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 AutoencoderKL, DiffusionPipeline, LMSDiscreteScheduler, PNDMScheduler, UNet2DConditionModel9from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion import StableDiffusionPipelineOutput10from torchvision import transforms11from transformers import CLIPFeatureExtractor, CLIPModel, CLIPTextModel, CLIPTokenizer12 13 14class MakeCutouts(nn.Module):15    def __init__(self, cut_size, cut_power=1.0):16        super().__init__()17 18        self.cut_size = cut_size19        self.cut_power = cut_power20 21    def forward(self, pixel_values, num_cutouts):22        sideY, sideX = pixel_values.shape[2:4]23        max_size = min(sideX, sideY)24        min_size = min(sideX, sideY, self.cut_size)25        cutouts = []26        for _ in range(num_cutouts):27            size = int(torch.rand([]) ** self.cut_power * (max_size - min_size) + min_size)28            offsetx = torch.randint(0, sideX - size + 1, ())29            offsety = torch.randint(0, sideY - size + 1, ())30            cutout = pixel_values[:, :, offsety : offsety + size, offsetx : offsetx + size]31            cutouts.append(F.adaptive_avg_pool2d(cutout, self.cut_size))32        return torch.cat(cutouts)33 34 35def spherical_dist_loss(x, y):36    x = F.normalize(x, dim=-1)37    y = F.normalize(y, dim=-1)38    return (x - y).norm(dim=-1).div(2).arcsin().pow(2).mul(2)39 40 41def set_requires_grad(model, value):42    for param in model.parameters():43        param.requires_grad = value44 45 46class CLIPGuidedStableDiffusion(DiffusionPipeline):47    """CLIP guided stable diffusion based on the amazing repo by @crowsonkb and @Jack00048    - https://github.com/Jack000/glid-3-xl49    - https://github.dev/crowsonkb/k-diffusion50    """51 52    def __init__(53        self,54        vae: AutoencoderKL,55        text_encoder: CLIPTextModel,56        clip_model: CLIPModel,57        tokenizer: CLIPTokenizer,58        unet: UNet2DConditionModel,59        scheduler: Union[PNDMScheduler, LMSDiscreteScheduler],60        feature_extractor: CLIPFeatureExtractor,61    ):62        super().__init__()63        self.register_modules(64            vae=vae,65            text_encoder=text_encoder,66            clip_model=clip_model,67            tokenizer=tokenizer,68            unet=unet,69            scheduler=scheduler,70            feature_extractor=feature_extractor,71        )72 73        self.normalize = transforms.Normalize(mean=feature_extractor.image_mean, std=feature_extractor.image_std)74        self.make_cutouts = MakeCutouts(feature_extractor.size)75 76        set_requires_grad(self.text_encoder, False)77        set_requires_grad(self.clip_model, False)78 79    def enable_attention_slicing(self, slice_size: Optional[Union[str, int]] = "auto"):80        if slice_size == "auto":81            # half the attention head size is usually a good trade-off between82            # speed and memory83            slice_size = self.unet.config.attention_head_dim // 284        self.unet.set_attention_slice(slice_size)85 86    def disable_attention_slicing(self):87        self.enable_attention_slicing(None)88 89    def freeze_vae(self):90        set_requires_grad(self.vae, False)91 92    def unfreeze_vae(self):93        set_requires_grad(self.vae, True)94 95    def freeze_unet(self):96        set_requires_grad(self.unet, False)97 98    def unfreeze_unet(self):99        set_requires_grad(self.unet, True)100 101    @torch.enable_grad()102    def cond_fn(103        self,104        latents,105        timestep,106        index,107        text_embeddings,108        noise_pred_original,109        text_embeddings_clip,110        clip_guidance_scale,111        num_cutouts,112        use_cutouts=True,113    ):114        latents = latents.detach().requires_grad_()115 116        if isinstance(self.scheduler, LMSDiscreteScheduler):117            sigma = self.scheduler.sigmas[index]118            # the model input needs to be scaled to match the continuous ODE formulation in K-LMS119            latent_model_input = latents / ((sigma**2 + 1) ** 0.5)120        else:121            latent_model_input = latents122 123        # predict the noise residual124        noise_pred = self.unet(latent_model_input, timestep, encoder_hidden_states=text_embeddings).sample125 126        if isinstance(self.scheduler, PNDMScheduler):127            alpha_prod_t = self.scheduler.alphas_cumprod[timestep]128            beta_prod_t = 1 - alpha_prod_t129            # compute predicted original sample from predicted noise also called130            # "predicted x_0" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf131            