diffusers/community-pipelines-mirror

1import inspect2from typing import List, Optional, Union3 4import torch5from torch import nn6from torch.nn import functional as F7 8from diffusers import (9    AutoencoderKL,10    DDIMScheduler,11    DiffusionPipeline,12    LMSDiscreteScheduler,13    PNDMScheduler,14    UNet2DConditionModel,15)16from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion import StableDiffusionPipelineOutput17from torchvision import transforms18from transformers import CLIPFeatureExtractor, CLIPModel, CLIPTextModel, CLIPTokenizer19 20 21class MakeCutouts(nn.Module):22    def __init__(self, cut_size, cut_power=1.0):23        super().__init__()24 25        self.cut_size = cut_size26        self.cut_power = cut_power27 28    def forward(self, pixel_values, num_cutouts):29        sideY, sideX = pixel_values.shape[2:4]30        max_size = min(sideX, sideY)31        min_size = min(sideX, sideY, self.cut_size)32        cutouts = []33        for _ in range(num_cutouts):34            size = int(torch.rand([]) ** self.cut_power * (max_size - min_size) + min_size)35            offsetx = torch.randint(0, sideX - size + 1, ())36            offsety = torch.randint(0, sideY - size + 1, ())37            cutout = pixel_values[:, :, offsety : offsety + size, offsetx : offsetx + size]38            cutouts.append(F.adaptive_avg_pool2d(cutout, self.cut_size))39        return torch.cat(cutouts)40 41 42def spherical_dist_loss(x, y):43    x = F.normalize(x, dim=-1)44    y = F.normalize(y, dim=-1)45    return (x - y).norm(dim=-1).div(2).arcsin().pow(2).mul(2)46 47 48def set_requires_grad(model, value):49    for param in model.parameters():50        param.requires_grad = value51 52 53class CLIPGuidedStableDiffusion(DiffusionPipeline):54    """CLIP guided stable diffusion based on the amazing repo by @crowsonkb and @Jack00055    - https://github.com/Jack000/glid-3-xl56    - https://github.dev/crowsonkb/k-diffusion57    """58 59    def __init__(60        self,61        vae: AutoencoderKL,62        text_encoder: CLIPTextModel,63        clip_model: CLIPModel,64        tokenizer: CLIPTokenizer,65        unet: UNet2DConditionModel,66        scheduler: Union[PNDMScheduler, LMSDiscreteScheduler, DDIMScheduler],67        feature_extractor: CLIPFeatureExtractor,68    ):69        super().__init__()70        self.register_modules(71            vae=vae,72            text_encoder=text_encoder,73            clip_model=clip_model,74            tokenizer=tokenizer,75            unet=unet,76            scheduler=scheduler,77            feature_extractor=feature_extractor,78        )79 80        self.normalize = transforms.Normalize(mean=feature_extractor.image_mean, std=feature_extractor.image_std)81        self.make_cutouts = MakeCutouts(feature_extractor.size)82 83        set_requires_grad(self.text_encoder, False)84        set_requires_grad(self.clip_model, False)85 86    def enable_attention_slicing(self, slice_size: Optional[Union[str, int]] = "auto"):87        if slice_size == "auto":88            # half the attention head size is usually a good trade-off between89            # speed and memory90            slice_size = self.unet.config.attention_head_dim // 291        self.unet.set_attention_slice(slice_size)92 93    def disable_attention_slicing(self):94        self.enable_attention_slicing(None)95 96    def freeze_vae(self):97        set_requires_grad(self.vae, False)98 99    def unfreeze_vae(self):100        set_requires_grad(self.vae, True)101 102    def freeze_unet(self):103        set_requires_grad(self.unet, False)104 105    def unfreeze_unet(self):106        set_requires_grad(self.unet, True)107 108    @torch.enable_grad()109    def cond_fn(110        self,111        latents,112        timestep,113        index,114        text_embeddings,115        noise_pred_original,116        text_embeddings_clip,117        clip_guidance_scale,118        num_cutouts,119        use_cutouts=True,120    ):121        latents = latents.detach().requires_grad_()122 123        if isinstance(self.scheduler, LMSDiscreteScheduler):124            sigma = self.scheduler.sigmas[index]125            # the model input needs to be scaled to match the continuous ODE formulation in K-LMS126            latent_model_input = latents / ((sigma**2 + 1) ** 0.5)127        else:128            latent_model_input = latents129 130        # predict the noise residual131        noise_pred = self.unet(latent_model_input, timestep, encoder_hidden_states=text_embeddings).sample132 133        if isinstance(self.scheduler, (PNDMScheduler, DDIMScheduler)):134            alpha_prod_t = self.scheduler.alphas_cumprod[timestep]135            beta_prod_t = 1 - alpha_prod_t136            # compute predicted original sample from predicted noise also called137            # "predicted x_0" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf138            pred_original_sample = (latents - beta_prod_t ** (0.5) * noise_pred) / alpha_prod_t ** (0.5)139 140            fac = torch.sqrt(beta_prod_t)141            sample = pred_original_sample * (fac) + latents * (1 - fac)142        elif isinstance(self.scheduler, LMSDiscreteScheduler):143            sigma = self.scheduler.sigmas[index]144            sample = latents - sigma * noise_pred145        else:146            raise ValueError(f"scheduler type {type(self.scheduler)} not supported")147 148        sample = 1 / 0.18215 * sample149        image = self.vae.decode(sample).sample150        image = (image / 2 + 0.5).clamp(0, 1)151 152        if use_cutouts:153            image = self.make_cutouts(image, num_cutouts)154        else:155            image = transforms.Resize(self.feature_extractor.size)(image)156        image = self.normalize(image).to(latents.dtype)157 158        image_embeddings_clip = self.clip_model.get_image_features(image)159        image_embeddings_clip = image_embeddings_clip / image_embeddings_clip.norm(p=2, dim=-1, keepdim=True)160 161        if use_cutouts:162            dists = spherical_dist_loss(image_embeddings_clip, text_embeddings_clip)163            dists = dists.view([num_cutouts, sample.shape[0], -1])164            loss = dists.sum(2).mean(0).sum() * clip_guidance_scale165        else:166            loss = spherical_dist_loss(image_embeddings_clip, text_embeddings_clip).mean() * clip_guidance_scale167 168        grads = -torch.autograd.grad(loss, latents)[0]169 170        if isinstance(self.scheduler, LMSDiscreteScheduler):171            latents = latents.detach() + grads * (sigma**2)172            noise_pred = noise_pred_original173        else:174            noise_pred = noise_pred_original - torch.sqrt(beta_prod_t) * grads175        return noise_pred, latents176 177    @torch.no_grad()178    def __call__(179        self,180        prompt: Union[str, List[str]],181        height: Optional[int] = 512,182        width: Optional[int] = 512,183        num_inference_steps: Optional[int] = 50,184        guidance_scale: Optional[float] = 7.5,185        num_images_per_prompt: Optional[int] = 1,186        eta: float = 0.0,187        clip_guidance_scale: Optional[float] = 100,188        clip_prompt: Optional[Union[str, List[str]]] = None,189        num_cutouts: Optional[int] = 4,190        use_cutouts: Optional[bool] = True,191        generator: Optional[torch.Generator] = None,192        latents: Optional[torch.FloatTensor] = None,193        output_type: Optional[str] = "pil",194        return_dict: bool = True,195    ):196        if isinstance(prompt, str):197            batch_size = 1198        elif isinstance(prompt, list):199            batch_size = len(prompt)200        else:201            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")202 203        if height % 8 != 0 or width % 8 != 0:204            raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")205 206        # get prompt text embeddings207        text_input = self.tokenizer(208            prompt,209            padding="max_length",210            max_length=self.tokenizer.model_max_length,211            truncation=True,212            return_tensors="pt",213        )214        text_embeddings = self.text_encoder(text_input.input_ids.to(self.device))[0]215        # duplicate text embeddings for each generation per prompt216        text_embeddings = text_embeddings.repeat_interleave(num_images_per_prompt, dim=0)217 218        if clip_guidance_scale > 0:219            if clip_prompt is not None:220                clip_text_input = self.tokenizer(221                    clip_prompt,222                    padding="max_length",223                    max_length=self.tokenizer.model_max_length,224                    truncation=True,225                    return_tensors="pt",226                ).input_ids.to(self.device)227            else:228                clip_text_input = text_input.input_ids.to(self.device)229            text_embeddings_clip = self.clip_model.get_text_features(clip_text_input)230            text_embeddings_clip = text_embeddings_clip / text_embeddings_clip.norm(p=2, dim=-1, keepdim=True)231            # duplicate text embeddings clip for each generation per prompt232            text_embeddings_clip = text_embeddings_clip.repeat_interleave(num_images_per_prompt, dim=0)233 234        # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)235        # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`236        # corresponds to doing no classifier free guidance.237        do_classifier_free_guidance = guidance_scale > 1.0238        # get unconditional embeddings for classifier free guidance239        if do_classifier_free_guidance:240            max_length = text_input.input_ids.shape[-1]241            uncond_input = self.tokenizer([""], padding="max_length", max_length=max_length, return_tensors="pt")242            uncond_embeddings = self.text_encoder(uncond_input.input_ids.to(self.device))[0]243            # duplicate unconditional embeddings for each generation per prompt244            uncond_embeddings = uncond_embeddings.repeat_interleave(num_images_per_prompt, dim=0)245 246            # For classifier free guidance, we need to do two forward passes.247            # Here we concatenate the unconditional and text embeddings into a single batch248            # to avoid doing two forward passes249            text_embeddings = torch.cat([uncond_embeddings, text_embeddings])250 251        # get the initial random noise unless the user supplied it252 253        # Unlike in other pipelines, latents need to be generated in the target device254        # for 1-to-1 results reproducibility with the CompVis implementation.255        # However this currently doesn't work in `mps`.256        latents_shape = (batch_size * num_images_per_prompt, self.unet.in_channels, height // 8, width // 8)257        latents_dtype = text_embeddings.dtype258        if latents is None:259            if self.device.type == "mps":260                # randn does not work reproducibly on mps261                latents = torch.randn(latents_shape, generator=generator, device="cpu", dtype=latents_dtype).to(262                    self.device263                )264            else:265                latents = torch.randn(latents_shape, generator=generator, device=self.device, dtype=latents_dtype)266        else:267            if latents.shape != latents_shape:268                raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {latents_shape}")269            latents = latents.to(self.device)270 271        # set timesteps272        accepts_offset = "offset" in set(inspect.signature(self.scheduler.set_timesteps).parameters.keys())273        extra_set_kwargs = {}274        if accepts_offset:275            extra_set_kwargs["offset"] = 1276 277        self.scheduler.set_timesteps(num_inference_steps, **extra_set_kwargs)278 279        # Some schedulers like PNDM have timesteps as arrays280        # It's more optimized to move all timesteps to correct device beforehand281        timesteps_tensor = self.scheduler.timesteps.to(self.device)282 283        # scale the initial noise by the standard deviation required by the scheduler284        latents = latents * self.scheduler.init_noise_sigma285 286        # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature287        # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.288        # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502289        # and should be between [0, 1]290        accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())291        extra_step_kwargs = {}292        if accepts_eta:293            extra_step_kwargs["eta"] = eta294 295        # check if the scheduler accepts generator296        accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())297        if accepts_generator:298            extra_step_kwargs["generator"] = generator299 300        for i, t in enumerate(self.progress_bar(timesteps_tensor)):301            # expand the latents if we are doing classifier free guidance302            latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents303            latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)304 305            # predict the noise residual306            noise_pred = self.unet(latent_model_input, t, encoder_hidden_states=text_embeddings).sample307 308            # perform classifier free guidance309            if do_classifier_free_guidance:310                noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)311                noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)312 313            # perform clip guidance314            if clip_guidance_scale > 0:315                text_embeddings_for_guidance = (316                    text_embeddings.chunk(2)[1] if do_classifier_free_guidance else text_embeddings317                )318                noise_pred, latents = self.cond_fn(319                    latents,320                    t,321                    i,322                    text_embeddings_for_guidance,323                    noise_pred,324                    text_embeddings_clip,325                    clip_guidance_scale,326                    num_cutouts,327                    use_cutouts,328                )329 330            # compute the previous noisy sample x_t -> x_t-1331            latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample332 333        # scale and decode the image latents with vae334        latents = 1 / 0.18215 * latents335        image = self.vae.decode(latents).sample336 337        image = (image / 2 + 0.5).clamp(0, 1)338        image = image.cpu().permute(0, 2, 3, 1).numpy()339 340        if output_type == "pil":341            image = self.numpy_to_pil(image)342 343        if not return_dict:344            return (image, None)345 346        return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=None)347