InstDrag/flowdiffusion/pipeline.py

# This file is partially based on the diffusers library, which licensed the code under the following license:

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

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

import PIL.Image
import torch
from transformers import CLIPImageProcessor

from diffusers.callbacks import MultiPipelineCallbacks, PipelineCallback
from diffusers.image_processor import PipelineImageInput, VaeImageProcessor
from diffusers.loaders import StableDiffusionLoraLoaderMixin
from diffusers.models import AutoencoderKL, UNet2DConditionModel
from diffusers.schedulers import KarrasDiffusionSchedulers
from diffusers.utils import deprecate, logging
from diffusers.utils.torch_utils import randn_tensor
from diffusers.pipelines.pipeline_utils import DiffusionPipeline, StableDiffusionMixin
from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput
from diffusers.pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker


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

# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents
def retrieve_latents(
    encoder_output: torch.Tensor, generator: Optional[torch.Generator] = None, sample_mode: str = "sample"
):
    if hasattr(encoder_output, "latent_dist") and sample_mode == "sample":
        return encoder_output.latent_dist.sample(generator)
    elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax":
        return encoder_output.latent_dist.mode()
    elif hasattr(encoder_output, "latents"):
        return encoder_output.latents
    else:
        raise AttributeError("Could not access latents of provided encoder_output")


class FlowDiffusionPipeline(
    DiffusionPipeline,
    StableDiffusionMixin,
    StableDiffusionLoraLoaderMixin,
):
    r"""
    Pipeline for pixel-level image editing given optical flow as condition.

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

    The pipeline also inherits the following loading methods:
        - [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`] for loading LoRA weights
        - [`~loaders.StableDiffusionLoraLoaderMixin.save_lora_weights`] for saving LoRA weights

    Args:
        vae ([`AutoencoderKL`]):
            Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations.
        unet ([`UNet2DConditionModel`]):
            A `UNet2DConditionModel` to denoise the encoded image latents.
        scheduler ([`SchedulerMixin`]):
            A scheduler to be used in combination with `unet` to denoise the encoded image latents.
        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 more details
            about a model's potential harms.
        feature_extractor ([`~transformers.CLIPImageProcessor`]):
            A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`.
    """

    model_cpu_offload_seq = "unet->vae"
    _optional_components = ["safety_checker", "feature_extractor"]
    _exclude_from_cpu_offload = ["safety_checker"]
    _callback_tensor_inputs = ["latents", "image_latents"]

    def __init__(
        self,
        vae: AutoencoderKL,
        unet: UNet2DConditionModel,
        scheduler: KarrasDiffusionSchedulers,
        safety_checker: StableDiffusionSafetyChecker,
        feature_extractor: CLIPImageProcessor,
        requires_safety_checker: bool = False,
        null_prompt: str = "../utils/null_prompt.pt"
    ):
        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."
            )

        self.register_modules(
            vae=vae,
            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.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
        self.register_to_config(requires_safety_checker=requires_safety_checker)
        self.null_prompt_embeds = torch.load(os.path.join(Path(__file__).parent.absolute(), null_prompt), map_location="cpu")

    @torch.no_grad()
    def __call__(
        self,
        image: PipelineImageInput = None,
        flow: torch.Tensor = None,
        num_inference_steps: int = 20,
        image_guidance_scale: float = 1.5,
        flow_guidance_scale: float = 1.5,
        eta: float = 0.0,
        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
        latents: Optional[torch.Tensor] = None,
        output_type: Optional[str] = "pil",
        return_dict: bool = True,
        callback_on_step_end: Optional[
            Union[Callable[[int, int, Dict], None], PipelineCallback, MultiPipelineCallbacks]
        ] = None,
        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
        **kwargs,
    ):
        r"""
        The call function to the pipeline for generation.

