pipeline_stable_diffusion_latent_guided_img2img.py

# Copyright 2022 The HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import inspect
from typing import Callable, List, Optional, Union

import numpy as np
import torch

import PIL
from diffusers.utils import is_accelerate_available
from packaging import version
from transformers import CLIPFeatureExtractor, CLIPTextModel, CLIPTokenizer

from diffusers.configuration_utils import FrozenDict
from diffusers.models import AutoencoderKL, UNet2DConditionModel
from diffusers.pipeline_utils import DiffusionPipeline
from diffusers.schedulers import (
    DDIMScheduler,
    DPMSolverMultistepScheduler,
    EulerAncestralDiscreteScheduler,
    EulerDiscreteScheduler,
    LMSDiscreteScheduler,
    PNDMScheduler,
)
from diffusers.utils import PIL_INTERPOLATION, deprecate, logging
from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput
from diffusers.pipelines.stable_diffusion import StableDiffusionSafetyChecker

from torchvision import transforms as tfms

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


def preprocess(image):
    if isinstance(image, torch.Tensor):
        return image
    elif isinstance(image, PIL.Image.Image):
        image = [image]

    if isinstance(image[0], PIL.Image.Image):
        w, h = image[0].size
        w, h = map(lambda x: x - x % 32, (w, h))  # resize to integer multiple of 32

        image = [
            np.array(i.resize((w, h), resample=PIL_INTERPOLATION["lanczos"]))[None, :]
            for i in image
        ]
        image = np.concatenate(image, axis=0)
        image = np.array(image).astype(np.float32) / 255.0
        image = image.transpose(0, 3, 1, 2)
        image = 2.0 * image - 1.0
        image = torch.from_numpy(image)
    elif isinstance(image[0], torch.Tensor):
        image = torch.cat(image, dim=0)
    return image


class StableDiffusionImg2ImgPipeline(DiffusionPipeline):
    r"""
    Pipeline for text-guided image to image generation using Stable Diffusion.

    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.)

    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.
        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 ([`CLIPFeatureExtractor`]):
            Model that extracts features from generated images to be used as inputs for the `safety_checker`.
    """
    _optional_components = ["safety_checker", "feature_extractor"]

    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.__init__
    def __init__(
        self,
        vae: AutoencoderKL,
        text_encoder: CLIPTextModel,
        tokenizer: CLIPTokenizer,
        unet: UNet2DConditionModel,
        scheduler: Union[
            DDIMScheduler,
            PNDMScheduler,
            LMSDiscreteScheduler,
            EulerDiscreteScheduler,
            EulerAncestralDiscreteScheduler,
            DPMSolverMultistepScheduler,
        ],
        safety_checker: StableDiffusionSafetyChecker,
        feature_extractor: CLIPFeatureExtractor,
        requires_safety_checker: bool = True,
    ):
        super().__init__()

        if (
            hasattr(scheduler.config, "steps_offset")
            and scheduler.config.steps_offset != 1
        ):
            deprecation_message = (
                f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`"
                f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure "
                "to update the config accordingly as leaving `steps_offset` might led to incorrect results"
                " in future versions. If you have downloaded this checkpoint from the Hugging Face Hub,"
                " it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`"
                " file"
            )
            deprecate(
                "steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False
            )
            new_config = dict(scheduler.config)
            new_config["steps_offset"] = 1
            scheduler._internal_dict = FrozenDict(new_config)

        if (
            hasattr(scheduler.config, "clip_sample")
            and scheduler.config.clip_sample is True
        ):
            deprecation_message = (
                f"The configuration file of this scheduler: {scheduler} has not set the configuration `clip_sample`."
                " `clip_sample` should be set to False in the configuration file. Please make sure to update the"
                " config accordingly as not setting `clip_sample` in the config might lead to incorrect results in"
                " future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it would be very"
                " nice if you could open a Pull request for the `scheduler/scheduler_config.json` file"
            )
            deprecate(
                "clip_sample not set", "1.0.0", deprecation_message, standard_warn=False
            )
            new_config = dict(scheduler.config)
            new_config["clip_sample"] = False
            scheduler._internal_dict = FrozenDict(new_config)