pred_original_sample = (latents - beta_prod_t ** (0.5) * noise_pred) / alpha_prod_t ** (0.5)132 133            fac = torch.sqrt(beta_prod_t)134            sample = pred_original_sample * (fac) + latents * (1 - fac)135        elif isinstance(self.scheduler, LMSDiscreteScheduler):136            sigma = self.scheduler.sigmas[index]137            sample = latents - sigma * noise_pred138        else:139            raise ValueError(f"scheduler type {type(self.scheduler)} not supported")140 141        sample = 1 / 0.18215 * sample142        image = self.vae.decode(sample).sample143        image = (image / 2 + 0.5).clamp(0, 1)144 145        if use_cutouts:146            image = self.make_cutouts(image, num_cutouts)147        else:148            image = transforms.Resize(self.feature_extractor.size)(image)149        image = self.normalize(image).to(latents.dtype)150 151        image_embeddings_clip = self.clip_model.get_image_features(image)152        image_embeddings_clip = image_embeddings_clip / image_embeddings_clip.norm(p=2, dim=-1, keepdim=True)153 154        if use_cutouts:155            dists = spherical_dist_loss(image_embeddings_clip, text_embeddings_clip)156            dists = dists.view([num_cutouts, sample.shape[0], -1])157            loss = dists.sum(2).mean(0).sum() * clip_guidance_scale158        else:159            loss = spherical_dist_loss(image_embeddings_clip, text_embeddings_clip).mean() * clip_guidance_scale160 161        grads = -torch.autograd.grad(loss, latents)[0]162 163        if isinstance(self.scheduler, LMSDiscreteScheduler):164            latents = latents.detach() + grads * (sigma**2)165            noise_pred = noise_pred_original166        else:167            noise_pred = noise_pred_original - torch.sqrt(beta_prod_t) * grads168        return noise_pred, latents169 170    @torch.no_grad()171    def __call__(172        self,173        prompt: Union[str, List[str]],174        height: Optional[int] = 512,175        width: Optional[int] = 512,176        num_inference_steps: Optional[int] = 50,177        guidance_scale: Optional[float] = 7.5,178        num_images_per_prompt: Optional[int] = 1,179        clip_guidance_scale: Optional[float] = 100,180        clip_prompt: Optional[Union[str, List[str]]] = None,181        num_cutouts: Optional[int] = 4,182        use_cutouts: Optional[bool] = True,183        generator: Optional[torch.Generator] = None,184        latents: Optional[torch.FloatTensor] = None,185        output_type: Optional[str] = "pil",186        return_dict: bool = True,187    ):188        if isinstance(prompt, str):189            batch_size = 1190        elif isinstance(prompt, list):191            batch_size = len(prompt)192        else:193            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")194 195        if height % 8 != 0 or width % 8 != 0:196            raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")197 198        # get prompt text embeddings199        text_input = self.tokenizer(200            prompt,201            padding="max_length",202            max_length=self.tokenizer.model_max_length,203            truncation=True,204            return_tensors="pt",205        )206        text_embeddings = self.text_encoder(text_input.input_ids.to(self.device))[0]207        # duplicate text embeddings for each generation per prompt208        text_embeddings = text_embeddings.repeat_interleave(num_images_per_prompt, dim=0)209 210        if clip_guidance_scale > 0:211            if clip_prompt is not None:212                clip_text_input = self.tokenizer(213                    clip_prompt,214                    padding="max_length",215                    max_length=self.tokenizer.model_max_length,216                    truncation=True,217                    return_tensors="pt",218                ).input_ids.to(self.device)219            else:220                clip_text_input = text_input.input_ids.to(self.device)221            text_embeddings_clip = self.clip_model.get_text_features(clip_text_input)222            text_embeddings_clip = text_embeddings_clip / text_embeddings_clip.norm(p=2, dim=-1, keepdim=True)223            # duplicate text embeddings clip for each generation per prompt224            text_embeddings_clip = text_embeddings_clip.repeat_interleave(num_images_per_prompt, dim=0)225 226        # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)227        # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`228        # corresponds to doing no classifier free guidance.229        do_classifier_free_guidance = guidance_scale > 1.0230        # get unconditional