pipelines/pipeline_ccsr.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.


import inspect
import os
import warnings
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 torchvision.utils import save_image
from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer

from diffusers.image_processor import VaeImageProcessor
from diffusers.loaders import TextualInversionLoaderMixin
# from diffusers.models import AutoencoderKL, ControlNetModel, UNet2DConditionModel
from diffusers.models import AutoencoderKL, UNet2DConditionModel
from models.controlnet import ControlNetModel
from diffusers.schedulers import KarrasDiffusionSchedulers
from diffusers.utils import (
    PIL_INTERPOLATION,
    is_accelerate_available,
    is_accelerate_version,
    logging,
    replace_example_docstring,
)
from diffusers.utils.torch_utils import is_compiled_module, randn_tensor

from diffusers.pipelines.pipeline_utils import DiffusionPipeline
from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput
from diffusers.pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker
from diffusers.pipelines.controlnet.multicontrolnet import MultiControlNetModel

from utils.vaehook import VAEHook, perfcount
from tqdm import tqdm
from torch import FloatTensor
from PIL import Image

import time

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


EXAMPLE_DOC_STRING = """
    Examples:
        ```py
        >>> # !pip install opencv-python transformers accelerate
        >>> from diffusers import StableDiffusionControlNetPipeline, ControlNetModel, UniPCMultistepScheduler
        >>> from diffusers.utils import load_image
        >>> import numpy as np
        >>> import torch

        >>> import cv2
        >>> from PIL import Image

        >>> # download an image
        >>> image = load_image(
        ...     "https://hf.co/datasets/huggingface/documentation-images/resolve/main/diffusers/input_image_vermeer.png"
        ... )
        >>> image = np.array(image)

        >>> # get canny image
        >>> image = cv2.Canny(image, 100, 200)
        >>> image = image[:, :, None]
        >>> image = np.concatenate([image, image, image], axis=2)
        >>> canny_image = Image.fromarray(image)

        >>> # load control net and stable diffusion v1-5
        >>> controlnet = ControlNetModel.from_pretrained("lllyasviel/sd-controlnet-canny", torch_dtype=torch.float16)
        >>> pipe = StableDiffusionControlNetPipeline.from_pretrained(
        ...     "runwayml/stable-diffusion-v1-5", controlnet=controlnet, torch_dtype=torch.float16
        ... )

        >>> # speed up diffusion process with faster scheduler and memory optimization
        >>> pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config)
        >>> # remove following line if xformers is not installed
        >>> pipe.enable_xformers_memory_efficient_attention()

        >>> pipe.enable_model_cpu_offload()

        >>> # generate image
        >>> generator = torch.manual_seed(0)
        >>> image = pipe(
        ...     "futuristic-looking woman", num_inference_steps=20, generator=generator, image=canny_image
        ... ).images[0]
        ```
"""


class StableDiffusionControlNetPipeline(DiffusionPipeline, TextualInversionLoaderMixin):
    r"""
    Pipeline for text-to-image generation using Stable Diffusion with ControlNet guidance.

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

    In addition the pipeline inherits the following loading methods:
        - *Textual-Inversion*: [`loaders.TextualInversionLoaderMixin.load_textual_inversion`]

    Args:
        vae ([`AutoencoderKL`]):
            Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
        text_encoder ([`CLIPTextModel`]):
            Frozen text-encoder. Stable Diffusion uses the text portion of
            [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically
            the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant.
        tokenizer (`CLIPTokenizer`):
            Tokenizer of class
            [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
        unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents.
        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`].
        safety_checker ([`StableDiffusionSafetyChecker`]):
            Classification module that estimates whether generated images could be considered offensive or harmful.
            Please, refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for details.
        feature_extractor ([`CLIPImageProcessor`]):
            Model that extracts features from generated images to be used as inputs for the `safety_checker`.
    """
    _optional_components = ["safety_checker", "feature_extractor"]

    def __init__(
        self,
        vae: AutoencoderKL,
        text_encoder: CLIPTextModel,
        tokenizer: CLIPTokenizer,
        unet: UNet2DConditionModel,
        controlnet: Union[ControlNetModel, List[ControlNetModel], Tuple[ControlNetModel], MultiControlNetModel],
        scheduler: KarrasDiffusionSchedulers,
        safety_checker: StableDiffusionSafetyChecker,
        feature_extractor: CLIPImageProcessor,
        requires_safety_checker: bool = True,
    ):
        super().__init__()

        if safety_checker is None and requires_safety_checker:
            logger.warning(
                f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
                " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
                " results in services or applications open to the public. Both the diffusers team and Hugging Face"
                " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
                " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
                " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
            )

        if safety_checker is not None and feature_extractor is None:
            raise ValueError(
                "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety"
                " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead."
            )

