hotshot_xl_controlnet_pipeline.py

# Copyright 2023 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.

# Modifications:
# Copyright 2023 Natural Synthetics Inc. All rights reserved.
# - Adapted the SDXL Controlnet Pipeline to work temporally

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

import numpy as np
import PIL.Image
import torch
import torch.nn.functional as F
from transformers import CLIPTextModel, CLIPTextModelWithProjection, CLIPTokenizer

from hotshot_xl import HotshotPipelineXLOutput

from diffusers.image_processor import VaeImageProcessor
from diffusers.loaders import FromSingleFileMixin, LoraLoaderMixin, TextualInversionLoaderMixin
from diffusers.models import AutoencoderKL, ControlNetModel
from diffusers.models.attention_processor import (
    AttnProcessor2_0,
    LoRAAttnProcessor2_0,
    LoRAXFormersAttnProcessor,
    XFormersAttnProcessor,
)
from diffusers.schedulers import KarrasDiffusionSchedulers
from diffusers.utils import (
    is_accelerate_available,
    is_accelerate_version,
    logging,
    replace_example_docstring,
)
from diffusers.pipelines.pipeline_utils import DiffusionPipeline
from diffusers.utils.torch_utils import randn_tensor, is_compiled_module

from ..models.unet import UNet3DConditionModel

from diffusers.pipelines.controlnet.multicontrolnet import MultiControlNetModel
from einops import rearrange
from tqdm import tqdm

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

def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0):
    """
    Rescale `noise_cfg` according to `guidance_rescale`. Based on findings of [Common Diffusion Noise Schedules and
    Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf). See Section 3.4
    """
    std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True)
    std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True)
    # rescale the results from guidance (fixes overexposure)
    noise_pred_rescaled = noise_cfg * (std_text / std_cfg)
    # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images
    noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg
    return noise_cfg

EXAMPLE_DOC_STRING = """
    Examples:
        ```py
        >>> import torch
        >>> from hotshot_xl import HotshotPipelineXL
        >>> from diffusers import ControlNetModel

        >>> pipe = HotshotXLPipeline.from_pretrained(
        ...     "hotshotco/Hotshot-XL",
        ...     controlnet=ControlNetModel.from_pretrained("diffusers/controlnet-canny-sdxl-1.0")
        ... )
        
        >>> def canny(image):
        >>>     image = cv2.Canny(image, 100, 200)
        >>>     image = image[:, :, None]
        >>>     image = np.concatenate([image, image, image], axis=2)
        >>>     return Image.fromarray(image)
        
        >>> # assuming you have 8 keyframes in current directory...
        
        >>> keyframes = [f"image_{i}.jpg" for i in range(8)]
        >>> control_images = [canny(Image.open(fp)) for fp in keyframes]
        
        >>> pipe = pipe.to("cuda")
        
        >>> prompt = "a photo of an astronaut riding a horse on mars"
        >>> video = pipe(prompt, 
        ...     width=672, height=384, 
        ...     original_size=(1920, 1080), 
        ...     target_size=(512, 512),
        ...     output_type="tensor",
        ...     controlnet_conditioning_scale=0.7,
        ...     control_images=control_images
        ).video
        ```
"""
class HotshotXLControlNetPipeline(
    DiffusionPipeline, TextualInversionLoaderMixin, LoraLoaderMixin, FromSingleFileMixin
):
    r"""
    Pipeline for text-to-image generation using Stable Diffusion XL with ControlNet guidance.

    This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
    implemented for all pipelines (downloading, saving, running on a particular device, etc.).

