ppdiffusers/pipelines/stable_diffusion/pipeline_cycle_diffusion.py

# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
# 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 paddle
import PIL
from packaging import version

from paddlenlp.transformers import CLIPFeatureExtractor, CLIPTextModel, CLIPTokenizer

from ...configuration_utils import FrozenDict
from ...models import AutoencoderKL, UNet2DConditionModel
from ...pipeline_utils import DiffusionPipeline
from ...schedulers import DDIMScheduler
from ...utils import PIL_INTERPOLATION, deprecate, logging
from . import StableDiffusionPipelineOutput
from .safety_checker import StableDiffusionSafetyChecker

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


# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.preprocess
def preprocess(image):
    if isinstance(image, paddle.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 = paddle.to_tensor(image)
    elif isinstance(image[0], paddle.Tensor):
        image = paddle.concat(image, axis=0)
    return image


def posterior_sample(scheduler, latents, timestep, clean_latents, generator, eta):
    # 1. get previous step value (=t-1)
    prev_timestep = timestep - scheduler.config.num_train_timesteps // scheduler.num_inference_steps

    if prev_timestep <= 0:
        return clean_latents

    # 2. compute alphas, betas
    alpha_prod_t = scheduler.alphas_cumprod[timestep]
    alpha_prod_t_prev = (
        scheduler.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else scheduler.final_alpha_cumprod
    )

    variance = scheduler._get_variance(timestep, prev_timestep)
    std_dev_t = eta * variance ** (0.5)

    # direction pointing to x_t
    e_t = (latents - alpha_prod_t ** (0.5) * clean_latents) / (1 - alpha_prod_t) ** (0.5)
    dir_xt = (1.0 - alpha_prod_t_prev - std_dev_t**2) ** (0.5) * e_t
    noise = std_dev_t * paddle.randn(clean_latents.shape, dtype=clean_latents.dtype, generator=generator)
    prev_latents = alpha_prod_t_prev ** (0.5) * clean_latents + dir_xt + noise

    return prev_latents


def compute_noise(scheduler, prev_latents, latents, timestep, noise_pred, eta):
    # 1. get previous step value (=t-1)
    prev_timestep = timestep - scheduler.config.num_train_timesteps // scheduler.num_inference_steps

    # 2. compute alphas, betas
    alpha_prod_t = scheduler.alphas_cumprod[timestep]
    alpha_prod_t_prev = (
        scheduler.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else scheduler.final_alpha_cumprod
    )

    beta_prod_t = 1 - alpha_prod_t

    # 3. compute predicted original sample from predicted noise also called
    # "predicted x_0" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
    pred_original_sample = (latents - beta_prod_t ** (0.5) * noise_pred) / alpha_prod_t ** (0.5)

    # 4. Clip "predicted x_0"
    if scheduler.config.clip_sample:
        pred_original_sample = pred_original_sample.clip(-1, 1)

    # 5. compute variance: "sigma_t(η)" -> see formula (16)
    # σ_t = sqrt((1 − α_t−1)/(1 − α_t)) * sqrt(1 − α_t/α_t−1)
    variance = scheduler._get_variance(timestep, prev_timestep)
    std_dev_t = eta * variance ** (0.5)

    # 6. compute "direction pointing to x_t" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
    pred_sample_direction = (1 - alpha_prod_t_prev - std_dev_t**2) ** (0.5) * noise_pred

    noise = (prev_latents - (alpha_prod_t_prev ** (0.5) * pred_original_sample + pred_sample_direction)) / (
        variance ** (0.5) * eta
    )
    return noise


class CycleDiffusionPipeline(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 xxxx, 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/CompVis/stable-diffusion-v1-4) 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"]

    def __init__(
        self,
        vae: AutoencoderKL,
        text_encoder: CLIPTextModel,
        tokenizer: CLIPTokenizer,
        unet: UNet2DConditionModel,
        scheduler: DDIMScheduler,
        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 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. PaddleNLP team, 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, "_ppdiffusers_version") and version.parse(
            version.parse(unet.config._ppdiffusers_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.register_to_config(requires_safety_checker=requires_safety_checker)

