                [`schedulers.DDIMScheduler`], will be ignored for others.
            generator (`torch.Generator`, *optional*):
                One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
                to make generation deterministic.
            latents (`torch.Tensor`, *optional*):
                Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
                tensor will ge generated by sampling using the supplied random `generator`.
            output_type (`str`, *optional*, defaults to `"pil"`):
                The output format of the generate image. Choose between
                [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
            return_dict (`bool`, *optional*, defaults to `True`):
                Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
                plain tuple.
            cross_attention_kwargs (`dict`, *optional*):
                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
                `self.processor` in
                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
            original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
                If `original_size` is not the same as `target_size` the image will appear to be down- or upsampled.
                `original_size` defaults to `(height, width)` if not specified. Part of SDXL's micro-conditioning as
                explained in section 2.2 of
                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
            crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)):
                `crops_coords_top_left` can be used to generate an image that appears to be "cropped" from the position
                `crops_coords_top_left` downwards. Favorable, well-centered images are usually achieved by setting
                `crops_coords_top_left` to (0, 0). Part of SDXL's micro-conditioning as explained in section 2.2 of
                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
            target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
                For most cases, `target_size` should be set to the desired height and width of the generated image. If
                not specified it will default to `(height, width)`. Part of SDXL's micro-conditioning as explained in
                section 2.2 of [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
            negative_original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
                To negatively condition the generation process based on a specific image resolution. Part of SDXL's
                micro-conditioning as explained in section 2.2 of
                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more
                information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208.
            negative_crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)):
                To negatively condition the generation process based on a specific crop coordinates. Part of SDXL's
                micro-conditioning as explained in section 2.2 of
                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more
                information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208.
            negative_target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
                To negatively condition the generation process based on a target image resolution. It should be as same
                as the `target_size` for most cases. Part of SDXL's micro-conditioning as explained in section 2.2 of
                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more
                information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208.
            aesthetic_score (`float`, *optional*, defaults to 6.0):
                Used to simulate an aesthetic score of the generated image by influencing the positive text condition.
                Part of SDXL's micro-conditioning as explained in section 2.2 of
                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
            negative_aesthetic_score (`float`, *optional*, defaults to 2.5):
                Part of SDXL's micro-conditioning as explained in section 2.2 of
                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). Can be used to
                simulate an aesthetic score of the generated image by influencing the negative text condition.
            clip_skip (`int`, *optional*):
                Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
                the output of the pre-final layer will be used for computing the prompt embeddings.
            callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*):
                A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of
                each denoising step during the inference. with the following arguments: `callback_on_step_end(self:
                DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a
                list of all tensors as specified by `callback_on_step_end_tensor_inputs`.
            callback_on_step_end_tensor_inputs (`List`, *optional*):
                The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
                will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
                `._callback_tensor_inputs` attribute of your pipeline class.

        Examples:

        Returns:
            [`~pipelines.stable_diffusion.StableDiffusionXLPipelineOutput`] or `tuple`:
            [`~pipelines.stable_diffusion.StableDiffusionXLPipelineOutput`] if `return_dict` is True, otherwise a
            `tuple. `tuple. When returning a tuple, the first element is a list with the generated images.
        """

        callback = kwargs.pop("callback", None)
        callback_steps = kwargs.pop("callback_steps", None)

        if callback is not None:
            deprecate(
                "callback",
                "1.0.0",
                "Passing `callback` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`",
            )
        if callback_steps is not None:
            deprecate(
                "callback_steps",
                "1.0.0",
                "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`",
            )

        if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)):
            callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs

        # 0. Default height and width to unet
        height = height or self.unet.ocnfig.sample_size * self.vae_scale_factor
        width = width or self.unet.config.sample_size * self.vae_scale_factor

        if prompt is None and prompt_embeds is None:
            prompt = ''
            guidance_scale = 0.0

