pytorch_svgrender/painter/vectorfusion/LSDS_SDXL_pipeline.py

# -*- coding: utf-8 -*-
# Copyright (c) XiMing Xing. All rights reserved.
# Author: XiMing Xing
# Description:

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

import torch
import torch.nn.functional as F
from torch.cuda.amp import custom_bwd, custom_fwd
from torchvision import transforms
from diffusers.pipelines.stable_diffusion_xl import StableDiffusionXLPipelineOutput
from diffusers.pipelines.stable_diffusion_xl import StableDiffusionXLPipeline


class LSDSSDXLPipeline(StableDiffusionXLPipeline):
    r"""
    Pipeline for text-to-image generation using Stable Diffusion XL.
    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"]

    @torch.no_grad()
    def __call__(
            self,
            prompt: Union[str, List[str]],
            prompt_2: Optional[Union[str, List[str]]] = None,
            height: Optional[int] = None,
            width: Optional[int] = None,
            num_inference_steps: int = 50,
            denoising_end: Optional[float] = None,
            guidance_scale: float = 5.0,
            negative_prompt: Optional[Union[str, List[str]]] = None,
            negative_prompt_2: Optional[Union[str, List[str]]] = None,
            num_images_per_prompt: Optional[int] = 1,
            eta: float = 0.0,
            generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
            latents: Optional[torch.FloatTensor] = None,
            output_type: Optional[str] = "pil",
            return_dict: bool = True,
            callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
            callback_steps: Optional[int] = 1,
            original_size: Optional[Tuple[int, int]] = None,
            crops_coords_top_left: Tuple[int, int] = (0, 0),
            target_size: Optional[Tuple[int, int]] = None,
    ):
        r"""
        Function invoked when calling the pipeline for generation.

        Args:
            prompt (`str` or `List[str]`, *optional*):
                The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
                instead.
            prompt_2 (`str` or `List[str]`, *optional*):
                The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
                used in both text-encoders
            height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
                The height in pixels of the generated image.
            width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
                The width in pixels of the generated image.
            num_inference_steps (`int`, *optional*, defaults to 50):
                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
                expense of slower inference.
            denoising_end (`float`, *optional*):
                When specified, determines the fraction (between 0.0 and 1.0) of the total denoising process to be
                completed before it is intentionally prematurely terminated. As a result, the returned sample will
                still retain a substantial amount of noise as determined by the discrete timesteps selected by the
                scheduler. The denoising_end parameter should ideally be utilized when this pipeline forms a part of a
                "Mixture of Denoisers" multi-pipeline setup, as elaborated in [**Refining the Image
                Output**](https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion/stable_diffusion_xl#refining-the-image-output)
            guidance_scale (`float`, *optional*, defaults to 5.0):
                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.
            negative_prompt (`str` or `List[str]`, *optional*):
                The prompt or prompts not to guide the image generation. If not defined, one has to pass
                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
                less than `1`).
            negative_prompt_2 (`str` or `List[str]`, *optional*):
                The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and
                `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders
            num_images_per_prompt (`int`, *optional*, defaults to 1):
                The number of images to generate per prompt.
            eta (`float`, *optional*, defaults to 0.0):
                Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
                [`schedulers.DDIMScheduler`], will be ignored for others.
            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
                One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
                to make generation deterministic.
            latents (`torch.FloatTensor`, *optional*):
                Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
                tensor will ge generated by sampling using the supplied random `generator`.
            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_xl.StableDiffusionXLPipelineOutput`] 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: torch.FloatTensor)`.
            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.
            original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
                If `original_size` is not the same as `target_size` the image will appear to be down- or upsampled.
                `original_size` defaults to `(width, height)` if not specified. Part of SDXL's micro-conditioning as
                explained in section 2.2 of
                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
            crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)):
                `crops_coords_top_left` can be used to generate an image that appears to be "cropped" from the position
                `crops_coords_top_left` downwards. Favorable, well-centered images are usually achieved by setting
                `crops_coords_top_left` to (0, 0). Part of SDXL's micro-conditioning as explained in section 2.2 of
                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
            target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
                For most cases, `target_size` should be set to the desired height and width of the generated image. If
                not specified it will default to `(width, height)`. Part of SDXL's micro-conditioning as explained in
                section 2.2 of [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).