        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."
            )

        is_unet_version_less_0_9_0 = hasattr(
            unet.config, "_diffusers_version"
        ) and version.parse(
            version.parse(unet.config._diffusers_version).base_version
        ) < version.parse(
            "0.9.0.dev0"
        )
        is_unet_sample_size_less_64 = (
            hasattr(unet.config, "sample_size") and unet.config.sample_size < 64
        )
        if is_unet_version_less_0_9_0 and is_unet_sample_size_less_64:
            deprecation_message = (
                "The configuration file of the unet has set the default `sample_size` to smaller than"
                " 64 which seems highly unlikely. If your checkpoint is a fine-tuned version of any of the"
                " following: \n- CompVis/stable-diffusion-v1-4 \n- CompVis/stable-diffusion-v1-3 \n-"
                " CompVis/stable-diffusion-v1-2 \n- CompVis/stable-diffusion-v1-1 \n- runwayml/stable-diffusion-v1-5"
                " \n- runwayml/stable-diffusion-inpainting \n you should change 'sample_size' to 64 in the"
                " configuration file. Please make sure to update the config accordingly as leaving `sample_size=32`"
                " in the config might lead to incorrect results in future versions. If you have downloaded this"
                " checkpoint from the Hugging Face Hub, it would be very nice if you could open a Pull request for"
                " the `unet/config.json` file"
            )
            deprecate(
                "sample_size<64", "1.0.0", deprecation_message, standard_warn=False
            )
            new_config = dict(unet.config)
            new_config["sample_size"] = 64
            unet._internal_dict = FrozenDict(new_config)

        self.register_modules(
            vae=vae,
            text_encoder=text_encoder,
            tokenizer=tokenizer,
            unet=unet,
            scheduler=scheduler,
            safety_checker=safety_checker,
            feature_extractor=feature_extractor,
        )
        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
        self.register_to_config(requires_safety_checker=requires_safety_checker)

    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_sequential_cpu_offload
    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 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.
        """
        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]:
            if cpu_offloaded_model is not None:
                cpu_offload(cpu_offloaded_model, device)

        if self.safety_checker is not None:
            # TODO(Patrick) - there is currently a bug with cpu offload of nn.Parameter in accelerate
            # fix by only offloading self.safety_checker for now
            cpu_offload(self.safety_checker.vision_model, device)

    @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 self.device != torch.device("meta") or 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,
    ):
        r"""
        Encodes the prompt into text encoder hidden states.

        Args:
            prompt (`str` or `list(int)`):
                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]`):
                The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
                if `guidance_scale` is less than `1`).
        """
        batch_size = len(prompt) if isinstance(prompt, list) else 1

        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

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

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

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

            max_length = text_input_ids.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

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

            # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
            seq_len = uncond_embeddings.shape[1]
            uncond_embeddings = uncond_embeddings.repeat(1, num_images_per_prompt, 1)
            uncond_embeddings = uncond_embeddings.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
            text_embeddings = torch.cat([uncond_embeddings, text_embeddings])

        return text_embeddings

    # 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 not None:
            safety_checker_input = self.feature_extractor(
                self.numpy_to_pil(image), return_tensors="pt"
            ).to(device)
            image, has_nsfw_concept = self.safety_checker(
                images=image, clip_input=safety_checker_input.pixel_values.to(dtype)
            )
        else:
            has_nsfw_concept = None
        return image, has_nsfw_concept

    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents
    def decode_latents(self, latents):
        latents = 1 / 0.18215 * latents
        image = self.vae.decode(latents).sample
        image = (image / 2 + 0.5).clamp(0, 1)
        # we always cast to float32 as this does not cause significant overhead and is compatible with bfloa16
        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
        return extra_step_kwargs

    def check_inputs(self, prompt, strength, callback_steps):
        if 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 strength < 0 or strength > 1:
            raise ValueError(
                f"The value of strength should in [1.0, 1.0] but is {strength}"
            )