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

        self.register_modules(
            vae=vae,
            text_encoder=text_encoder,
            tokenizer=tokenizer,
            unet=unet,
            controlnet=controlnet,
            scheduler=scheduler,
            safety_checker=safety_checker,
            feature_extractor=feature_extractor,
        )
        self.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)
        self.register_to_config(requires_safety_checker=requires_safety_checker)

    def _init_tiled_vae(self,
            encoder_tile_size = 256,
            decoder_tile_size = 256,
            fast_decoder = False,
            fast_encoder = False,
            color_fix = False,
            vae_to_gpu = True):
        # save original forward (only once)
        if not hasattr(self.vae.encoder, 'original_forward'):
            setattr(self.vae.encoder, 'original_forward', self.vae.encoder.forward)
        if not hasattr(self.vae.decoder, 'original_forward'):
            setattr(self.vae.decoder, 'original_forward', self.vae.decoder.forward)

        encoder = self.vae.encoder
        decoder = self.vae.decoder

        self.vae.encoder.forward = VAEHook(
            encoder, encoder_tile_size, is_decoder=False, fast_decoder=fast_decoder, fast_encoder=fast_encoder, color_fix=color_fix, to_gpu=vae_to_gpu)
        self.vae.decoder.forward = VAEHook(
            decoder, decoder_tile_size, is_decoder=True, fast_decoder=fast_decoder, fast_encoder=fast_encoder, color_fix=color_fix, to_gpu=vae_to_gpu)


    # 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 invoked, 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 to save a large amount of memory and to allow the processing of 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 invoked, this method will go back to
        computing decoding in one step.
        """
        self.vae.disable_tiling()

    def enable_sequential_cpu_offload(self, gpu_id=0):
        r"""
        Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet,
        text_encoder, vae, controlnet, and safety checker have their state dicts saved to CPU and then are moved to a
        `torch.device('meta') and loaded to GPU only when their specific submodule has its `forward` method called.
        Note that offloading happens on a submodule basis. Memory savings are higher than with
        `enable_model_cpu_offload`, but performance is lower.
        """
        if is_accelerate_available():
            from accelerate import cpu_offload
        else:
            raise ImportError("Please install accelerate via `pip install accelerate`")

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

        for cpu_offloaded_model in [self.unet, self.text_encoder, self.vae, self.controlnet]:
            cpu_offload(cpu_offloaded_model, device)

        if self.safety_checker is not None:
            cpu_offload(self.safety_checker, execution_device=device, offload_buffers=True)

    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}")

        hook = None
        for cpu_offloaded_model in [self.text_encoder, self.unet, self.vae]:
            _, hook = cpu_offload_with_hook(cpu_offloaded_model, device, prev_module_hook=hook)

        if self.safety_checker is not None:
            # the safety checker can offload the vae again
            _, hook = cpu_offload_with_hook(self.safety_checker, device, prev_module_hook=hook)

        # control net hook has be manually offloaded as it alternates with unet
        cpu_offload_with_hook(self.controlnet, device)

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

    @property
    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._execution_device
    def _execution_device(self):
        r"""
        Returns the device on which the pipeline's models will be executed. After calling
        `pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module
        hooks.
        """
        if not hasattr(self.unet, "_hf_hook"):
            return self.device
        for module in self.unet.modules():
            if (
                hasattr(module, "_hf_hook")
                and hasattr(module._hf_hook, "execution_device")
                and module._hf_hook.execution_device is not None
            ):
                return torch.device(module._hf_hook.execution_device)
        return self.device

    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt
    def _encode_prompt(
        self,
        prompt,
        device,
        num_images_per_prompt,
        do_classifier_free_guidance,
        negative_prompt=None,
        prompt_embeds: Optional[torch.FloatTensor] = None,
        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
    ):
        r"""
        Encodes the prompt into text encoder hidden states.