    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt
    def _encode_prompt(self, prompt, 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
            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="pd",
        )
        text_input_ids = text_inputs.input_ids
        untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pd").input_ids

        if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not paddle.equal_all(
            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
        else:
            attention_mask = None

        text_embeddings = self.text_encoder(
            text_input_ids,
            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.tile([1, num_images_per_prompt, 1])
        text_embeddings = text_embeddings.reshape([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="pd",
            )

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

            uncond_embeddings = self.text_encoder(
                uncond_input.input_ids,
                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.tile([1, num_images_per_prompt, 1])
            uncond_embeddings = uncond_embeddings.reshape([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 = paddle.concat([uncond_embeddings, text_embeddings])

        return text_embeddings

    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.StableDiffusionImg2ImgPipeline.check_inputs
    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)}."
            )

    # 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

    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker
    def run_safety_checker(self, image, dtype):
        if self.safety_checker is not None:
            safety_checker_input = self.feature_extractor(self.numpy_to_pil(image), return_tensors="pd")
            image, has_nsfw_concept = self.safety_checker(
                images=image, clip_input=safety_checker_input.pixel_values.cast(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).clip(0, 1)
        # we always cast to float32 as this does not cause significant overhead and is compatible with bfloa16
        image = image.transpose([0, 2, 3, 1]).cast("float32").numpy()
        return image

    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.StableDiffusionImg2ImgPipeline.get_timesteps
    def get_timesteps(self, num_inference_steps, strength):
        # 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, generator=None):
        image = image.cast(dtype=dtype)

        batch_size = image.shape[0]
        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 = paddle.concat(init_latents, axis=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 = paddle.concat([init_latents] * additional_image_per_prompt * num_images_per_prompt, axis=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 = paddle.concat([init_latents] * num_images_per_prompt, axis=0)

        # add noise to latents using the timestep
        shape = init_latents.shape
        if isinstance(generator, list):
            shape = [
                1,
            ] + shape[1:]
            noise = [paddle.randn(shape, generator=generator[i], dtype=dtype) for i in range(batch_size)]
            noise = paddle.concat(noise, axis=0)
        else:
            noise = paddle.randn(shape, generator=generator, dtype=dtype)

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

        return latents, clean_latents

    @paddle.no_grad()
    def __call__(
        self,
        prompt: Union[str, List[str]],
        source_prompt: Union[str, List[str]],
        image: Union[paddle.Tensor, PIL.Image.Image] = None,
        strength: float = 0.8,
        num_inference_steps: Optional[int] = 50,
        guidance_scale: Optional[float] = 7.5,
        source_guidance_scale: Optional[float] = 1,
        num_images_per_prompt: Optional[int] = 1,
        eta: Optional[float] = 0.1,
        generator: Optional[Union[paddle.Generator, List[paddle.Generator]]] = None,
        output_type: Optional[str] = "pil",
        return_dict: bool = True,
        callback: Optional[Callable[[int, int, paddle.Tensor], None]] = None,
        callback_steps: Optional[int] = 1,
    ):
        r"""
        Function invoked when calling the pipeline for generation.