    The pipeline also inherits the following loading methods:
        - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
        - [`loaders.LoraLoaderMixin.load_lora_weights`] for loading LoRA weights
        - [`loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files

    Args:
        vae ([`AutoencoderKL`]):
            Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations.
        text_encoder ([`~transformers.CLIPTextModel`]):
            Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
        text_encoder_2 ([`~transformers.CLIPTextModelWithProjection`]):
            Second frozen text-encoder
            ([laion/CLIP-ViT-bigG-14-laion2B-39B-b160k](https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k)).
        tokenizer ([`~transformers.CLIPTokenizer`]):
            A `CLIPTokenizer` to tokenize text.
        tokenizer_2 ([`~transformers.CLIPTokenizer`]):
            A `CLIPTokenizer` to tokenize text.
        unet ([`UNet3DConditionModel`]):
            A `UNet3DConditionModel` to denoise the encoded image latents.
        controlnet ([`ControlNetModel`] or `List[ControlNetModel]`):
            Provides additional conditioning to the `unet` during the denoising process. If you set multiple
            ControlNets as a list, the outputs from each ControlNet are added together to create one combined
            additional conditioning.
        scheduler ([`SchedulerMixin`]):
            A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
            [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
        force_zeros_for_empty_prompt (`bool`, *optional*, defaults to `"True"`):
            Whether the negative prompt embeddings should always be set to 0. Also see the config of
            `stabilityai/stable-diffusion-xl-base-1-0`.
        add_watermarker (`bool`, *optional*):
            Whether to use the [invisible_watermark](https://github.com/ShieldMnt/invisible-watermark/) library to
            watermark output images. If not defined, it defaults to `True` if the package is installed; otherwise no
            watermarker is used.
    """

    def __init__(
        self,
        vae: AutoencoderKL,
        text_encoder: CLIPTextModel,
        text_encoder_2: CLIPTextModelWithProjection,
        tokenizer: CLIPTokenizer,
        tokenizer_2: CLIPTokenizer,
        unet: UNet3DConditionModel,
        controlnet: Union[ControlNetModel, List[ControlNetModel], Tuple[ControlNetModel], MultiControlNetModel],
        scheduler: KarrasDiffusionSchedulers,
        force_zeros_for_empty_prompt: bool = True,
        add_watermarker: Optional[bool] = None,
    ):
        super().__init__()

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

        self.register_modules(
            vae=vae,
            text_encoder=text_encoder,
            text_encoder_2=text_encoder_2,
            tokenizer=tokenizer,
            tokenizer_2=tokenizer_2,
            unet=unet,
            controlnet=controlnet,
            scheduler=scheduler,
        )
        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
        self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor, do_convert_rgb=True)
        self.control_image_processor = VaeImageProcessor(
            vae_scale_factor=self.vae_scale_factor, do_convert_rgb=True, do_normalize=False
        )

        self.watermark = None

        self.register_to_config(force_zeros_for_empty_prompt=force_zeros_for_empty_prompt)

    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_slicing
    def enable_vae_slicing(self):
        r"""
        Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to
        compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
        """
        self.vae.enable_slicing()

    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_slicing
    def disable_vae_slicing(self):
        r"""
        Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to
        computing decoding in one step.
        """
        self.vae.disable_slicing()

    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_tiling
    def enable_vae_tiling(self):
        r"""
        Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to
        compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow
        processing larger images.
        """
        self.vae.enable_tiling()

    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_tiling
    def disable_vae_tiling(self):
        r"""
        Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to
        computing decoding in one step.
        """
        self.vae.disable_tiling()

    def enable_model_cpu_offload(self, gpu_id=0):
        r"""
        Offloads all models to CPU using accelerate, reducing memory usage with a low impact on performance. Compared
        to `enable_sequential_cpu_offload`, this method moves one whole model at a time to the GPU when its `forward`
        method is called, and the model remains in GPU until the next model runs. Memory savings are lower than with
        `enable_sequential_cpu_offload`, but performance is much better due to the iterative execution of the `unet`.
        """
        if is_accelerate_available() and is_accelerate_version(">=", "0.17.0.dev0"):
            from accelerate import cpu_offload_with_hook
        else:
            raise ImportError("`enable_model_cpu_offload` requires `accelerate v0.17.0` or higher.")

