mvadapter/pipelines/pipeline_mvadapter_i2mv_sdxl.py

# Copyright 2024 The HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

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

import numpy as np
import PIL
import torch
import torch.nn as nn
from diffusers.image_processor import PipelineImageInput, VaeImageProcessor
from diffusers.models import (
    AutoencoderKL,
    ImageProjection,
    T2IAdapter,
    UNet2DConditionModel,
)
from diffusers.pipelines.stable_diffusion_xl.pipeline_output import (
    StableDiffusionXLPipelineOutput,
)
from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl import (
    StableDiffusionXLPipeline,
    rescale_noise_cfg,
    retrieve_timesteps,
)
from diffusers.schedulers import KarrasDiffusionSchedulers
from diffusers.utils import deprecate, logging
from diffusers.utils.torch_utils import randn_tensor
from einops import rearrange
from transformers import (
    CLIPImageProcessor,
    CLIPTextModel,
    CLIPTextModelWithProjection,
    CLIPTokenizer,
    CLIPVisionModelWithProjection,
)

from ..loaders import CustomAdapterMixin
from ..models.attention_processor import (
    DecoupledMVRowSelfAttnProcessor2_0,
    set_unet_2d_condition_attn_processor,
)

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


def retrieve_latents(
    encoder_output: torch.Tensor,
    generator: Optional[torch.Generator] = None,
    sample_mode: str = "sample",
):
    if hasattr(encoder_output, "latent_dist") and sample_mode == "sample":
        return encoder_output.latent_dist.sample(generator)
    elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax":
        return encoder_output.latent_dist.mode()
    elif hasattr(encoder_output, "latents"):
        return encoder_output.latents
    else:
        raise AttributeError("Could not access latents of provided encoder_output")


class MVAdapterI2MVSDXLPipeline(StableDiffusionXLPipeline, CustomAdapterMixin):
    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,
        force_zeros_for_empty_prompt: bool = True,
        add_watermarker: Optional[bool] = None,
    ):
        super().__init__(
            vae=vae,
            text_encoder=text_encoder,
            text_encoder_2=text_encoder_2,
            tokenizer=tokenizer,
            tokenizer_2=tokenizer_2,
            unet=unet,
            scheduler=scheduler,
            image_encoder=image_encoder,
            feature_extractor=feature_extractor,
            force_zeros_for_empty_prompt=force_zeros_for_empty_prompt,
            add_watermarker=add_watermarker,
        )

        self.control_image_processor = VaeImageProcessor(
            vae_scale_factor=self.vae_scale_factor,
            do_convert_rgb=True,
            do_normalize=False,
        )

    # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl_img2img.prepare_latents
    def prepare_image_latents(
        self,
        image,
        timestep,
        batch_size,
        num_images_per_prompt,
        dtype,
        device,
        generator=None,
        add_noise=True,
    ):
        if not isinstance(image, (torch.Tensor, PIL.Image.Image, list)):
            raise ValueError(
                f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}"
            )

        latents_mean = latents_std = None
        if (
            hasattr(self.vae.config, "latents_mean")
            and self.vae.config.latents_mean is not None
        ):
            latents_mean = torch.tensor(self.vae.config.latents_mean).view(1, 4, 1, 1)
        if (
            hasattr(self.vae.config, "latents_std")
            and self.vae.config.latents_std is not None
        ):
            latents_std = torch.tensor(self.vae.config.latents_std).view(1, 4, 1, 1)

        # Offload text encoder if `enable_model_cpu_offload` was enabled
        if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
            self.text_encoder_2.to("cpu")
            torch.cuda.empty_cache()