        # 1. Check inputs
        self.check_inputs(
            prompt,
            prompt_2,
            image,
            mask_image,
            height,
            width,
            strength,
            callback_steps,
            output_type,
            negative_prompt,
            negative_prompt_2,
            prompt_embeds,
            negative_prompt_embeds,
            ip_adapter_image,
            ip_adapter_image_embeds,
            callback_on_step_end_tensor_inputs,
            padding_mask_crop,
        )

        self._guidance_scale = guidance_scale
        self._guidance_rescale = guidance_rescale
        self._clip_skip = clip_skip
        self._cross_attention_kwargs = cross_attention_kwargs
        self._denoising_end = denoising_end
        self._denoising_start = denoising_start
        self._interrupt = False

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

        device = self._execution_device

        # 3. Encode input prompt
        text_encoder_lora_scale = (
            self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None
        )

        (
            prompt_embeds,
            negative_prompt_embeds,
            pooled_prompt_embeds,
            negative_pooled_prompt_embeds,
        ) = self.encode_prompt(
            prompt=prompt,
            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(dnv, float) and 0 < dnv < 1

        timesteps, num_inference_steps = retrieve_timesteps(
            self.scheduler, num_inference_steps, device, timesteps, sigmas
        )
        timesteps, num_inference_steps = self.get_timesteps(
            num_inference_steps,
            strength,
            device,
            denoising_start=self.denoising_start if denoising_value_valid(self.denoising_start) else None,
        )
        # check that number of inference steps is not < 1 - as this doesn't make sense
        if num_inference_steps < 1:
            raise ValueError(
                f"After adjusting the num_inference_steps by strength parameter: {strength}, the number of pipeline"
                f"steps is {num_inference_steps} which is < 1 and not appropriate for this pipeline."
            )
        # at which timestep to set the initial noise (n.b. 50% if strength is 0.5)
        latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt)
        # create a boolean to check if the strength is set to 1. if so then initialise the latents with pure noise
        is_strength_max = strength == 1.0

        # 5. Preprocess mask and image
        if padding_mask_crop is not None:
            crops_coords = self.mask_processor.get_crop_region(
                mask_image,
                width,
                height,
                pad=padding_mask_crop
            )
            resize_mode = "fill"
        else:
            crops_coords = None
            resize_mode = "default"

        original_image = image
        init_image = self.image_processor.preprocess(
            image,
            height=height,
            width=width,
            crops_coords=crops_coords,
            resize_mode=resize_mode
        )
        init_image = init_image.to(dtype=torch.float32)

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

        if masked_image_latents is not None:
            masked_image = masked_image_latents
        elif init_image.shape[1] == 4:
            # if images are in latent space, we can't mask it
            masked_image = None
        else:
            # masked_image = init_image * (mask < 0.5)
            masked_image = torch.where(mask < 1.0, init_image, 0.5)

        # 6. Prepare latent variables
        num_channels_latents = self.vae.config.latent_channels
        num_channels_unet = self.unet.config.in_channels
        # return_image_latents = num_channels_unet == 4
        return_image_latents = ensure_image_consistency

        add_noise = True if self.denoising_start is None else False
        latents_outputs = self.prepare_latents(
            batch_size * num_images_per_prompt,
            num_channels_latents,
            height,
            width,
            prompt_embeds.dtype,
            device,
            generator,
            latents,
            image=init_image,
            timestep=latent_timestep,
            is_strength_max=is_strength_max,
            add_noise=add_noise,
            return_noise=True,
            return_image_latents=return_image_latents,
        )

        if return_image_latents:
            latents, noise, image_latents = latents_outputs
        else:
            latents, noise = latents_outputs

        # 7. Prepare mask latent variables
        latent_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,
            dilate_latent_mask
        )