        device = torch.device(f"cuda:{gpu_id}")

        if self.device.type != "cpu":
            self.to("cpu", silence_dtype_warnings=True)
            torch.cuda.empty_cache()  # otherwise we don't see the memory savings (but they probably exist)

        model_sequence = (
            [self.text_encoder, self.text_encoder_2] if self.text_encoder is not None else [self.text_encoder_2]
        )
        model_sequence.extend([self.unet, self.vae])

        hook = None
        for cpu_offloaded_model in model_sequence:
            _, hook = cpu_offload_with_hook(cpu_offloaded_model, device, prev_module_hook=hook)

        cpu_offload_with_hook(self.controlnet, device)

        # We'll offload the last model manually.
        self.final_offload_hook = hook

    # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.encode_prompt
    def encode_prompt(
        self,
        prompt: str,
        prompt_2: Optional[str] = None,
        device: Optional[torch.device] = None,
        num_images_per_prompt: int = 1,
        do_classifier_free_guidance: bool = True,
        negative_prompt: Optional[str] = None,
        negative_prompt_2: Optional[str] = None,
        prompt_embeds: Optional[torch.FloatTensor] = None,
        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
        pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
        negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
        lora_scale: Optional[float] = None,
    ):
        r"""
        Encodes the prompt into text encoder hidden states.

        Args:
            prompt (`str` or `List[str]`, *optional*):
                prompt to be encoded
            prompt_2 (`str` or `List[str]`, *optional*):
                The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
                used in both text-encoders
            device: (`torch.device`):
                torch device
            num_images_per_prompt (`int`):
                number of images that should be generated per prompt
            do_classifier_free_guidance (`bool`):
                whether to use classifier free guidance or not
            negative_prompt (`str` or `List[str]`, *optional*):
                The prompt or prompts not to guide the image generation. If not defined, one has to pass
                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
                less than `1`).
            negative_prompt_2 (`str` or `List[str]`, *optional*):
                The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and
                `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders
            prompt_embeds (`torch.FloatTensor`, *optional*):
                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
                provided, text embeddings will be generated from `prompt` input argument.
            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
                argument.
            pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
                Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
                If not provided, pooled text embeddings will be generated from `prompt` input argument.
            negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
                Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
                weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt`
                input argument.
            lora_scale (`float`, *optional*):
                A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
        """
        device = device or self._execution_device

        # set lora scale so that monkey patched LoRA
        # function of text encoder can correctly access it
        if lora_scale is not None and isinstance(self, LoraLoaderMixin):
            self._lora_scale = lora_scale

        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]

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

        if prompt_embeds is None:
            prompt_2 = prompt_2 or prompt
            # textual inversion: procecss multi-vector tokens if necessary
            prompt_embeds_list = []
            prompts = [prompt, prompt_2]
            for prompt, tokenizer, text_encoder in zip(prompts, tokenizers, text_encoders):
                if isinstance(self, TextualInversionLoaderMixin):
                    prompt = self.maybe_convert_prompt(prompt, tokenizer)

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

                text_input_ids = text_inputs.input_ids
                untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids

                if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
                    text_input_ids, untruncated_ids
                ):
                    removed_text = tokenizer.batch_decode(untruncated_ids[:, tokenizer.model_max_length - 1 : -1])
                    logger.warning(
                        "The following part of your input was truncated because CLIP can only handle sequences up to"
                        f" {tokenizer.model_max_length} tokens: {removed_text}"
                    )

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

                # We are only ALWAYS interested in the pooled output of the final text encoder
                pooled_prompt_embeds = prompt_embeds[0]
                prompt_embeds = prompt_embeds.hidden_states[-2]

                prompt_embeds_list.append(prompt_embeds)