        Args:
             prompt (`str` or `List[str]`, *optional*):
                prompt to be encoded
            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`).
            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.
        """
        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]

        if prompt_embeds is None:
            # textual inversion: procecss multi-vector tokens if necessary
            if isinstance(self, TextualInversionLoaderMixin):
                prompt = self.maybe_convert_prompt(prompt, self.tokenizer)

            text_inputs = self.tokenizer(
                prompt,
                padding="max_length",
                max_length=self.tokenizer.model_max_length,
                truncation=True,
                return_tensors="pt",
            )
            text_input_ids = text_inputs.input_ids
            untruncated_ids = self.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 = self.tokenizer.batch_decode(
                    untruncated_ids[:, self.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" {self.tokenizer.model_max_length} tokens: {removed_text}"
                )

            if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
                attention_mask = text_inputs.attention_mask.to(device)
            else:
                attention_mask = None

            prompt_embeds = self.text_encoder(
                text_input_ids.to(device),
                attention_mask=attention_mask,
            )
            prompt_embeds = prompt_embeds[0]

        prompt_embeds = prompt_embeds.to(dtype=self.text_encoder.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)

        # get unconditional embeddings for classifier free guidance
        if do_classifier_free_guidance and negative_prompt_embeds is None:
            uncond_tokens: List[str]
            if negative_prompt is None:
                uncond_tokens = [""] * batch_size
            elif 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]
            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

            # textual inversion: procecss multi-vector tokens if necessary
            if isinstance(self, TextualInversionLoaderMixin):
                uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer)

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

            if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
                attention_mask = uncond_input.attention_mask.to(device)
            else:
                attention_mask = None

            negative_prompt_embeds = self.text_encoder(
                uncond_input.input_ids.to(device),
                attention_mask=attention_mask,
            )
            negative_prompt_embeds = negative_prompt_embeds[0]

        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.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)

            # For classifier free guidance, we need to do two forward passes.
            # Here we concatenate the unconditional and text embeddings into a single batch
            # to avoid doing two forward passes
            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])

        return prompt_embeds

    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker
    def run_safety_checker(self, image, device, dtype):
        if self.safety_checker is None:
            has_nsfw_concept = None
        else:
            if torch.is_tensor(image):
                feature_extractor_input = self.image_processor.postprocess(image, output_type="pil")
            else:
                feature_extractor_input = self.image_processor.numpy_to_pil(image)
            safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device)
            image, has_nsfw_concept = self.safety_checker(
                images=image, clip_input=safety_checker_input.pixel_values.to(dtype)
            )
        return image, has_nsfw_concept

    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents
    def decode_latents(self, latents):
        warnings.warn(
            "The decode_latents method is deprecated and will be removed in a future version. Please"
            " use VaeImageProcessor instead",
            FutureWarning,
        )
        latents = 1 / self.vae.config.scaling_factor * latents
        image = self.vae.decode(latents, return_dict=False)[0]
        image = (image / 2 + 0.5).clamp(0, 1)
        # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
        image = image.cpu().permute(0, 2, 3, 1).float().numpy()
        return image

    # 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
        #extra_step_kwargs["generator"] = generator
        return extra_step_kwargs

    def check_inputs(
        self,
        prompt,
        image,
        height,
        width,
        callback_steps,
        negative_prompt=None,
        prompt_embeds=None,
        negative_prompt_embeds=None,
        controlnet_conditioning_scale=1.0,
    ):
        if height % 8 != 0 or width % 8 != 0:
            raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")

        if (callback_steps is 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 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)}")

        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."
            )

        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}."
                )

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

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

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

    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_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)

        if not image_is_pil and not image_is_tensor and not image_is_pil_list and not image_is_tensor_list:
            raise TypeError(
                "image must be passed and be one of PIL image, torch tensor, list of PIL images, or list of torch tensors"
            )

        if image_is_pil:
            image_batch_size = 1
        elif image_is_tensor:
            image_batch_size = image.shape[0]
        elif image_is_pil_list:
            image_batch_size = len(image)
        elif image_is_tensor_list:
            image_batch_size = len(image)