        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)}."
            )

    def get_timesteps(self, num_inference_steps, strength, device):
        # get the original timestep using init_timestep
        init_timestep = min(int(num_inference_steps * strength), num_inference_steps)

        t_start = max(num_inference_steps - init_timestep, 0)
        timesteps = self.scheduler.timesteps[t_start:]

        return timesteps, num_inference_steps - t_start

    def prepare_latents(
        self,
        image,
        timestep,
        batch_size,
        num_images_per_prompt,
        dtype,
        device,
        generator=None,
    ):
        image = image.to(device=device, dtype=dtype)

        batch_size = batch_size * num_images_per_prompt
        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 isinstance(generator, list):
            init_latents = [
                self.vae.encode(image[i : i + 1]).latent_dist.sample(generator[i])
                for i in range(batch_size)
            ]
            init_latents = torch.cat(init_latents, dim=0)
        else:
            init_latents = self.vae.encode(image).latent_dist.sample(generator)

        init_latents = 0.18215 * init_latents

        if (
            batch_size > init_latents.shape[0]
            and batch_size % init_latents.shape[0] == 0
        ):
            # expand init_latents for batch_size
            deprecation_message = (
                f"You have passed {batch_size} text prompts (`prompt`), but only {init_latents.shape[0]} initial"
                " images (`image`). Initial images are now duplicating to match the number of text prompts. Note"
                " that this behavior is deprecated and will be removed in a version 1.0.0. Please make sure to update"
                " your script to pass as many initial images as text prompts to suppress this warning."
            )
            deprecate(
                "len(prompt) != len(image)",
                "1.0.0",
                deprecation_message,
                standard_warn=False,
            )
            additional_image_per_prompt = batch_size // init_latents.shape[0]
            init_latents = torch.cat(
                [init_latents] * additional_image_per_prompt, dim=0
            )
        elif (
            batch_size > init_latents.shape[0]
            and batch_size % init_latents.shape[0] != 0
        ):
            raise ValueError(
                f"Cannot duplicate `image` of batch size {init_latents.shape[0]} to {batch_size} text prompts."
            )
        else:
            init_latents = torch.cat([init_latents], dim=0)

        rand_device = "cpu" if device.type == "mps" else device
        shape = init_latents.shape
        if isinstance(generator, list):
            shape = (1,) + shape[1:]
            noise = [
                torch.randn(
                    shape, generator=generator[i], device=rand_device, dtype=dtype
                )
                for i in range(batch_size)
            ]
            noise = torch.cat(noise, dim=0).to(device)
        else:
            noise = torch.randn(
                shape, generator=generator, device=rand_device, dtype=dtype
            ).to(device)

        # get latents
        init_latents = self.scheduler.add_noise(init_latents, noise, timestep)
        latents = init_latents

        return latents

    def latent_loss(self, latents, target_latents):
        loss = torch.abs(latents - target_latents).mean()
        return loss

    def latents_to_pil(self, latents):
        # bath of latents -> list of images
        latents = (1 / 0.18215) * latents
        with torch.no_grad():
            image = self.vae.decode(latents).sample
        image = (image / 2 + 0.5).clamp(0, 1)
        image = image.detach().cpu().permute(0, 2, 3, 1).numpy()
        images = (image * 255).round().astype("uint8")
        pil_images = [PIL.Image.fromarray(image) for image in images]
        return pil_images

    def pil_to_latent(self, input_im):
        # Single image -> single latent in a batch (so size 1, 4, 64, 64)
        # TODO: only work with fp16
        tensor = (
            tfms.ToTensor()(input_im).unsqueeze(0).to("cuda").half() * 2 - 1
        )  # Note scaling
        print(tensor.shape)
        print(tensor)
        with torch.no_grad():
            latent = self.vae.encode(tensor)
            return 0.18215 * latent.latent_dist.sample()

    def __call__(
        self,
        prompt: Union[str, List[str]],
        image: Union[torch.FloatTensor, PIL.Image.Image] = None,
        latent_target_image: Union[torch.FloatTensor, PIL.Image.Image] = None,
        strength: float = 0.8,
        latent_loss_scale: int = 100,
        num_inference_steps: Optional[int] = 50,
        guidance_scale: Optional[float] = 7.5,
        negative_prompt: Optional[Union[str, List[str]]] = None,
        num_images_per_prompt: Optional[int] = 1,
        eta: Optional[float] = 0.0,
        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
        output_type: Optional[str] = "pil",
        return_dict: bool = True,
        callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
        callback_steps: Optional[int] = 1,
        **kwargs,
    ):
        # 1. Check inputs
        self.check_inputs(prompt, strength, callback_steps)