        Examples:

        Returns:
            [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] or `tuple`:
            [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] if `return_dict` is True, otherwise a
            `tuple`. When returning a tuple, the first element is a list with the generated images.
        """

        # 0. Default height and width to unet
        height = height or self.unet.config.sample_size * self.vae_scale_factor
        width = width or self.unet.config.sample_size * self.vae_scale_factor

        original_size = original_size or (height, width)
        target_size = target_size or (height, width)

        # 1. Check inputs. Raise error if not correct
        self.check_inputs(prompt, prompt_2, height, width, 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
        (
            text_embeddings,
            negative_text_embeddings,
            pooled_text_embeddings,
            negative_pooled_text_embeddings,
        ) = self.encode_prompt(
            prompt=prompt,
            prompt_2=prompt_2,
            device=device,
            num_images_per_prompt=num_images_per_prompt,
            do_classifier_free_guidance=do_classifier_free_guidance,
            negative_prompt=negative_prompt,
            negative_prompt_2=negative_prompt_2,
        )

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

        # 5. Prepare latent variables
        try:
            num_channels_latents = self.unet.config.in_channels
        except Exception or Warning:
            num_channels_latents = self.unet.in_channels

        latents = self.prepare_latents(
            batch_size * num_images_per_prompt,
            num_channels_latents,
            height,
            width,
            text_embeddings.dtype,
            device,
            generator,
            latents,
        )

        # 6. Prepare extra step kwargs. inherit TODO: Logic should ideally just be moved out of the pipeline
        extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)

        # 7. Prepare added time ids & embeddings
        add_text_embeddings = pooled_text_embeddings
        add_time_ids = self._get_add_time_ids(
            original_size, crops_coords_top_left, target_size, dtype=text_embeddings.dtype
        )

        if do_classifier_free_guidance:
            text_embeddings = torch.cat([negative_text_embeddings, text_embeddings], dim=0)
            add_text_embeddings = torch.cat([negative_pooled_text_embeddings, add_text_embeddings], dim=0)
            add_time_ids = torch.cat([add_time_ids, add_time_ids], dim=0)

        text_embeddings = text_embeddings.to(device)
        add_text_embeddings = add_text_embeddings.to(device)
        add_time_ids = add_time_ids.to(device).repeat(batch_size * num_images_per_prompt, 1)

        # 8. Denoising loop
        num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)

        # 8.1 Apply denoising_end
        if denoising_end is not None and type(denoising_end) == float and denoising_end > 0 and denoising_end < 1:
            discrete_timestep_cutoff = int(
                round(
                    self.scheduler.config.num_train_timesteps
                    - (denoising_end * self.scheduler.config.num_train_timesteps)
                )
            )
            num_inference_steps = len(list(filter(lambda ts: ts >= discrete_timestep_cutoff, timesteps)))
            timesteps = timesteps[:num_inference_steps]

        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
                added_cond_kwargs = {"text_embeds": add_text_embeddings, "time_ids": add_time_ids}
                noise_pred = self.unet(
                    latent_model_input,
                    t,
                    encoder_hidden_states=text_embeddings,
                    added_cond_kwargs=added_cond_kwargs
                ).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)

                # 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)

        # 9. Post-processing

        # The decode_latents method is deprecated and has been removed in sdxl
        # image = self.decode_latents(latents)

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

        if not output_type == "latent":
            image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
        else:
            image = latents
            return StableDiffusionXLPipelineOutput(images=image)

        # apply watermark if available
        if self.watermark is not None:
            image = self.watermark.apply_watermark(image)

        image = self.image_processor.postprocess(image, output_type=output_type)

        if not return_dict:
            return (image,)

        return StableDiffusionXLPipelineOutput(images=image)

    def encode_(self, images):
        images = (2 * images - 1).clamp(-1.0, 1.0)  # images: [B, 3, H, W]

        # encode images
        latents = self.vae.encode(images).latent_dist.sample()
        latents = self.vae.config.scaling_factor * latents

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

        return latents

    def x_augment(self, x: torch.Tensor, img_size: int = 1024):
        augment_compose = transforms.Compose([
            transforms.RandomPerspective(distortion_scale=0.5, p=0.7),
            transforms.RandomCrop(size=(img_size, img_size), pad_if_needed=True, padding_mode='reflect')
        ])
        return augment_compose(x)

    def score_distillation_sampling(self,
                                    pred_rgb: torch.Tensor,
                                    im_size: int,
                                    prompt: Union[List, str],
                                    prompt_2: Optional[Union[List, str]] = None,
                                    height: Optional[int] = None,
                                    width: Optional[int] = None,
                                    negative_prompt: Union[List, str] = None,
                                    negative_prompt_2: Optional[Union[List, str]] = None,
                                    guidance_scale: float = 100,
                                    as_latent: bool = False,
                                    grad_scale: float = 1,
                                    t_range: Union[List[float], Tuple[float]] = (0.05, 0.95),
                                    original_size: Optional[Tuple[int, int]] = None,
                                    crops_coords_top_left: Tuple[int, int] = (0, 0),
                                    target_size: Optional[Tuple[int, int]] = None):
        height = height or self.unet.config.sample_size * self.vae_scale_factor
        width = width or self.unet.config.sample_size * self.vae_scale_factor