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

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

    def prepare_image(
        self,
        image,
        width,
        height,
        batch_size,
        num_images_per_prompt,
        device,
        dtype,
        do_classifier_free_guidance=False,
        guess_mode=False,
    ):
        if not isinstance(image, torch.Tensor):
            if isinstance(image, PIL.Image.Image):
                image = [image]

            if isinstance(image[0], PIL.Image.Image):
                images = []

                for image_ in image:
                    image_ = image_.convert("RGB")
                    #image_ = image_.resize((width, height), resample=PIL_INTERPOLATION["lanczos"])
                    image_ = np.array(image_)
                    image_ = image_[None, :]
                    images.append(image_)

                image = images

                image = np.concatenate(image, axis=0)
                image = np.array(image).astype(np.float32) / 255.0
                image = image.transpose(0, 3, 1, 2)
                image = torch.from_numpy(image)#.flip(1)
            elif isinstance(image[0], torch.Tensor):
                image = torch.cat(image, 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

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

        if 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

    def _default_height_width(self, height, width, image):
        # NOTE: It is possible that a list of images have different
        # dimensions for each image, so just checking the first image
        # is not _exactly_ correct, but it is simple.
        while isinstance(image, list):
            image = image[0]

        if height is None:
            if isinstance(image, PIL.Image.Image):
                height = image.height
            elif isinstance(image, torch.Tensor):
                height = image.shape[2]

            height = (height // 8) * 8  # round down to nearest multiple of 8

        if width is None:
            if isinstance(image, PIL.Image.Image):
                width = image.width
            elif isinstance(image, torch.Tensor):
                width = image.shape[3]

            width = (width // 8) * 8  # round down to nearest multiple of 8

        return height, width

    # override DiffusionPipeline
    def save_pretrained(
        self,
        save_directory: Union[str, os.PathLike],
        safe_serialization: bool = False,
        variant: Optional[str] = None,
    ):
        if isinstance(self.controlnet, ControlNetModel):
            super().save_pretrained(save_directory, safe_serialization, variant)
        else:
            raise NotImplementedError("Currently, the `save_pretrained()` is not implemented for Multi-ControlNet.")
        
    def previous_timestep(self, timestep):
        if self.scheduler.custom_timesteps:
            index = (self.scheduler.timesteps == timestep).nonzero(as_tuple=True)[0][0]
            if index == self.scheduler.timesteps.shape[0] - 1:
                prev_t = torch.tensor(-1)
            else:
                prev_t = self.scheduler.timesteps[index + 1]
        else:
            num_inference_steps = (
                self.scheduler.num_inference_steps if self.scheduler.num_inference_steps else self.scheduler.config.num_train_timesteps
            )
            prev_t = timestep - self.scheduler.config.num_train_timesteps // num_inference_steps

        return prev_t   

    def predict_start_from_noise(self, sample, t, model_output):
        t = t.to(self.scheduler.alphas_cumprod.device)
        prev_t = self.previous_timestep(t)

        # 1. compute alphas, betas
        alpha_prod_t = self.scheduler.alphas_cumprod[t].to(sample.device)
        alpha_prod_t_prev = self.scheduler.alphas_cumprod[prev_t] if prev_t >= 0 else self.scheduler.one
        alpha_prod_t_prev = alpha_prod_t_prev.to(sample.device)
        beta_prod_t = 1 - alpha_prod_t
        beta_prod_t_prev = 1 - alpha_prod_t_prev
        current_alpha_t = alpha_prod_t / alpha_prod_t_prev
        current_beta_t = 1 - current_alpha_t