        # 2. Define call parameters
        batch_size = 1 if isinstance(prompt, str) else len(prompt)
        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

        # 3. Encode input prompt
        with torch.no_grad():
            text_embeddings = self._encode_prompt(
                prompt,
                device,
                num_images_per_prompt,
                do_classifier_free_guidance,
                negative_prompt,
            )

        # 4. Preprocess image
        image = preprocess(image)

        # 5. set timesteps
        self.scheduler.set_timesteps(num_inference_steps, device=device)
        timesteps, num_inference_steps = self.get_timesteps(
            num_inference_steps, strength, device
        )
        latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt)

        # 6. Prepare latent variables
        latents = self.prepare_latents(
            image,
            latent_timestep,
            batch_size,
            num_images_per_prompt,
            text_embeddings.dtype,
            device,
            generator,
        )

        # TODO: don't use original functions
        if latent_target_image is not None:
            target_latents = self.pil_to_latent(latent_target_image)

        # 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)

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

                # predict the noise residual
                with torch.no_grad():
                    noise_pred = self.unet(
                        latent_model_input, t, encoder_hidden_states=text_embeddings
                    ).sample

                # 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
                    )

                #### ADDITIONAL GUIDANCE ###
                if i % 5 == 0 and latent_target_image is not None:
                    # Requires grad on the latents
                    latents = latents.detach().requires_grad_()

                    # Get the predicted x0:
                    # latents_x0 = latents - sigma * noise_pred
                    latents_x0 = self.scheduler.step(
                        noise_pred, t, latents
                    ).pred_original_sample

                    # Calculate loss
                    loss = (
                        self.latent_loss(latents_x0, target_latents) * latent_loss_scale
                    )

                    # Occasionally print it out
                    print(i, "loss:", loss.item())

                    # Get gradient
                    cond_grad = torch.autograd.grad(loss, latents)[0]

                    # Modify the latents based on this gradient
                    latents = (
                        latents.detach() - cond_grad # * self.scheduler.sigmas[i] ** 2
                    )

                # compute the previous noisy sample x_t -> x_t-1
                latents = self.scheduler.step(
                    noise_pred, t, latents, **extra_step_kwargs
                ).prev_sample

                # 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)

        # TODO: don't use original functions
        image = self.latents_to_pil(latents)

        return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=None)


if __name__ == "__main__":
    img = PIL.Image.open("/content/drive/MyDrive/data/Miku_dance/0095.png").resize(
        (512, 512)
    )
    latent_target_img = PIL.Image.open(
        "/content/drive/MyDrive/data/Miku_dance/0093.png"
    ).resize((512, 512))
    prompt = "Hatsune Miku, masterpiece, best quality, beautiful detailed, kawaii moe sexy bishoujo, 1girl, bare shoulders, upper body, dynamic pose, black outline, blush, sketch written by pencil"
    negative_prompt = "bad anatomy, mutated hands, mutated, poorly, long arm, part of head, mutation, poorly drawn, malformed, text, 3d"
    strength = 0.3
    latent_loss_scale = 3000

    pipe = StableDiffusionImg2ImgPipeline.from_pretrained(
        "/content/drive/MyDrive/models/diffusers/waifu_diffusion_1_3"
    ).to("cuda")
    out = pipe(
        prompt=prompt,
        negative_prompt=negative_prompt,
        image=img,
        latent_target_image=latent_target_img,
        strength=strength,
        latent_loss_scale=latent_loss_scale,
    ).images[0]

    out.save("result.png")
    print("done")