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

        # get unconditional embeddings for classifier free guidance
        zero_out_negative_prompt = negative_prompt is None and self.config.force_zeros_for_empty_prompt
        if do_classifier_free_guidance and negative_prompt_embeds is None and zero_out_negative_prompt:
            negative_prompt_embeds = torch.zeros_like(prompt_embeds)
            negative_pooled_prompt_embeds = torch.zeros_like(pooled_prompt_embeds)
        elif do_classifier_free_guidance and negative_prompt_embeds is None:
            negative_prompt = negative_prompt or ""
            negative_prompt_2 = negative_prompt_2 or negative_prompt

            uncond_tokens: List[str]
            if prompt is not None and type(prompt) is not type(negative_prompt):
                raise TypeError(
                    f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
                    f" {type(prompt)}."
                )
            elif isinstance(negative_prompt, str):
                uncond_tokens = [negative_prompt, negative_prompt_2]
            elif batch_size != len(negative_prompt):
                raise ValueError(
                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
                    f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
                    " the batch size of `prompt`."
                )
            else:
                uncond_tokens = [negative_prompt, negative_prompt_2]

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

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

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

                negative_prompt_embeds_list.append(negative_prompt_embeds)

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

        prompt_embeds = prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device)
        bs_embed, seq_len, _ = prompt_embeds.shape
        # duplicate text embeddings for each generation per prompt, using mps friendly method
        prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
        prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)

        if do_classifier_free_guidance:
            # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
            seq_len = negative_prompt_embeds.shape[1]
            negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device)
            negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
            negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)

        pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt).view(
            bs_embed * num_images_per_prompt, -1
        )
        if do_classifier_free_guidance:
            negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.repeat(1, num_images_per_prompt).view(
                bs_embed * num_images_per_prompt, -1
            )

        return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds

    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
    def prepare_extra_step_kwargs(self, generator, eta):
        # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
        # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
        # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
        # and should be between [0, 1]

        accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
        extra_step_kwargs = {}
        if accepts_eta:
            extra_step_kwargs["eta"] = eta

        # check if the scheduler accepts generator
        accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
        if accepts_generator:
            extra_step_kwargs["generator"] = generator
        return extra_step_kwargs

    def check_inputs(
        self,
        prompt,
        prompt_2,
        control_images,
        video_length,
        callback_steps,
        negative_prompt=None,
        negative_prompt_2=None,
        prompt_embeds=None,
        negative_prompt_embeds=None,
        pooled_prompt_embeds=None,
        negative_pooled_prompt_embeds=None,
        controlnet_conditioning_scale=1.0,
        control_guidance_start=0.0,
        control_guidance_end=1.0,
    ):
        if (callback_steps is None) or (
            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 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 prompt_embeds is not None and pooled_prompt_embeds is None:
            raise ValueError(
                "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`."
            )

        if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None:
            raise ValueError(
                "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`."
            )

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

        # Check `image`
        is_compiled = hasattr(F, "scaled_dot_product_attention") and isinstance(
            self.controlnet, torch._dynamo.eval_frame.OptimizedModule
        )
        if (
            isinstance(self.controlnet, ControlNetModel)
            or is_compiled
            and isinstance(self.controlnet._orig_mod, ControlNetModel)
        ):

            assert len(control_images) == video_length
            # for image in control_images:
            #     self.check_image(image, prompt, prompt_embeds)
        elif (
            isinstance(self.controlnet, MultiControlNetModel)
            or is_compiled
            and isinstance(self.controlnet._orig_mod, MultiControlNetModel)
        ):
            ...
            # todo
            #
            # if not isinstance(image, list):
            #     raise TypeError("For multiple controlnets: `image` must be type `list`")
            #
            # # When `image` is a nested list:
            # # (e.g. [[canny_image_1, pose_image_1], [canny_image_2, pose_image_2]])
            # elif any(isinstance(i, list) for i in image):
            #     raise ValueError("A single batch of multiple conditionings are supported at the moment.")
            # elif len(image) != len(self.controlnet.nets):
            #     raise ValueError(
            #         f"For multiple controlnets: `image` must have the same length as the number of controlnets, but got {len(image)} images and {len(self.controlnet.nets)} ControlNets."
            #     )
            #
            # for image_ in image:
            #     self.check_image(image_, prompt, prompt_embeds)
        else:
            assert False