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

        batch_size = batch_size * num_images_per_prompt

        if image.shape[1] == 4:
            init_latents = image

        else:
            # make sure the VAE is in float32 mode, as it overflows in float16
            if self.vae.config.force_upcast:
                image = image.float()
                self.vae.to(dtype=torch.float32)

            if isinstance(generator, list) and len(generator) != batch_size:
                raise ValueError(
                    f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
                    f" size of {batch_size}. Make sure the batch size matches the length of the generators."
                )

            elif isinstance(generator, list):
                if image.shape[0] < batch_size and batch_size % image.shape[0] == 0:
                    image = torch.cat([image] * (batch_size // image.shape[0]), dim=0)
                elif image.shape[0] < batch_size and batch_size % image.shape[0] != 0:
                    raise ValueError(
                        f"Cannot duplicate `image` of batch size {image.shape[0]} to effective batch_size {batch_size} "
                    )

                init_latents = [
                    retrieve_latents(
                        self.vae.encode(image[i : i + 1]), generator=generator[i]
                    )
                    for i in range(batch_size)
                ]
                init_latents = torch.cat(init_latents, dim=0)
            else:
                init_latents = retrieve_latents(
                    self.vae.encode(image), generator=generator
                )

            if self.vae.config.force_upcast:
                self.vae.to(dtype)

            init_latents = init_latents.to(dtype)
            if latents_mean is not None and latents_std is not None:
                latents_mean = latents_mean.to(device=device, dtype=dtype)
                latents_std = latents_std.to(device=device, dtype=dtype)
                init_latents = (
                    (init_latents - latents_mean)
                    * self.vae.config.scaling_factor
                    / latents_std
                )
            else:
                init_latents = self.vae.config.scaling_factor * init_latents

        if (
            batch_size > init_latents.shape[0]
            and batch_size % init_latents.shape[0] == 0
        ):
            # expand init_latents for batch_size
            additional_image_per_prompt = batch_size // init_latents.shape[0]
            init_latents = torch.cat(
                [init_latents] * additional_image_per_prompt, dim=0
            )
        elif (
            batch_size > init_latents.shape[0]
            and batch_size % init_latents.shape[0] != 0
        ):
            raise ValueError(
                f"Cannot duplicate `image` of batch size {init_latents.shape[0]} to {batch_size} text prompts."
            )
        else:
            init_latents = torch.cat([init_latents], dim=0)

        if add_noise:
            shape = init_latents.shape
            noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
            # get latents
            init_latents = self.scheduler.add_noise(init_latents, noise, timestep)

        latents = init_latents

        return latents

    def prepare_control_image(
        self,
        image,
        width,
        height,
        batch_size,
        num_images_per_prompt,
        device,
        dtype,
        do_classifier_free_guidance=False,
        num_empty_images=0,  # for concat in batch like ImageDream
    ):
        assert hasattr(
            self, "control_image_processor"
        ), "control_image_processor is not initialized"

        image = self.control_image_processor.preprocess(
            image, height=height, width=width
        ).to(dtype=torch.float32)

        if num_empty_images > 0:
            image = torch.cat(
                [image, torch.zeros_like(image[:num_empty_images])], dim=0
            )

        image_batch_size = image.shape[0]

        if image_batch_size == 1:
            repeat_by = batch_size
        else:
            # image batch size is the same as prompt batch size
            repeat_by = num_images_per_prompt  # always 1 for control image

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

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

        if do_classifier_free_guidance:
            image = torch.cat([image] * 2)

