artistic_portrait/pipeline.py

from diffusers import StableDiffusionXLControlNetPipeline
from diffusers.pipelines.controlnet.pipeline_controlnet_sd_xl import *
from .pulid_encoder import PuLIDEncoder
from csd_clip import create_model_and_transforms as create_csd_clip_model_and_transforms
from csd_clip import CSD_CLIP
from ip_adapter_diffusers.ip_adapter import *
from transformers import CLIPVisionModelWithProjection


class ArtisticPortraitXLPipeline(StableDiffusionXLControlNetPipeline):
    def __init__(
        self,
        vae: AutoencoderKL,
        text_encoder: CLIPTextModel,
        text_encoder_2: CLIPTextModelWithProjection,
        tokenizer: CLIPTokenizer,
        tokenizer_2: CLIPTokenizer,
        unet: UNet2DConditionModel,
        controlnet: Union[
            ControlNetModel,
            List[ControlNetModel],
            Tuple[ControlNetModel],
            MultiControlNetModel,
        ],
        scheduler: KarrasDiffusionSchedulers,
        force_zeros_for_empty_prompt: bool = True,
        add_watermarker: Optional[bool] = None,
        feature_extractor: CLIPImageProcessor = None,
        image_encoder: CLIPVisionModelWithProjection = None,
        style_adapter_path=None,
        id_adapter_path=None,
        style_image_encoder_path="models/h94/IP-Adapter/sdxl_models/image_encoder",
        device=None,
    ):
        super().__init__(
            vae=vae,
            text_encoder=text_encoder,
            text_encoder_2=text_encoder_2,
            tokenizer=tokenizer,
            tokenizer_2=tokenizer_2,
            unet=unet,
            controlnet=controlnet,
            scheduler=scheduler,
            force_zeros_for_empty_prompt=force_zeros_for_empty_prompt,
            add_watermarker=add_watermarker,
            feature_extractor=feature_extractor,
            image_encoder=image_encoder,
        )
        self.id_image_encoder = PuLIDEncoder(device=device)
        if "art" in style_adapter_path:
            self.style_image_encoder = create_csd_clip_model_and_transforms()[0]
        else:
            self.style_image_encoder = CLIPVisionModelWithProjection.from_pretrained(
                style_image_encoder_path
            )

        self.style_image_processor = CLIPImageProcessor()
        load_multi_ip_adapter(
            self.unet,
            paths=[style_adapter_path, id_adapter_path],
        )
        self.style_image_projection_layer = (
            self.unet.encoder_hid_proj.image_projection_layers[0]
        )
        self.id_image_projection_layer = (
            self.unet.encoder_hid_proj.image_projection_layers[1]
        )

    def load_style_adapter_to_controlnet(self, style_adapter_path):
        load_ip_adapter(self.controlnet, style_adapter_path)

    def get_id_hidden_states(self, image):
        if not isinstance(image, list):
            image = [image]
        image = [
            (
                single_image
                if isinstance(single_image, np.ndarray)
                else np.array(single_image)
            )
            for single_image in image
        ]
        id_cond, id_vit_hidden, id_uncond, id_vit_hidden_uncond = (
            self.id_image_encoder.get_id_embedding(image)
        )

        id_vit_hidden = [x.to(dtype=self.unet.dtype) for x in id_vit_hidden]
        id_vit_hidden_uncond = [
            x.to(dtype=self.unet.dtype) for x in id_vit_hidden_uncond
        ]
        uncond_id_embedding = self.id_image_projection_layer(
            id_uncond.to(self.unet.device, self.unet.dtype),
            id_vit_hidden_uncond,
        )
        id_embedding = self.id_image_projection_layer(
            id_cond.to(self.unet.device, self.unet.dtype), id_vit_hidden
        )
        id_hidden_states = torch.concat([uncond_id_embedding, id_embedding], dim=0)
        torch.cuda.empty_cache()
        return id_hidden_states

    def get_style_hidden_states(self, image):
        if isinstance(self.style_image_encoder, CSD_CLIP):
            self.style_image_encoder = self.style_image_encoder.to(
                self._execution_device, dtype=torch.float32
            )
            style_pixel_values = self.style_image_processor.preprocess(
                image, return_tensors="pt"
            ).pixel_values
            _, __, style_image_embeds = self.style_image_encoder(
                style_pixel_values.to(self._execution_device, torch.float32)
            )
            style_image_embeds = torch.stack(
                [
                    torch.zeros_like(style_image_embeds).to(self._execution_device),
                    style_image_embeds,
                ]
            ).to(self._execution_device, torch.float16)
            style_ip_adapter_hidden_states = self.style_image_projection_layer(
                style_image_embeds
            )