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

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

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

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

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

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

    # Copied from diffusers.pipelines.controlnet.pipeline_controlnet.StableDiffusionControlNetPipeline.check_image
    def check_image(self, image, prompt, prompt_embeds):
        image_is_pil = isinstance(image, PIL.Image.Image)
        image_is_tensor = isinstance(image, torch.Tensor)
        image_is_np = isinstance(image, np.ndarray)
        image_is_pil_list = isinstance(image, list) and isinstance(image[0], PIL.Image.Image)
        image_is_tensor_list = isinstance(image, list) and isinstance(image[0], torch.Tensor)
        image_is_np_list = isinstance(image, list) and isinstance(image[0], np.ndarray)

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

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

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

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

    # Copied from diffusers.pipelines.controlnet.pipeline_controlnet.StableDiffusionControlNetPipeline.prepare_image
    def prepare_images(
        self,
        images,
        width,
        height,
        batch_size,
        num_images_per_prompt,
        device,
        dtype,
        do_classifier_free_guidance=False,
        guess_mode=False,
    ):
        images_pre_processed = [self.control_image_processor.preprocess(image, height=height, width=width).to(dtype=torch.float32) for image in images]

        images_pre_processed = torch.cat(images_pre_processed, dim=0)

        repeat_factor = [1] * len(images_pre_processed.shape)
        repeat_factor[0] = batch_size * num_images_per_prompt
        images_pre_processed = images_pre_processed.repeat(*repeat_factor)

        images = images_pre_processed.unsqueeze(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

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

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

        if do_classifier_free_guidance and not guess_mode:
            repeat_factor = [1] * len(images.shape)
            repeat_factor[0] = 2
            images = images.repeat(*repeat_factor)

        return images

    # def prepare_images(self,
    #     images: list,
    #     width,
    #     height,
    #     batch_size,
    #     num_images_per_prompt,
    #     device,
    #     dtype,
    #     do_classifier_free_guidance=False,
    #     guess_mode=False):
    #
    #     images = [self.control_image_processor.preprocess(image, height=height, width=width).to(dtype=torch.float32) for image in images]
    #
    #     image_batch_size = image.shape[0]
    #
    #     if image_batch_size == 1:
    #         repeat_by = batch_size
    #     else:
    #         # image batch size is the same as prompt batch size
    #         repeat_by = num_images_per_prompt
    #
    #     image = image.repeat_interleave(repeat_by, dim=0)
    #
    #     image = image.to(device=device, dtype=dtype)
    #
    #     if do_classifier_free_guidance and not guess_mode:
    #         image = torch.cat([image] * 2)
    #
    #     return image

    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents
    def prepare_latents(self, batch_size, num_channels_latents, video_length, height, width, dtype, device, generator, latents=None):
        #shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
        shape = (batch_size, num_channels_latents, video_length, height // self.vae_scale_factor, width // self.vae_scale_factor)
        if isinstance(generator, list) and len(generator) != batch_size:
            raise ValueError(
                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
            )

        if latents is None:
            latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
        else:
            latents = latents.to(device)

        # scale the initial noise by the standard deviation required by the scheduler
        latents = latents * self.scheduler.init_noise_sigma
        return latents

    # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline._get_add_time_ids
    def _get_add_time_ids(self, original_size, crops_coords_top_left, target_size, dtype):
        add_time_ids = list(original_size + crops_coords_top_left + target_size)

        passed_add_embed_dim = (
            self.unet.config.addition_time_embed_dim * len(add_time_ids) + self.text_encoder_2.config.projection_dim
        )
        expected_add_embed_dim = self.unet.add_embedding.linear_1.in_features

        if expected_add_embed_dim != passed_add_embed_dim:
            raise ValueError(
                f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. The model has an incorrect config. Please check `unet.config.time_embedding_type` and `text_encoder_2.config.projection_dim`."
            )

        add_time_ids = torch.tensor([add_time_ids], dtype=dtype)
        return add_time_ids