        original_size = original_size or (height, width)
        target_size = target_size or (height, width)

        batch_size = 1 if isinstance(prompt, str) else len(prompt)

        num_train_timesteps = self.scheduler.config.num_train_timesteps
        min_step = int(num_train_timesteps * t_range[0])
        max_step = int(num_train_timesteps * t_range[1])
        alphas = self.scheduler.alphas_cumprod.to(self.device)  # for convenience

        # input augmentation
        pred_rgb_a = self.x_augment(pred_rgb, im_size)

        # interp to im_size x im_size to be fed into vae.
        if as_latent:
            latents = F.interpolate(pred_rgb_a, (128, 128), mode='bilinear', align_corners=False) * 2 - 1
        else:
            # encode image into latents with vae, requires grad!
            latents = self.encode_(pred_rgb_a)

        #  Encode input prompt
        num_images_per_prompt = 1  # the number of images to generate per prompt
        do_classifier_free_guidance = guidance_scale > 1.0
        (
            text_embeddings,
            negative_text_embeddings,
            pooled_text_embeddings,
            negative_pooled_text_embeddings,
        ) = self.encode_prompt(
            prompt=prompt,
            prompt_2=prompt_2,
            device=self.device,
            num_images_per_prompt=num_images_per_prompt,
            do_classifier_free_guidance=do_classifier_free_guidance,
            negative_prompt=negative_prompt,
            negative_prompt_2=negative_prompt_2,
        )

        # timestep ~ U(0.05, 0.95) to avoid very high/low noise level
        t = torch.randint(min_step, max_step + 1, [1], dtype=torch.long, device=self.device)

        # 7. Prepare added time ids & embeddings
        add_text_embeddings = pooled_text_embeddings
        add_time_ids = self._get_add_time_ids(
            original_size, crops_coords_top_left, target_size, dtype=text_embeddings.dtype
        )

        if do_classifier_free_guidance:
            text_embeddings = torch.cat([negative_text_embeddings, text_embeddings], dim=0)
            add_text_embeddings = torch.cat([negative_pooled_text_embeddings, add_text_embeddings], dim=0)
            add_time_ids = torch.cat([add_time_ids, add_time_ids], dim=0)

        text_embeddings = text_embeddings.to(self.device)
        add_text_embeddings = add_text_embeddings.to(self.device)
        add_time_ids = add_time_ids.to(self.device).repeat(batch_size * num_images_per_prompt, 1)

        # predict the noise residual with unet, stop gradient
        with torch.no_grad():
            # add noise
            noise = torch.randn_like(latents)
            latents_noisy = self.scheduler.add_noise(latents, noise, t)
            # pred noise
            latent_model_input = torch.cat([latents_noisy] * 2) if do_classifier_free_guidance else latents_noisy
            # predict the noise residual
            added_cond_kwargs = {"text_embeds": add_text_embeddings, "time_ids": add_time_ids}
            noise_pred = self.unet(
                latent_model_input,
                t,
                encoder_hidden_states=text_embeddings,
                added_cond_kwargs=added_cond_kwargs
            ).sample

        # perform guidance (high scale from paper!)
        if do_classifier_free_guidance:
            noise_pred_uncond, noise_pred_pos = noise_pred.chunk(2)
            noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_pos - noise_pred_uncond)

        # w(t), sigma_t^2
        w = (1 - alphas[t])
        grad = grad_scale * w * (noise_pred - noise)
        grad = torch.nan_to_num(grad)

        # since we omitted an item in grad, we need to use the custom function to specify the gradient
        loss = SpecifyGradient.apply(latents, grad)

        return loss, grad.mean()


class SpecifyGradient(torch.autograd.Function):

    @staticmethod
    @custom_fwd
    def forward(ctx, input_tensor, gt_grad):
        ctx.save_for_backward(gt_grad)
        # we return a dummy value 1, which will be scaled by amp's scaler so we get the scale in backward.
        return torch.ones([1], device=input_tensor.device, dtype=input_tensor.dtype)

    @staticmethod
    @custom_bwd
    def backward(ctx, grad_scale):
        gt_grad, = ctx.saved_tensors
        gt_grad = gt_grad * grad_scale
        return gt_grad, None