        return image

    @torch.no_grad()
    def __call__(
        self,
        prompt: Union[str, List[str]] = None,
        prompt_2: Optional[Union[str, List[str]]] = None,
        height: Optional[int] = None,
        width: Optional[int] = None,
        num_inference_steps: int = 50,
        timesteps: List[int] = None,
        denoising_end: Optional[float] = None,
        guidance_scale: float = 5.0,
        negative_prompt: Optional[Union[str, List[str]]] = None,
        negative_prompt_2: Optional[Union[str, List[str]]] = None,
        num_images_per_prompt: Optional[int] = 1,
        eta: float = 0.0,
        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
        latents: Optional[torch.FloatTensor] = None,
        prompt_embeds: Optional[torch.FloatTensor] = None,
        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
        pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
        negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
        ip_adapter_image: Optional[PipelineImageInput] = None,
        ip_adapter_image_embeds: Optional[List[torch.FloatTensor]] = 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: Optional[Tuple[int, int]] = None,
        crops_coords_top_left: Tuple[int, int] = (0, 0),
        target_size: Optional[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[Callable[[int, int, Dict], None]] = None,
        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
        # NEW
        mv_scale: float = 1.0,
        # Camera or geometry condition
        control_image: Optional[PipelineImageInput] = None,
        control_conditioning_scale: Optional[float] = 1.0,
        control_conditioning_factor: float = 1.0,
        # Image condition
        reference_image: Optional[PipelineImageInput] = None,
        reference_conditioning_scale: Optional[float] = 1.0,
        **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
            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.
            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.
            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):
                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
            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` or `List[torch.Generator]`, *optional*):
                One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
                to make generation deterministic.
            latents (`torch.FloatTensor`, *optional*):
                Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
                tensor will ge generated by sampling using the supplied random `generator`.
            prompt_embeds (`torch.FloatTensor`, *optional*):
                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
                provided, text embeddings will be generated from `prompt` input argument.
            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
                argument.
            pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
                Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
                If not provided, pooled text embeddings will be generated from `prompt` input argument.
            negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
                Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
                weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt`
                input argument.
            ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters.
            ip_adapter_image_embeds (`List[torch.FloatTensor]`, *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 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_xl.StableDiffusionXLPipelineOutput`] 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).
            guidance_rescale (`float`, *optional*, defaults to 0.0):
                Guidance rescale factor proposed by [Common Diffusion Noise Schedules and Sample Steps are
                Flawed](https://arxiv.org/pdf/2305.08891.pdf) `guidance_scale` is defined as `φ` in equation 16. of
                [Common Diffusion Noise Schedules and Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf).
                Guidance rescale factor should fix overexposure when using zero terminal SNR.
            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.
            callback_on_step_end (`Callable`, *optional*):
                A function that calls at the end of each denoising steps during the inference. The function is called
                with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
                callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
                `callback_on_step_end_tensor_inputs`.
            callback_on_step_end_tensor_inputs (`List`, *optional*):
                The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
                will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
                `._callback_tensor_inputs` attribute of your pipeline class.

        Examples:

        Returns:
            [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] or `tuple`:
            [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] if `return_dict` is True, otherwise a
            `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`",
            )

        # 0. Default height and width to unet
        height = height or self.default_sample_size * self.vae_scale_factor
        width = width or self.default_sample_size * self.vae_scale_factor

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

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

        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._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
        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=lora_scale,
            clip_skip=self.clip_skip,
        )

        # 4. Prepare timesteps
        timesteps, num_inference_steps = retrieve_timesteps(
            self.scheduler, num_inference_steps, device, timesteps
        )

        # 5. 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. 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. Prepare added time ids & embeddings
        add_text_embeds = pooled_prompt_embeds
        if self.text_encoder_2 is None:
            text_encoder_projection_dim = int(pooled_prompt_embeds.shape[-1])
        else:
            text_encoder_projection_dim = self.text_encoder_2.config.projection_dim

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

        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,
            )

        # Preprocess reference image
        reference_image = self.image_processor.preprocess(reference_image)
        reference_latents = self.prepare_image_latents(
            reference_image,
            timesteps[:1].repeat(batch_size * num_images_per_prompt),  # no use
            batch_size,
            1,
            prompt_embeds.dtype,
            device,
            generator,
            add_noise=False,
        )

        with torch.no_grad():
            ref_timesteps = torch.zeros_like(timesteps[0])
            ref_hidden_states = {}

            self.unet(
                reference_latents,
                ref_timesteps,
                encoder_hidden_states=prompt_embeds[-1:],
                added_cond_kwargs={
                    "text_embeds": add_text_embeds[-1:],
                    "time_ids": add_time_ids[-1:],
                },
                cross_attention_kwargs={
                    "cache_hidden_states": ref_hidden_states,
                    "use_mv": False,
                    "use_ref": False,
                },
                return_dict=False,
            )
            ref_hidden_states = {
                k: v.repeat_interleave(num_images_per_prompt, dim=0)
                for k, v in ref_hidden_states.items()
            }
        if self.do_classifier_free_guidance:
            ref_hidden_states = {
                k: torch.cat([torch.zeros_like(v), v], dim=0)
                for k, v in ref_hidden_states.items()
            }

        cross_attention_kwargs = {
            "mv_scale": mv_scale,
            "ref_hidden_states": {k: v.clone() for k, v in ref_hidden_states.items()},
            "ref_scale": reference_conditioning_scale,
            **(self.cross_attention_kwargs or {}),
        }