    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_upscale.StableDiffusionUpscalePipeline.upcast_vae
    def upcast_vae(self):
        dtype = self.vae.dtype
        self.vae.to(dtype=torch.float32)
        use_torch_2_0_or_xformers = isinstance(
            self.vae.decoder.mid_block.attentions[0].processor,
            (
                AttnProcessor2_0,
                XFormersAttnProcessor,
                LoRAXFormersAttnProcessor,
                LoRAAttnProcessor2_0,
            ),
        )
        # if xformers or torch_2_0 is used attention block does not need
        # to be in float32 which can save lots of memory
        if use_torch_2_0_or_xformers:
            self.vae.post_quant_conv.to(dtype)
            self.vae.decoder.conv_in.to(dtype)
            self.vae.decoder.mid_block.to(dtype)

    @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,
        video_length: Optional[int] = 8,
        control_images: List[PIL.Image.Image] = None,
        height: Optional[int] = None,
        width: Optional[int] = None,
        num_inference_steps: int = 50,
        guidance_scale: float = 5.0,
        negative_prompt: Optional[Union[str, List[str]]] = None,
        negative_prompt_2: Optional[Union[str, List[str]]] = None,
        num_images_per_prompt: Optional[int] = 1,
        eta: float = 0.0,
        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
        latents: Optional[torch.FloatTensor] = None,
        prompt_embeds: Optional[torch.FloatTensor] = None,
        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
        pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
        negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
        output_type: Optional[str] = "pil",
        return_dict: bool = True,
        callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
        callback_steps: int = 1,
        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
        guidance_rescale: float = 0.0,
        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,
        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,
    ):
        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.FloatTensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.FloatTensor]`, `List[PIL.Image.Image]`, `List[np.ndarray]`,:
                    `List[List[torch.FloatTensor]]`, `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.FloatTensor`, 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.
            width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
                The width in pixels of the generated image.
            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.
            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.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 is 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 (prompt weighting). If not
                provided, text embeddings are generated from the `prompt` input argument.
            negative_prompt_embeds (`torch.FloatTensor`, *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.FloatTensor`, *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.FloatTensor`, *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.
            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.
            callback (`Callable`, *optional*):
                A function that calls every `callback_steps` steps during inference. The function is called with the
                following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
            callback_steps (`int`, *optional*, defaults to 1):
                The frequency at which the `callback` function is called. If not specified, the callback is called at
                every step.
            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 `(width, height)` if not specified. Part of SDXL's micro-conditioning as
                explained in section 2.2 of
                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
            crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)):
                `crops_coords_top_left` can be used to generate an image that appears to be "cropped" from the position
                `crops_coords_top_left` downwards. Favorable, well-centered images are usually achieved by setting
                `crops_coords_top_left` to (0, 0). Part of SDXL's micro-conditioning as explained in section 2.2 of
                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
            target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
                For most cases, `target_size` should be set to the desired height and width of the generated image. If
                not specified it will default to `(width, height)`. Part of SDXL's micro-conditioning as explained in
                section 2.2 of [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
            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.

        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.
        """


        if video_length > 1 and num_images_per_prompt > 1:
            print(f"Warning - setting num_images_per_prompt = 1 because video_length = {video_length}")
            num_images_per_prompt = 1

        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_images,
            video_length,
            callback_steps,
            negative_prompt,
            negative_prompt_2,
            prompt_embeds,
            negative_prompt_embeds,
            pooled_prompt_embeds,
            negative_pooled_prompt_embeds,
            controlnet_conditioning_scale,
            control_guidance_start,
            control_guidance_end,
        )

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

        device = self._execution_device
        # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
        # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
        # corresponds to doing no classifier free guidance.
        do_classifier_free_guidance = guidance_scale > 1.0

        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. Encode input prompt
        text_encoder_lora_scale = (
            cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None
        )
        (
            prompt_embeds,
            negative_prompt_embeds,
            pooled_prompt_embeds,
            negative_pooled_prompt_embeds,
        ) = self.encode_prompt(
            prompt,
            prompt_2,
            device,
            num_images_per_prompt,
            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,
        )