        elif isinstance(self.style_image_encoder, CLIPVisionModelWithProjection):
            self.style_image_encoder = self.style_image_encoder.to(
                self._execution_device, dtype=torch.float16
            )
            style_pixel_values = self.style_image_processor.preprocess(
                image, return_tensors="pt"
            ).pixel_values
            style_image_embeds = self.style_image_encoder(
                style_pixel_values.to(self._execution_device, torch.float16)
            ).image_embeds
            style_image_embeds = torch.stack(
                [
                    torch.zeros_like(style_image_embeds).to(self._execution_device),
                    style_image_embeds,
                ]
            ).to(self._execution_device, torch.float16)
            style_ip_adapter_hidden_states = self.style_image_projection_layer(
                style_image_embeds
            )

        torch.cuda.empty_cache()
        self.style_image_encoder = self.style_image_encoder.to("cpu")

        return style_ip_adapter_hidden_states

    def set_style_adapter_scale(self, style_adapter_scale):
        for name, processor in self.unet.attn_processors.items():
            if (
                isinstance(processor, torch.nn.Module)
                and "up_blocks.0.attentions.1" in name
            ):
                processor.scale = [style_adapter_scale, 0.0]

    def set_id_adapter_scale(self, id_adapter_scale):
        for name, processor in self.unet.attn_processors.items():
            if (
                isinstance(processor, torch.nn.Module)
                and "up_blocks.0.attentions.1" not in name
            ):
                processor.scale = [0.0, id_adapter_scale]

    @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,
        control_image: PipelineImageInput = None,
        style_image: PipelineImageInput = None,
        id_image: PipelineImageInput = None,
        height: Optional[int] = None,
        width: Optional[int] = None,
        num_inference_steps: int = 50,
        timesteps: List[int] = None,
        sigmas: List[float] = None,
        denoising_end: Optional[float] = None,
        guidance_scale: float = 5.0,
        id_adapter_scale=1.0,
        style_adapter_scale=1.0,
        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.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,
        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,
        style_guidance_start=0.0,
        style_guidance_end=1.0,
        id_guidance_start=0.0,
        id_guidance_end=1.0,
        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,
        clip_skip: Optional[int] = None,
        callback_on_step_end: Optional[
            Union[
                Callable[[int, int, Dict], None],
                PipelineCallback,
                MultiPipelineCallbacks,
            ]
        ] = None,
        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
        **kwargs,
    ):
        r"""
        The call function to the pipeline for generation.

        Args:
            prompt (`str` or `List[str]`, *optional*):
                The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`.
            prompt_2 (`str` or `List[str]`, *optional*):
                The prompt or prompts to be sent to `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
                used in both text-encoders.
            image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.Tensor]`, `List[PIL.Image.Image]`, `List[np.ndarray]`,:
                    `List[List[torch.Tensor]]`, `List[List[np.ndarray]]` or `List[List[PIL.Image.Image]]`):
                The ControlNet input condition to provide guidance to the `unet` for generation. If the type is
                specified as `torch.Tensor`, it is passed to ControlNet as is. `PIL.Image.Image` can also be accepted
                as an image. The dimensions of the output image defaults to `image`'s dimensions. If height and/or
                width are passed, `image` is resized accordingly. If multiple ControlNets are specified in `init`,
                images must be passed as a list such that each element of the list can be correctly batched for input
                to a single ControlNet.
            height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
                The height in pixels of the generated image. 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. 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.
            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_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 as determined by the discrete timesteps selected by 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 5.0):
                A higher guidance scale value encourages the model to generate images closely linked to the text
                `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
            negative_prompt (`str` or `List[str]`, *optional*):
                The prompt or prompts to guide what to not include in image generation. If not defined, you need to
                pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
            negative_prompt_2 (`str` or `List[str]`, *optional*):
                The prompt or prompts to guide what to not include in image generation. This is sent to `tokenizer_2`
                and `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders.
            num_images_per_prompt (`int`, *optional*, defaults to 1):
                The number of images to generate per prompt.
            eta (`float`, *optional*, defaults to 0.0):
                Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
                to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
                A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
                generation deterministic.
            latents (`torch.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 is generated by sampling using the supplied random `generator`.
            prompt_embeds (`torch.Tensor`, *optional*):
                Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
                provided, text embeddings are generated from the `prompt` input argument.
            negative_prompt_embeds (`torch.Tensor`, *optional*):
                Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
                not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
            pooled_prompt_embeds (`torch.Tensor`, *optional*):
                Pre-generated pooled text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
                not provided, pooled text embeddings are 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 (prompt
                weighting). If not provided, pooled `negative_prompt_embeds` are 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.
            output_type (`str`, *optional*, defaults to `"pil"`):
                The output format of the generated image. Choose between `PIL.Image` or `np.array`.
            return_dict (`bool`, *optional*, defaults to `True`):
                Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
                plain tuple.
            cross_attention_kwargs (`dict`, *optional*):
                A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in
                [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
            controlnet_conditioning_scale (`float` or `List[float]`, *optional*, defaults to 1.0):
                The outputs of the ControlNet are multiplied by `controlnet_conditioning_scale` before they are added
                to the residual in the original `unet`. If multiple ControlNets are specified in `init`, you can set
                the corresponding scale as a list.
            guess_mode (`bool`, *optional*, defaults to `False`):
                The ControlNet encoder tries to recognize the content of the input image even if you remove all
                prompts. A `guidance_scale` value between 3.0 and 5.0 is recommended.
            control_guidance_start (`float` or `List[float]`, *optional*, defaults to 0.0):
                The percentage of total steps at which the ControlNet starts applying.
            control_guidance_end (`float` or `List[float]`, *optional*, defaults to 1.0):
                The percentage of total steps at which the ControlNet stops applying.
            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.
            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.StableDiffusionPipelineOutput`] or `tuple`:
                If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned,
                otherwise a `tuple` is returned containing the output 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`",
            )