        # 2. compute predicted original sample from predicted noise also called
        # "predicted x_0" of formula (15) from https://arxiv.org/pdf/2006.11239.pdf
        if self.scheduler.config.prediction_type == "epsilon":
            pred_original_sample = (sample - beta_prod_t ** (0.5) * model_output) / alpha_prod_t ** (0.5)
        elif self.scheduler.config.prediction_type == "sample":
            pred_original_sample = model_output
        elif self.scheduler.config.prediction_type == "v_prediction":
            pred_original_sample = (alpha_prod_t**0.5) * sample - (beta_prod_t**0.5) * model_output
        else:
            raise ValueError(
                f"prediction_type given as {noise_scheduler.config.prediction_type} must be one of `epsilon`, `sample` or"
                " `v_prediction`  for the DDPMScheduler."
            )

        return pred_original_sample
    
    def _sliding_windows(self,h: int, w: int, tile_size: int, tile_stride: int) -> Tuple[int, int, int, int]:
        hi_list = list(range(0, h - tile_size + 1, tile_stride))
        if (h - tile_size) % tile_stride != 0:
            hi_list.append(h - tile_size)
        
        wi_list = list(range(0, w - tile_size + 1, tile_stride))
        if (w - tile_size) % tile_stride != 0:
            wi_list.append(w - tile_size)
        
        coords = []
        for hi in hi_list:
            for wi in wi_list:
                coords.append((hi, hi + tile_size, wi, wi + tile_size))
        return coords

    # Helper methods within the class
    def _prepare_controlnet_inputs(self, latent_model_input, latents, prompt_embeds, do_classifier_free_guidance, guess_mode):
        if guess_mode and do_classifier_free_guidance:
            return latents, prompt_embeds.chunk(2)[1]
        return latent_model_input, prompt_embeds

    def _predict_noise(self, latent_model_input, t, image, prompt_embeds, cross_attention_kwargs, vae_conditions, tile_diffusion, tile_size, tile_stride, conditioning_scale, guess_mode):
        if not tile_diffusion:
            noise_pred = self._unet_predict(latent_model_input, t, image, prompt_embeds, cross_attention_kwargs, vae_conditions)
        else:
            noise_pred = self._tile_predict(latent_model_input, t, image, prompt_embeds, cross_attention_kwargs, vae_conditions, tile_size, tile_stride, conditioning_scale, guess_mode)
        return noise_pred

    def _unet_predict(self, latent_model_input, t, image, prompt_embeds, cross_attention_kwargs, vae_conditions):
        down_res_samples, mid_res_sample = self.controlnet(
            latent_model_input, t, encoder_hidden_states=prompt_embeds, controlnet_cond=image,
            conditioning_scale=1.0, guess_mode=False,
            return_dict=False, vae_encode_condition_hidden_states=vae_conditions
        )
        noise_pred = self.unet(
            latent_model_input, t, encoder_hidden_states=prompt_embeds,
            cross_attention_kwargs=cross_attention_kwargs,
            down_block_additional_residuals=down_res_samples,
            mid_block_additional_residual=mid_res_sample,
            return_dict=False,
        )[0]
        return noise_pred

    def _tile_predict(self, latent_model_input, t, image, prompt_embeds, cross_attention_kwargs, vae_conditions, tile_size, tile_stride, conditioning_scale, guess_mode):
        tile_weight = self.gaussian_weights(int(tile_size//8), int(tile_size//8), 1).to(latent_model_input.device)
        noise_pred = torch.zeros_like(latent_model_input, dtype=torch.float32)
        count = torch.zeros_like(latent_model_input, dtype=torch.float32)
        h, w = latent_model_input.shape[2:4]

        for hi, hi_end, wi, wi_end in self._sliding_windows(h, w, int(tile_size // 8), int(tile_stride // 8)):
            tile = latent_model_input[:, :, hi:hi_end, wi:wi_end]
            tile_cond = vae_conditions[:, :, hi:hi_end, wi:wi_end] if vae_conditions is not None else None
            tile_image = image[:, :, hi*8:hi_end*8, wi*8:wi_end*8]
            # tile_cond = self.vae.encode(tile_image * 2 - 1).latent_dist.sample() * self.vae.config.scaling_factor
            