        # 4. Prepare image
        if isinstance(controlnet, ControlNetModel):

            assert len(control_images) == video_length * batch_size

            images = self.prepare_images(
                    images=control_images,
                    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=do_classifier_free_guidance,
                    guess_mode=guess_mode,
                )

            height, width = images.shape[-2:]
        elif isinstance(controlnet, MultiControlNetModel):

            raise Exception("not supported yet")

            # images = []
            #
            # for image_ in control_images:
            #     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=do_classifier_free_guidance,
            #         guess_mode=guess_mode,
            #     )
            #
            #     images.append(image_)
            #
            # image = images
            # height, width = image[0].shape[-2:]
        else:
            assert False

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

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

        # 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 isinstance(image, list):
        #     original_size = original_size or image[0].shape[-2:]
        # else:
        original_size = original_size or images.shape[-2:]
        target_size = target_size or (height, width)

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

        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,
            )
        else:
            negative_add_time_ids = add_time_ids

        if do_classifier_free_guidance:
            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
            add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0)
            add_time_ids = torch.cat([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

        images = rearrange(images, "b f c h w -> (b f) c h w")

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

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

                # controlnet(s) inference
                if guess_mode and 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]


                # this will be non interlaced when arranged!
                control_model_input = rearrange(control_model_input, "b c f h w -> (b f) c h w")
                # if we chunked this by 2 - the top 8 frames will be positive for cfg
                # the bottom half will be negative for cfg...

                if video_length > 1:
                    # use repeat_interleave as we need to match the rearrangement above.

                    controlnet_prompt_embeds = controlnet_prompt_embeds.repeat_interleave(video_length, dim=0)
                    controlnet_added_cond_kwargs = {
                        "text_embeds": controlnet_added_cond_kwargs['text_embeds'].repeat_interleave(video_length, dim=0),
                        "time_ids": controlnet_added_cond_kwargs['time_ids'].repeat_interleave(video_length, dim=0)
                    }

                # if type(image) is list:
                #     image = torch.cat(image, dim=0)

                # todo - check if video_length > 1 this needs to produce num_frames * batch_size samples...
                down_block_res_samples, mid_block_res_sample = self.controlnet(
                    control_model_input,
                    t,
                    encoder_hidden_states=controlnet_prompt_embeds,
                    controlnet_cond=images,
                    conditioning_scale=cond_scale,
                    guess_mode=guess_mode,
                    added_cond_kwargs=controlnet_added_cond_kwargs,
                    return_dict=False,
                )

                for j, sample in enumerate(down_block_res_samples):
                    down_block_res_samples[j] = rearrange(sample, "(b f) c h w -> b c f h w", f=video_length)

                mid_block_res_sample = rearrange(mid_block_res_sample, "(b f) c h w -> b c f h w", f=video_length)

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

                # predict the noise residual
                noise_pred = self.unet(
                    latent_model_input,
                    t,
                    encoder_hidden_states=prompt_embeds,
                    cross_attention_kwargs=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,
                    enable_temporal_attentions=video_length > 1
                )[0]

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

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

                # compute the previous noisy sample x_t -> x_t-1
                latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]

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

        # make sure the VAE is in float32 mode, as it overflows in float16
        if self.vae.dtype == torch.float16 and self.vae.config.force_upcast:
            self.upcast_vae()
            latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)

        # If we do sequential model offloading, let's offload unet and controlnet
        # manually for max memory savings
        if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
            self.unet.to("cpu")
            self.controlnet.to("cpu")
            torch.cuda.empty_cache()

        # if not output_type == "latent":
        #     # make sure the VAE is in float32 mode, as it overflows in float16
        #     needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
        #
        #     if needs_upcasting:
        #         self.upcast_vae()
        #         latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
        #
        #     image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
        #
        #     # cast back to fp16 if needed
        #     if needs_upcasting:
        #         self.vae.to(dtype=torch.float16)
        # else:
        #     image = latents
        #     return StableDiffusionXLPipelineOutput(images=image)

        video = self.decode_latents(latents)