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

        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],
            )

        # 1. Check inputs. Raise error if not correct
        # self.check_inputs(
        #     prompt,
        #     prompt_2,
        #     control_image,
        #     callback_steps,
        #     negative_prompt,
        #     negative_prompt_2,
        #     prompt_embeds,
        #     negative_prompt_embeds,
        #     pooled_prompt_embeds,
        #     ip_adapter_image,
        #     ip_adapter_image_embeds,
        #     negative_pooled_prompt_embeds,
        #     controlnet_conditioning_scale,
        #     control_guidance_start,
        #     control_guidance_end,
        #     callback_on_step_end_tensor_inputs,
        # )

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

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

        global_pool_conditions = (
            controlnet.config.global_pool_conditions
            if isinstance(controlnet, ControlNetModel)
            else controlnet.nets[0].config.global_pool_conditions
        )
        guess_mode = guess_mode or global_pool_conditions

        # 3.1 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_2,
            device,
            num_images_per_prompt,
            self.do_classifier_free_guidance,
            negative_prompt,
            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,
        )

        # 3.2 Encode ip_adapter_image
        style_hidden_states = self.get_style_hidden_states(style_image)
        id_hidden_states = self.get_id_hidden_states(id_image)
        set_multi_ip_hidden_states(
            self.unet,
            [
                style_hidden_states,
                id_hidden_states,
            ],
        )
        set_ip_hidden_states(self.controlnet, style_hidden_states)
        self.set_id_adapter_scale(id_adapter_scale)
        self.set_style_adapter_scale(style_adapter_scale)
        # 4. Prepare image
        if isinstance(controlnet, ControlNetModel) and control_image is not None:
            control_image = self.prepare_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,
                do_classifier_free_guidance=self.do_classifier_free_guidance,
                guess_mode=guess_mode,
            )
            height, width = control_image.shape[-2:]
        elif isinstance(controlnet, MultiControlNetModel) and control_image is not None:
            images = []

            for image_ in control_image:
                image_ = self.prepare_image(
                    image=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,
                    do_classifier_free_guidance=self.do_classifier_free_guidance,
                    guess_mode=guess_mode,
                )

                images.append(image_)

            control_image = images
            height, width = control_image[0].shape[-2:]

        # 5. Prepare timesteps
        timesteps, num_inference_steps = retrieve_timesteps(
            self.scheduler, num_inference_steps, device, timesteps, sigmas
        )
        self._num_timesteps = len(timesteps)

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

        # 6.5 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)

        # 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
        extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)

        # 7.1 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)
            ]
            controlnet_keep.append(
                keeps[0] if isinstance(controlnet, ControlNetModel) else keeps
            )

        # 7.2 Prepare added time ids & embeddings
        if control_image is None:
            original_size = original_size
            original_size = original_size or (height, width)
            target_size = target_size or (height, width)
        else:
            if isinstance(control_image, list):
                original_size = original_size or control_image[0].shape[-2:]
            else:
                original_size = original_size or control_image.shape[-2:]
        target_size = target_size or (height, width)