            down_block_res_samples, mid_block_res_sample = [None]*10, None
            down_res_samples, mid_res_sample = self.controlnet(
                tile, t, encoder_hidden_states=prompt_embeds, controlnet_cond=tile_image,
                conditioning_scale=1.0, guess_mode=False,
                return_dict=False, vae_encode_condition_hidden_states=tile_cond
            )
            tile_noise = self.unet(
                tile, t, encoder_hidden_states=prompt_embeds,
                cross_attention_kwargs=cross_attention_kwargs,
                down_block_additional_residuals=down_res_samples,
                mid_block_additional_residual=mid_res_sample,
                return_dict=False,
            )[0]
            noise_pred[:, :, hi:hi_end, wi:wi_end] += tile_noise * tile_weight
            count[:, :, hi:hi_end, wi:wi_end] += tile_weight
        noise_pred /= count
        return noise_pred.to(torch.float16)

    def _initial_step(self, do_classifier_free_guidance, latents, t, timesteps, prompt_embeds, image, vae_conditions, tile_diffusion, tile_size, tile_stride):
        if do_classifier_free_guidance:
            prompt_embeds = prompt_embeds.chunk(2)[0]
            image = image.chunk(2)[0]
            vae_conditions = vae_conditions.chunk(2)[0]
        
        noise_pred = self._predict_noise(latents, t, image, prompt_embeds, None, vae_conditions, tile_diffusion, tile_size, tile_stride, 1.0, False)
        x0_T = self.predict_start_from_noise(latents, t, noise_pred)
        noise_tao = torch.randn_like(latents)
        latents = self.scheduler.add_noise(x0_T, noise_tao, timesteps)
        return latents, x0_T

    def _postprocess_latents(self, latents, output_type, do_denormalize):
        latents = latents.to(torch.float16)
        if output_type != "latent":
            image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0].to(torch.float32)
            image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize)
        else:
            image = latents
        return image

    def gaussian_weights(self, tile_width: int, tile_height: int, nbatches: int) -> torch.Tensor:
        """Generates a gaussian mask of weights for tile contributions"""
        from numpy import pi, exp, sqrt
        import numpy as np

        latent_width = tile_width
        latent_height = tile_height

        var = 0.01
        midpoint = (latent_width - 1) / 2  # -1 because index goes from 0 to latent_width - 1
        x_probs = [exp(-(x-midpoint)*(x-midpoint)/(latent_width*latent_width)/(2*var)) / sqrt(2*pi*var) for x in range(latent_width)]
        midpoint = latent_height / 2
        y_probs = [exp(-(y-midpoint)*(y-midpoint)/(latent_height*latent_height)/(2*var)) / sqrt(2*pi*var) for y in range(latent_height)]

        weights = np.outer(y_probs, x_probs)
        return torch.tile(torch.tensor(weights, device=self.device), (nbatches, 4, 1, 1))
    
    @torch.no_grad()
    @replace_example_docstring(EXAMPLE_DOC_STRING)
    def __call__(
        self,
        t_max: float,
        t_min: float,
        tile_diffusion: bool, 
        tile_size: float,
        tile_stride: float,
        prompt: Union[str, List[str]] = None,
        image: Union[FloatTensor, Image.Image, List[FloatTensor], List[Image.Image]] = None,
        height: Optional[int] = None,
        width: Optional[int] = None,
        num_inference_steps: int = 50,
        guidance_scale: float = 7.5,
        negative_prompt: 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[FloatTensor] = None,
        prompt_embeds: Optional[FloatTensor] = None,
        negative_prompt_embeds: Optional[FloatTensor] = None,
        output_type: Optional[str] = "pil",
        return_dict: bool = True,
        callback: Optional[Callable[[int, int, FloatTensor], None]] = None,
        callback_steps: int = 1,
        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
        conditioning_scale: Union[float, List[float]] = 1.0,
        guess_mode: bool = False,
        start_steps: int = 999,
        use_vae_encode_condition: bool = False,
        start_point: str = 'noise',
    ) -> Union[StableDiffusionPipelineOutput, tuple]:
        r"""
        Optimized diffusion pipeline call for image super-resolution.
        For 'Improving the Stability and Efficiency of Diffusion Models for Content Consistent Super-Resolution'.