        Args:
            image (`torch.Tensor` `np.ndarray`, `PIL.Image.Image`, `List[torch.Tensor]`, `List[PIL.Image.Image]`, or `List[np.ndarray]`):
                `Image` or tensor representing an image batch to be repainted according to `prompt`. Can also accept
                image latents as `image`, but if passing latents directly it is not encoded again. We only support batch size of 1 for now.
            flow: torch.Tensor = None,
                Optical flow tensor to be used as a condition for the image generation. We only support batch size of 1 for now.
            num_inference_steps (`int`, *optional*, defaults to 20):
                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
                expense of slower inference.
            image_guidance_scale (`float`, *optional*, defaults to 1.5):
                Push the generated image towards the initial `image`. Image guidance scale is enabled by setting
                `image_guidance_scale > 1`. Higher image guidance scale encourages generated images that are closely
                linked to the source `image`, usually at the expense of lower image quality. This pipeline requires a
                value of at least `1`.
            flow_guidance_scale (`float`, *optional*, defaults to 1.5):
                Apply the flow guidance to the image generation. Higher values of `flow_guidance_scale` encourage
                the model to follow the flow stronger.
            eta (`float`, *optional*, defaults to 0.0):
                Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
                to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
            generator (`torch.Generator`, *optional*):
                A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
                generation deterministic.
            latents (`torch.Tensor`, *optional*):
                Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
                tensor is generated by sampling using the supplied random `generator`.
            output_type (`str`, *optional*, defaults to `"pil"`):
                The output format of the generated image. Choose between `PIL.Image` or `np.array`.
            return_dict (`bool`, *optional*, defaults to `True`):
                Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
                plain tuple.
            callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*):
                A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of
                each denoising step during the inference. with the following arguments: `callback_on_step_end(self:
                DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a
                list of all tensors as specified by `callback_on_step_end_tensor_inputs`.
            callback_on_step_end_tensor_inputs (`List`, *optional*):
                The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
                will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
                `._callback_tensor_inputs` attribute of your pipeline class.
            cross_attention_kwargs (`dict`, *optional*):
                A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in
                [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
        """

        callback = kwargs.pop("callback", None)
        callback_steps = kwargs.pop("callback_steps", None)

        if callback is not None:
            deprecate(
                "callback",
                "1.0.0",
                "Passing `callback` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`",
            )
        if callback_steps is not None:
            deprecate(
                "callback_steps",
                "1.0.0",
                "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`",
            )

        if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)):
            callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs

        # 0. Check inputs
        self.check_inputs(
            callback_steps,
            callback_on_step_end_tensor_inputs,
        )
        self._image_guidance_scale = image_guidance_scale
        self._flow_guidance_scale = flow_guidance_scale

        device = self._execution_device

        if image is None or flow is None:
            raise ValueError("`image` or `flow` input cannot be undefined.")

        # 1. Define call parameters

        # 2. Encode input prompt
        prompt_embeds = self._encode_prompt(
            device,
            self.do_classifier_free_guidance,
        )

        # 3. Preprocess image
        image = self.image_processor.preprocess(image)
        assert image.shape[0] == 1 and flow.shape[0] == 1, "Batch size must be 1 for now."

        # 4. set timesteps
        self.scheduler.set_timesteps(num_inference_steps, device=device)
        timesteps = self.scheduler.timesteps

        # 5. Prepare Image latents
        image_latents = self.prepare_image_latents(
            image,
            flow,
            prompt_embeds.dtype,
            device,
            self.do_classifier_free_guidance,
        )

        height, width = image_latents.shape[-2:]
        height = height * self.vae_scale_factor
        width = width * self.vae_scale_factor

        # 6. Prepare latent variables
        num_channels_latents = self.vae.config.latent_channels
        latents = self.prepare_latents(
            num_channels_latents,
            height,
            width,
            prompt_embeds.dtype,
            device,
            generator,
            latents,
        )

        # 7. Check that shapes of latents and image match the UNet channels
        num_channels_image = image_latents.shape[1]
        if num_channels_latents + num_channels_image != self.unet.config.in_channels:
            raise ValueError(
                f"Incorrect configuration settings! The config of `pipeline.unet`: {self.unet.config} expects"
                f" {self.unet.config.in_channels} but received `num_channels_latents`: {num_channels_latents} +"
                f" `num_channels_image`: {num_channels_image} "
                f" = {num_channels_latents+num_channels_image}. Please verify the config of"
                " `pipeline.unet` or your `image` input."
            )