        # 8. Check that sizes of mask, masked image and latents match
        if num_channels_unet == 9:
            # default case for runwayml/stable-diffusion-inpainting
            num_channels_mask = mask.shape[1]
            num_channels_masked_image = masked_image_latents.shape[1]
            if num_channels_latents + num_channels_mask + num_channels_masked_image != self.unet.config.in_channels:
                raise ValueError(
                    f"Incorrect configuration settings! The config of `pipeline.unet`: {self.unet.config} expects"
                    f" {self.unet.config.in_channels} but received `num_channels_latents`: {num_channels_latents} +"
                    f" `num_channels_mask`: {num_channels_mask} + `num_channels_masked_image`: {num_channels_masked_image}"
                    f" = {num_channels_latents + num_channels_masked_image + num_channels_mask}. Please verify the config of"
                    " `pipeline.unet` or your `mask_image` or `image` input."
                )
        elif num_channels_unet != 4:
            raise ValueError(
                f"The unet {self.unet.__class__} should have either 4 or 9 input channels, not {self.unet.config.in_channels}."
            )
        # 8.1 Prepare extra step kwargs.
        extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)

        # 9. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
        height, width = latents.shape[-2:]
        height = height * self.vae_scale_factor
        width = width * self.vae_scale_factor

        original_size = original_size or (height, width)
        target_size = target_size or (height, width)

        # 10. Prepare added time ids & embeddings
        if negative_original_size is None:
            negative_original_size = original_size
        if negative_target_size is None:
            negative_target_size = target_size

        add_text_embeds = pooled_prompt_embeds
        if self.text_encoder_2 is None:
            text_encoder_projection_dim = int(pooled_prompt_embeds.shape[-1])
        else:
            text_encoder_projection_dim = self.text_encoder_2.config.projection_dim

        add_time_ids, add_neg_time_ids = self._get_add_time_ids(
            original_size,
            crops_coords_top_left,
            target_size,
            aesthetic_score,
            negative_aesthetic_score,
            negative_original_size,
            negative_crops_coords_top_left,
            negative_target_size,
            dtype=prompt_embeds.dtype,
            text_encoder_projection_dim=text_encoder_projection_dim,
        )
        add_time_ids = add_time_ids.repeat(batch_size * num_images_per_prompt, 1)

        if self.do_classifier_free_guidance:
            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
            add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0)
            add_neg_time_ids = add_neg_time_ids.repeat(batch_size * num_images_per_prompt, 1)
            add_time_ids = torch.cat([add_neg_time_ids, add_time_ids], dim=0)

        prompt_embeds = prompt_embeds.to(device)
        add_text_embeds = add_text_embeds.to(device)
        add_time_ids = add_time_ids.to(device)

        if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
            image_embeds = self.prepare_ip_adapter_image_embeds(
                ip_adapter_image,
                ip_adapter_image_embeds,
                device,
                batch_size * num_images_per_prompt,
                self.do_classifier_free_guidance,
            )

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

        if (
            self.denoising_end is not None
            and self.denoising_start is not None
            and denoising_value_valid(self.denoising_end)
            and denoising_value_valid(self.denoising_start)
            and self.denoising_start >= self.denoising_end
        ):
            raise ValueError(
                f"`denoising_start`: {self.denoising_start} cannot be larger than or equal to `denoising_end`: "
                + f" {self.denoising_end} when using type float."
            )
        elif self.denoising_end is not None and denoising_value_valid(self.denoising_end):
            discrete_timestep_cutoff = int(
                round(
                    self.scheduler.config.num_train_timesteps
                    - (self.denoising_end * self.scheduler.config.num_train_timesteps)
                )
            )
            num_inference_steps = len(list(filter(lambda ts: ts >= discrete_timestep_cutoff, timesteps)))
            timesteps = timesteps[:num_inference_steps]

        # 11.1 Optionally get Guidance Scale Embedding
        timestep_cond = None
        if self.unet.config.time_cond_proj_dim is not None:
            guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt)
            timestep_cond = self.get_guidance_scale_embedding(
                guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim
            ).to(device=device, dtype=latents.dtype)

        self._num_timesteps = len(timesteps)
        with self.progress_bar(total=num_inference_steps) as progress_bar:
            for i, t in enumerate(timesteps):
                if self.interrupt:
                    continue
                # expand the latents if we are doing classifier free guidance
                latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents

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

                # predict the noise residual
                added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids}
                if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
                    added_cond_kwargs["image_embeds"] = image_embeds
                
                noise_pred = self.unet(
                    latent_model_input,
                    t,
                    encoder_hidden_states=prompt_embeds,
                    timestep_cond=timestep_cond,
                    cross_attention_kwargs=self.cross_attention_kwargs,
                    added_cond_kwargs=added_cond_kwargs,
                    return_dict=False
                )[0]

                # perform guidance
                if self.do_classifier_free_guidance:
                    noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
                    noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond)

                if self.do_classifier_free_guidance and self.guidance_rescale > 0.0:
                    # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf
                    noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=self.guidance_rescale)

                # compute the previous noisy sample x_t -> x_t-1
                latents_dtype = latents.dtype
                latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
                if latents.dtype != latents_dtype:
                    if torch.backends.mps.is_available():
                        # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272
                        latents = latents.to(latents_dtype)

                if ensure_image_consistency:
                    init_latents_proper = image_latents
                    if self.do_classifier_free_guidance:
                        init_mask, _ = latent_mask.chunk(2)
                    else:
                        init_mask = latent_mask

                    init_mask = init_mask.bool().to(latents)
                
                    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)
                    add_text_embeds = callback_outputs.pop("add_text_embeds", add_text_embeds)
                    add_time_ids = callback_outputs.pop("add_time_ids", add_time_ids)
                    mask = callback_outputs.pop("mask", mask)
                    masked_image_latents = callback_outputs.pop("masked_image_latents", masked_image_latents)

                # call the callback, if provided
                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
                    progress_bar.update()
                    if callback is not None and i % callback_steps == 0:
                        step_idx = i // getattr(self.scheduler, "order", 1)
                        callback(step_idx, t, latents)

                if XLA_AVAILABLE:
                    xm.mark_step()

        if not output_type == "latent":
            # make sure the VAE is in float32 mode, as it overflows in float16
            needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast

            if needs_upcasting:
                self.upcast_vae()
                latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
            elif latents.dtype != self.vae.dtype:
                if torch.backends.mps.is_available():
                    # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272
                    self.vae = self.vae.to(latents.dtype)

            # unscale/denormalize the latents
            # denormalize with the mean and std if available and not None
            has_latents_mean = hasattr(self.vae.config, "latents_mean") and self.vae.config.latents_mean is not None
            has_latents_std = hasattr(self.vae.config, "latents_std") and self.vae.config.latents_std is not None
            if has_latents_mean and has_latents_std:
                latents_mean = (
                    torch.tensor(self.vae.config.latents_mean).view(1, 4, 1, 1).to(latents.device, latents.dtype)
                )
                latents_std = (
                    torch.tensor(self.vae.config.latents_std).view(1, 4, 1, 1).to(latents.device, latents.dtype)
                )
                latents = latents * latents_std / self.vae.config.scaling_factor + latents_mean
            else:
                latents = latents / self.vae.config.scaling_factor

            image = self.vae.decode(latents, return_dict=False)[0]

            if ensure_image_consistency:
                init_image = init_image.to(image)
                init_mask = mask.bool().to(image)
                kernel_size = 11
                init_mask = F.max_pool2d(init_mask, kernel_size=kernel_size, stride=1, padding=kernel_size // 2)
                init_mask = F.avg_pool2d(init_mask, kernel_size=kernel_size, stride=1, padding=kernel_size // 2)
                image = (1 - init_mask) * init_image + init_mask * image.clamp(-1, 1)

            # cast back to fp16 if needed
            if needs_upcasting:
                self.vae.to(dtype=torch.float16)
        else:
            return StableDiffusionXLPipelineOutput(images=latents)

        # apply watermark if available
        if self.watermark is not None:
            image = self.watermark.apply_watermark(image)

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

        if padding_mask_crop is not None:
            image = [self.image_processor.apply_overlay(mask_image, original_image, i, crops_coords) for i in image]

        # Offload all models
        self.maybe_free_model_hooks()

        if not return_dict:
            return (image,)

        return StableDiffusionXLPipelineOutput(images=image)