        Examples:
            # Example usage:
            # pipeline(t_max=0.6667, t_min=0.5, tile_diffusion=True, tile_size=256, tile_stride=128, prompt="", num_inference_steps=6)
            pass
        """
        # 0. Set default height and width
        height, width = self._default_height_width(height, width, image)
        
        # 1. Determine batch size
        if prompt is not None:
            batch_size = 1 if isinstance(prompt, str) else len(prompt)
        else:
            batch_size = prompt_embeds.shape[0]
        
        device = self._execution_device
        do_classifier_free_guidance = guidance_scale > 1.0
        
        controlnet = self.controlnet._orig_mod if is_compiled_module(self.controlnet) else self.controlnet

        # 2. Prepare image
        image = self.prepare_image(
            image=image, width=width, height=height,
            batch_size=batch_size * num_images_per_prompt,
            num_images_per_prompt=num_images_per_prompt,
            device=device, dtype=controlnet.dtype,
            do_classifier_free_guidance=do_classifier_free_guidance,
            guess_mode=guess_mode
        )

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

        # 4. Prepare extra step kwargs
        extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
        
        ### calculate the running time for each inference step
        torch.cuda.synchronize()
        start_time = time.time()
        # 5. Encode prompts
        prompt_embeds = self._encode_prompt(
            prompt, device, num_images_per_prompt, do_classifier_free_guidance,
            negative_prompt, prompt_embeds=prompt_embeds,
            negative_prompt_embeds=negative_prompt_embeds
        ) 

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

        # 7. Initialize latent variables based on start_point
        latents_condition_image = self.vae.encode(image * 2 - 1).latent_dist.sample() * self.vae.config.scaling_factor
        if start_point != 'noise':
            start_steps_tensor = torch.randint(start_steps, start_steps + 1, (latents.shape[0],), device=latents.device).long()
            latents = self.scheduler.add_noise(latents_condition_image[0:1, ...], latents, start_steps_tensor)

        # 8. Optionally prepare VAE-encoded condition
        vae_encode_condition_hidden_states = (
            latents_condition_image if use_vae_encode_condition else None
        )

        # 9. Initial prediction at t_max if needed
        total_steps = len(timesteps)
        t_tao = timesteps[-round(total_steps * t_max)]
          
        
        if t_max != 1:
            t = torch.randint(start_steps, start_steps+1, (batch_size,), device=latents.device)
            latents = latents.to(torch.float16)
            # we do not do the classifier free guidance in this step
            latent_model_input = self.scheduler.scale_model_input(latents, t)
            latents, x0_T = self._initial_step(do_classifier_free_guidance, latent_model_input, t, t_tao, prompt_embeds, image, vae_encode_condition_hidden_states, tile_diffusion, tile_size, tile_stride)
        # redefine timesteps
        timesteps = timesteps[-round(total_steps * t_max):]
        timesteps = timesteps[:-round(total_steps * t_min)] if t_min > 0 else timesteps
        
        # 10. Denoising loop
        if num_inference_steps==1:
            latents = x0_T
        else:
            with self.progress_bar(total=len(timesteps)) as progress_bar:
                for i, t in enumerate(timesteps):
                    latents = latents.to(torch.float16)
                    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)
            
                    controlnet_latent_model_input, controlnet_prompt_embeds = self._prepare_controlnet_inputs(latent_model_input, latents, prompt_embeds, do_classifier_free_guidance, guess_mode)
            
                    noise_pred = self._predict_noise(
                        controlnet_latent_model_input, t, image, controlnet_prompt_embeds, cross_attention_kwargs, 
                        vae_encode_condition_hidden_states, tile_diffusion, tile_size, tile_stride, conditioning_scale, guess_mode
                    )
            
                    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)

                    latents_old = latents
                    latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
                    

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

        # Predict x0 for t_min
        if t_min:
            x0_tmin = self.predict_start_from_noise(latents_old, t, noise_pred)
            latents = x0_tmin

        # 11. Post-processing
        has_nsfw_concept = None
        if has_nsfw_concept is None:
            do_denormalize = [True] * image.shape[0]
        else:
            do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept]

        image = self._postprocess_latents(latents, output_type, do_denormalize)
        
        ## cauculate the inference time for each inference step
        torch.cuda.synchronize()
        end_time = time.time()
        total_time = end_time - start_time
        
        return total_time, StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)