        # 8. 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)

        # 9. Denoising loop
        num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
        self._num_timesteps = len(timesteps)
        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.
                # The latents are expanded 3 times because for image / flow guidance
                latent_model_input = torch.cat([latents] * 3) if self.do_classifier_free_guidance else latents

                # concat latents, image_latents in the channel dimension
                scaled_latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
                scaled_latent_model_input = torch.cat([scaled_latent_model_input, image_latents], dim=1)

                # predict the noise residual
                noise_pred = self.unet(
                    scaled_latent_model_input,
                    t,
                    encoder_hidden_states=prompt_embeds,
                    added_cond_kwargs=None,
                    cross_attention_kwargs=cross_attention_kwargs,
                    return_dict=False,
                )[0]

                # perform guidance
                if self.do_classifier_free_guidance:
                    noise_pred_flow, noise_pred_image, noise_pred_uncond = noise_pred.chunk(3)
                    noise_pred = (
                        noise_pred_uncond
                        + self._image_guidance_scale * (noise_pred_image - noise_pred_uncond)
                        + self._flow_guidance_scale * (noise_pred_flow - noise_pred_image)
                    )

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

                if callback_on_step_end is not None:
                    callback_kwargs = {}
                    for k in callback_on_step_end_tensor_inputs:
                        callback_kwargs[k] = locals()[k]
                    callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)

                    latents = callback_outputs.pop("latents", latents)
                    prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
                    image_latents = callback_outputs.pop("image_latents", image_latents)

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

        if not output_type == "latent":
            image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
            image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype)
        else:
            image = latents
            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.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize)

        # Offload all models
        self.maybe_free_model_hooks()

        if not return_dict:
            return (image, has_nsfw_concept)

        return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)

    def _encode_prompt(
        self,
        device,
        do_classifier_free_guidance,
    ):
        prompt_embeds = self.null_prompt_embeds.to(dtype=torch.float16, device=device) # 1 77 512
        
        if do_classifier_free_guidance: # We are only doing cfg for image and flow
            prompt_embeds = torch.cat([prompt_embeds, prompt_embeds, prompt_embeds]) # 3 77 512
        
        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.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.decode_latents
    def decode_latents(self, latents):
        deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead"
        deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)

        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

    def check_inputs(
        self,
        callback_steps,
        callback_on_step_end_tensor_inputs=None,
    ):
        if 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 callback_on_step_end_tensor_inputs is not None and not all(
            k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
        ):
            raise ValueError(
                f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
            )

    def prepare_latents(self, num_channels_latents, height, width, dtype, device, generator, latents=None):
        shape = (
            1,
            num_channels_latents,
            int(height) // self.vae_scale_factor,
            int(width) // self.vae_scale_factor,
        )
        if isinstance(generator, list) and len(generator) != 1:
            raise ValueError(
                f"You have passed a list of generators of length {len(generator)}, but we only support a single batch for now."
            )

        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 prepare_image_latents(
        self, image, flow, dtype, device, do_classifier_free_guidance, generator=None
    ):
        if not isinstance(image, (torch.Tensor, PIL.Image.Image, list)):
            raise ValueError(
                f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}"
            )

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

        if image.shape[1] == 4:
            image_latents = image
        else:
            image_latents = retrieve_latents(self.vae.encode(image), sample_mode="argmax")
        
        image_latents_flow_cond = torch.cat([image_latents, flow.to(device)], dim=1)

        if do_classifier_free_guidance:
            image_latents_flow_uncond = torch.cat([image_latents, torch.zeros_like(flow).to(device)], dim=1)
            image_latents_uncond = torch.zeros_like(image_latents_flow_cond)
            image_latents_final = torch.cat([image_latents_flow_cond, image_latents_flow_uncond, image_latents_uncond], dim=0)
        else:
            image_latents_final = image_latents_flow_cond

        return image_latents_final

    @property
    def image_guidance_scale(self):
        return self._image_guidance_scale

    @property
    def flow_guidance_scale(self):
        return self._flow_guidance_scale

    @property
    def num_timesteps(self):
        return self._num_timesteps

    @property
    def do_classifier_free_guidance(self):
        return self._image_guidance_scale > 1 or self._flow_guidance_scale > 1