        # Convert to tensor
        if output_type == "tensor":
            video = torch.from_numpy(video)

        if not return_dict:
            return video

        return HotshotPipelineXLOutput(videos=video)

    def decode_latents(self, latents):
        video_length = latents.shape[2]
        latents = 1 / self.vae.config.scaling_factor * latents
        latents = rearrange(latents, "b c f h w -> (b f) c h w")
        # video = self.vae.decode(latents).sample
        video = []
        for frame_idx in tqdm(range(latents.shape[0])):
            video.append(self.vae.decode(
                latents[frame_idx:frame_idx+1]).sample)
        video = torch.cat(video)
        video = rearrange(video, "(b f) c h w -> b c f h w", f=video_length)
        video = (video / 2.0 + 0.5).clamp(0, 1)
        # we always cast to float32 as this does not cause significant overhead and is compatible with bfloa16
        video = video.cpu().float().numpy()
        return video

    # Overrride to properly handle the loading and unloading of the additional text encoder.
    # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.load_lora_weights
    def load_lora_weights(self, pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]], **kwargs):
        # We could have accessed the unet config from `lora_state_dict()` too. We pass
        # it here explicitly to be able to tell that it's coming from an SDXL
        # pipeline.
        state_dict, network_alphas = self.lora_state_dict(
            pretrained_model_name_or_path_or_dict,
            unet_config=self.unet.config,
            **kwargs,
        )
        self.load_lora_into_unet(state_dict, network_alphas=network_alphas, unet=self.unet)

        text_encoder_state_dict = {k: v for k, v in state_dict.items() if "text_encoder." in k}
        if len(text_encoder_state_dict) > 0:
            self.load_lora_into_text_encoder(
                text_encoder_state_dict,
                network_alphas=network_alphas,
                text_encoder=self.text_encoder,
                prefix="text_encoder",
                lora_scale=self.lora_scale,
            )

        text_encoder_2_state_dict = {k: v for k, v in state_dict.items() if "text_encoder_2." in k}
        if len(text_encoder_2_state_dict) > 0:
            self.load_lora_into_text_encoder(
                text_encoder_2_state_dict,
                network_alphas=network_alphas,
                text_encoder=self.text_encoder_2,
                prefix="text_encoder_2",
                lora_scale=self.lora_scale,
            )

    @classmethod
    # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.save_lora_weights
    def save_lora_weights(
        self,
        save_directory: Union[str, os.PathLike],
        unet_lora_layers: Dict[str, Union[torch.nn.Module, torch.Tensor]] = None,
        text_encoder_lora_layers: Dict[str, Union[torch.nn.Module, torch.Tensor]] = None,
        text_encoder_2_lora_layers: Dict[str, Union[torch.nn.Module, torch.Tensor]] = None,
        is_main_process: bool = True,
        weight_name: str = None,
        save_function: Callable = None,
        safe_serialization: bool = True,
    ):
        state_dict = {}

        def pack_weights(layers, prefix):
            layers_weights = layers.state_dict() if isinstance(layers, torch.nn.Module) else layers
            layers_state_dict = {f"{prefix}.{module_name}": param for module_name, param in layers_weights.items()}
            return layers_state_dict

        state_dict.update(pack_weights(unet_lora_layers, "unet"))

        if text_encoder_lora_layers and text_encoder_2_lora_layers:
            state_dict.update(pack_weights(text_encoder_lora_layers, "text_encoder"))
            state_dict.update(pack_weights(text_encoder_2_lora_layers, "text_encoder_2"))

        self.write_lora_layers(
            state_dict=state_dict,
            save_directory=save_directory,
            is_main_process=is_main_process,
            weight_name=weight_name,
            save_function=save_function,
            safe_serialization=safe_serialization,
        )

    # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline._remove_text_encoder_monkey_patch
    def _remove_text_encoder_monkey_patch(self):
        self._remove_text_encoder_monkey_patch_classmethod(self.text_encoder)
        self._remove_text_encoder_monkey_patch_classmethod(self.text_encoder_2)