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

        add_time_ids = self._get_add_time_ids(
            original_size,
            crops_coords_top_left,
            target_size,
            dtype=prompt_embeds.dtype,
            text_encoder_projection_dim=text_encoder_projection_dim,
        )

        if negative_original_size is not None and negative_target_size is not None:
            negative_add_time_ids = self._get_add_time_ids(
                negative_original_size,
                negative_crops_coords_top_left,
                negative_target_size,
                dtype=prompt_embeds.dtype,
                text_encoder_projection_dim=text_encoder_projection_dim,
            )
        else:
            negative_add_time_ids = add_time_ids

        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_time_ids = torch.cat([negative_add_time_ids, add_time_ids], dim=0)

        prompt_embeds = prompt_embeds.to(device)
        add_text_embeds = add_text_embeds.to(device)
        add_time_ids = add_time_ids.to(device).repeat(
            batch_size * num_images_per_prompt, 1
        )

        # 8. Denoising loop
        num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order

        # 8.1 Apply denoising_end
        if (
            self.denoising_end is not None
            and isinstance(self.denoising_end, float)
            and self.denoising_end > 0
            and self.denoising_end < 1
        ):
            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]

        is_unet_compiled = is_compiled_module(self.unet)
        is_controlnet_compiled = is_compiled_module(self.controlnet)
        is_torch_higher_equal_2_1 = is_torch_version(">=", "2.1")
        with self.progress_bar(total=num_inference_steps) as progress_bar:
            for i, t in enumerate(timesteps):
                if self.interrupt:
                    continue

                # Relevant thread:
                # https://dev-discuss.pytorch.org/t/cudagraphs-in-pytorch-2-0/1428
                if (
                    is_unet_compiled and is_controlnet_compiled
                ) and is_torch_higher_equal_2_1:
                    torch._inductor.cudagraph_mark_step_begin()
                # 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
                )
                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]

                if control_image is not None and controlnet_conditioning_scale != 0.0:
                    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,
                    )
                else:
                    down_block_res_samples = None
                    mid_block_res_sample = None

                if (
                    guess_mode
                    and self.do_classifier_free_guidance
                    and control_image is not None
                ):
                    # Inferred 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 (
                #     i / num_inference_steps >= style_guidance_start
                #     and i / num_inference_steps <= style_guidance_end
                # ):
                #     self.set_style_adapter_scale(style_adapter_scale)
                # else:
                #     self.set_style_adapter_scale(0.0)

                # if (
                #     i / num_inference_steps >= id_guidance_start
                #     and i / num_inference_steps <= id_guidance_end
                # ):
                #     self.set_id_adapter_scale(id_adapter_scale)
                # else:
                #     self.set_id_adapter_scale(0.0)
                # predict the noise residual
                noise_pred = self.unet(
                    latent_model_input,
                    t,
                    encoder_hidden_states=prompt_embeds,
                    timestep_cond=timestep_cond,
                    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
                    )

                # 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 callback_on_step_end is not None:
                    callback_kwargs = {}
                    for k in callback_on_step_end_tensor_inputs:
                        callback_kwargs[k] = locals()[k]
                    callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)

                    latents = callback_outputs.pop("latents", latents)
                    prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
                    negative_prompt_embeds = callback_outputs.pop(
                        "negative_prompt_embeds", negative_prompt_embeds
                    )
                    add_text_embeds = callback_outputs.pop(
                        "add_text_embeds", add_text_embeds
                    )
                    negative_pooled_prompt_embeds = callback_outputs.pop(
                        "negative_pooled_prompt_embeds", negative_pooled_prompt_embeds
                    )
                    add_time_ids = callback_outputs.pop("add_time_ids", add_time_ids)
                    negative_add_time_ids = callback_outputs.pop(
                        "negative_add_time_ids", negative_add_time_ids
                    )

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

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

            if needs_upcasting:
                self.upcast_vae()
                latents = latents.to(
                    next(iter(self.vae.post_quant_conv.parameters())).dtype
                )

            # 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

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

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

        if not output_type == "latent":
            # apply watermark if available
            if self.watermark is not None:
                control_image = self.watermark.apply_watermark(control_image)

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

        # Offload all models
        self.maybe_free_model_hooks()

        if not return_dict:
            return (control_image,)

        return StableDiffusionXLPipelineOutput(images=control_image)