        Args:
            prompt (`str` or `List[str]`):
                The prompt or prompts to guide the image generation.
            image (`paddle.Tensor` or `PIL.Image.Image`):
                `Image`, or tensor representing an image batch, that will be used as the starting point for the
                process.
            strength (`float`, *optional*, defaults to 0.8):
                Conceptually, indicates how much to transform the reference `image`. Must be between 0 and 1.
                `image` will be used as a starting point, adding more noise to it the larger the `strength`. The
                number of denoising steps depends on the amount of noise initially added. When `strength` is 1, added
                noise will be maximum and the denoising process will run for the full number of iterations specified in
                `num_inference_steps`. A value of 1, therefore, essentially ignores `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. This parameter will be modulated by `strength`.
            guidance_scale (`float`, *optional*, defaults to 7.5):
                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
                `guidance_scale` is defined as `w` of equation 2. of [Imagen
                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
                usually at the expense of lower image quality.
            source_guidance_scale (`float`, *optional*, defaults to 1):
                Guidance scale for the source prompt. This is useful to control the amount of influence the source
                prompt for encoding.
            num_images_per_prompt (`int`, *optional*, defaults to 1):
                The number of images to generate per prompt.
            eta (`float`, *optional*, defaults to 0.1):
                Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
                [`schedulers.DDIMScheduler`], will be ignored for others.
            generator (`paddle.Generator`, *optional*):
                One or a list of paddle generator(s) to make generation deterministic.
            output_type (`str`, *optional*, defaults to `"pil"`):
                The output format of the generate image. Choose between
                [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
            return_dict (`bool`, *optional*, defaults to `True`):
                Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
                plain tuple.
            callback (`Callable`, *optional*):
                A function that will be called every `callback_steps` steps during inference. The function will be
                called with the following arguments: `callback(step: int, timestep: int, latents: paddle.Tensor)`.
            callback_steps (`int`, *optional*, defaults to 1):
                The frequency at which the `callback` function will be called. If not specified, the callback will be
                called at every step.

        Returns:
            [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
            [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple.
            When returning a tuple, the first element is a list with the generated images, and the second element is a
            list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work"
            (nsfw) content, according to the `safety_checker`.
        """
        # 1. Check inputs
        self.check_inputs(prompt, strength, callback_steps)

        # 2. Define call parameters
        batch_size = 1 if isinstance(prompt, str) else len(prompt)
        # 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
        text_embeddings = self._encode_prompt(prompt, num_images_per_prompt, do_classifier_free_guidance, None)
        source_text_embeddings = self._encode_prompt(
            source_prompt, num_images_per_prompt, do_classifier_free_guidance, None
        )

        # 4. Preprocess image
        image = preprocess(image)

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

        # 6. Prepare latent variables
        latents, clean_latents = self.prepare_latents(
            image, latent_timestep, batch_size, num_images_per_prompt, text_embeddings.dtype, generator
        )
        source_latents = 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)
        generator = extra_step_kwargs.pop("generator", None)

        # 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 = paddle.concat([latents] * 2)
                source_latent_model_input = paddle.concat([source_latents] * 2)
                latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
                source_latent_model_input = self.scheduler.scale_model_input(source_latent_model_input, t)

                # predict the noise residual
                concat_latent_model_input = paddle.stack(
                    [
                        source_latent_model_input[0],
                        latent_model_input[0],
                        source_latent_model_input[1],
                        latent_model_input[1],
                    ],
                    axis=0,
                )
                concat_text_embeddings = paddle.stack(
                    [
                        source_text_embeddings[0],
                        text_embeddings[0],
                        source_text_embeddings[1],
                        text_embeddings[1],
                    ],
                    axis=0,
                )
                concat_noise_pred = self.unet(
                    concat_latent_model_input, t, encoder_hidden_states=concat_text_embeddings
                ).sample

                # perform guidance
                (
                    source_noise_pred_uncond,
                    noise_pred_uncond,
                    source_noise_pred_text,
                    noise_pred_text,
                ) = concat_noise_pred.chunk(4, axis=0)
                noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
                source_noise_pred = source_noise_pred_uncond + source_guidance_scale * (
                    source_noise_pred_text - source_noise_pred_uncond
                )

                # Sample source_latents from the posterior distribution.
                prev_source_latents = posterior_sample(
                    self.scheduler, source_latents, t, clean_latents, generator=generator, **extra_step_kwargs
                )
                # Compute noise.
                noise = compute_noise(
                    self.scheduler, prev_source_latents, source_latents, t, source_noise_pred, **extra_step_kwargs
                )
                source_latents = prev_source_latents

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

        # 9. Post-processing
        image = self.decode_latents(latents)

        # 10. Run safety checker
        image, has_nsfw_concept = self.run_safety_checker(image, text_embeddings.dtype)

        # 11. Convert to PIL
        if output_type == "pil":
            image = self.numpy_to_pil(image)

        if not return_dict:
            return (image, has_nsfw_concept)

        return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)