embeddings for classifier free guidance231        if do_classifier_free_guidance:232            max_length = text_input.input_ids.shape[-1]233            uncond_input = self.tokenizer([""], padding="max_length", max_length=max_length, return_tensors="pt")234            uncond_embeddings = self.text_encoder(uncond_input.input_ids.to(self.device))[0]235            # duplicate unconditional embeddings for each generation per prompt236            uncond_embeddings = uncond_embeddings.repeat_interleave(num_images_per_prompt, dim=0)237 238            # For classifier free guidance, we need to do two forward passes.239            # Here we concatenate the unconditional and text embeddings into a single batch240            # to avoid doing two forward passes241            text_embeddings = torch.cat([uncond_embeddings, text_embeddings])242 243        # get the initial random noise unless the user supplied it244 245        # Unlike in other pipelines, latents need to be generated in the target device246        # for 1-to-1 results reproducibility with the CompVis implementation.247        # However this currently doesn't work in `mps`.248        latents_shape = (batch_size * num_images_per_prompt, self.unet.in_channels, height // 8, width // 8)249        latents_dtype = text_embeddings.dtype250        if latents is None:251            if self.device.type == "mps":252                # randn does not exist on mps253                latents = torch.randn(latents_shape, generator=generator, device="cpu", dtype=latents_dtype).to(254                    self.device255                )256            else:257                latents = torch.randn(latents_shape, generator=generator, device=self.device, dtype=latents_dtype)258        else:259            if latents.shape != latents_shape:260                raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {latents_shape}")261            latents = latents.to(self.device)262 263        # set timesteps264        accepts_offset = "offset" in set(inspect.signature(self.scheduler.set_timesteps).parameters.keys())265        extra_set_kwargs = {}266        if accepts_offset:267            extra_set_kwargs["offset"] = 1268 269        self.scheduler.set_timesteps(num_inference_steps, **extra_set_kwargs)270 271        # Some schedulers like PNDM have timesteps as arrays272        # It's more optimized to move all timesteps to correct device beforehand273        timesteps_tensor = self.scheduler.timesteps.to(self.device)274 275        # scale the initial noise by the standard deviation required by the scheduler276        latents = latents * self.scheduler.init_noise_sigma277 278        for i, t in enumerate(self.progress_bar(timesteps_tensor)):279            # expand the latents if we are doing classifier free guidance280            latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents281            latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)282 283            # predict the noise residual284            noise_pred = self.unet(latent_model_input, t, encoder_hidden_states=text_embeddings).sample285 286            # perform classifier free guidance287            if do_classifier_free_guidance:288                noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)289                noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)290 291            # perform clip guidance292            if clip_guidance_scale > 0:293                text_embeddings_for_guidance = (294                    text_embeddings.chunk(2)[1] if do_classifier_free_guidance else text_embeddings295                )296                noise_pred, latents = self.cond_fn(297                    latents,298                    t,299                    i,300                    text_embeddings_for_guidance,301                    noise_pred,302                    text_embeddings_clip,303                    clip_guidance_scale,304                    num_cutouts,305                    use_cutouts,306                )307 308            # compute the previous noisy sample x_t -> x_t-1309            latents = self.scheduler.step(noise_pred, t, latents).prev_sample310 311        # scale and decode the image latents with vae312        latents = 1 / 0.18215 * latents313        image = self.vae.decode(latents).sample314 315        image = (image / 2 + 0.5).clamp(0, 1)316        image = image.cpu().permute(0, 2, 3, 1).numpy()317 318        if output_type == "pil":319            image = self.numpy_to_pil(image)320 321        if not return_dict:322            return (image, None)323 324        return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=None)325