        # Preprocess control image
        control_image_feature = self.prepare_control_image(
            image=control_image,
            width=width,
            height=height,
            batch_size=batch_size * num_images_per_prompt,
            num_images_per_prompt=1,  # NOTE: always 1 for control images
            device=device,
            dtype=latents.dtype,
            do_classifier_free_guidance=self.do_classifier_free_guidance,
        )
        control_image_feature = control_image_feature.to(
            device=device, dtype=latents.dtype
        )

        adapter_state = self.cond_encoder(control_image_feature)
        for i, state in enumerate(adapter_state):
            adapter_state[i] = state * control_conditioning_scale

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

        # 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]

        # 9. 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
                )

                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,
                }
                if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
                    added_cond_kwargs["image_embeds"] = image_embeds

                if i < int(num_inference_steps * control_conditioning_factor):
                    down_intrablock_additional_residuals = [
                        state.clone() for state in adapter_state
                    ]
                else:
                    down_intrablock_additional_residuals = None

                # predict the noise residual
                noise_pred = self.unet(
                    latent_model_input,
                    t,
                    encoder_hidden_states=prompt_embeds,
                    timestep_cond=timestep_cond,
                    cross_attention_kwargs=cross_attention_kwargs,
                    down_intrablock_additional_residuals=down_intrablock_additional_residuals,
                    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 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
                )
            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]

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

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

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

        # Offload all models
        self.maybe_free_model_hooks()

        if not return_dict:
            return (image,)

        return StableDiffusionXLPipelineOutput(images=image)

    ### NEW: adapters ###
    def _init_custom_adapter(
        self,
        # Multi-view adapter
        num_views: int,
        self_attn_processor: Any = DecoupledMVRowSelfAttnProcessor2_0,
        # Condition encoder
        cond_in_channels: int = 6,
        # For training
        copy_attn_weights: bool = True,
        zero_init_module_keys: List[str] = [],
    ):
        # Condition encoder
        self.cond_encoder = T2IAdapter(
            in_channels=cond_in_channels,
            channels=(320, 640, 1280, 1280),
            num_res_blocks=2,
            downscale_factor=16,
            adapter_type="full_adapter_xl",
        )

        # set custom attn processor for multi-view attention and image cross-attention
        self.unet: UNet2DConditionModel
        set_unet_2d_condition_attn_processor(
            self.unet,
            set_self_attn_proc_func=lambda name, hs, cad, ap: self_attn_processor(
                query_dim=hs,
                inner_dim=hs,
                num_views=num_views,
                name=name,
                use_mv=True,
                use_ref=True,
            ),
        )

        # copy decoupled attention weights from original unet
        if copy_attn_weights:
            state_dict = self.unet.state_dict()
            for key in state_dict.keys():
                if "_mv" in key:
                    compatible_key = key.replace("_mv", "").replace("processor.", "")
                elif "_ref" in key:
                    compatible_key = key.replace("_ref", "").replace("processor.", "")
                else:
                    compatible_key = key

                is_zero_init_key = any([k in key for k in zero_init_module_keys])
                if is_zero_init_key:
                    state_dict[key] = torch.zeros_like(state_dict[compatible_key])
                else:
                    state_dict[key] = state_dict[compatible_key].clone()
            self.unet.load_state_dict(state_dict)

    def _load_custom_adapter(self, state_dict):
        self.unet.load_state_dict(state_dict, strict=False)
        self.cond_encoder.load_state_dict(state_dict, strict=False)

    def _save_custom_adapter(
        self,
        include_keys: Optional[List[str]] = None,
        exclude_keys: Optional[List[str]] = None,
    ):
        def include_fn(k):
            is_included = False

            if include_keys is not None:
                is_included = is_included or any([key in k for key in include_keys])
            if exclude_keys is not None:
                is_included = is_included and not any(
                    [key in k for key in exclude_keys]
                )

            return is_included

        state_dict = {k: v for k, v in self.unet.state_dict().items() if include_fn(k)}
        state_dict.update(self.cond_encoder.state_dict())

        return state_dict