1 files changed, 457 insertions, 0 deletions

diff --git a/pipelines/stable_diffusion/clip_guided_stable_diffusion.py b/pipelines/stable_diffusion/clip_guided_stable_diffusion.py
new file mode 100644
index 0000000..306d9a9
--- /dev/null
+++ b/pipelines/stable_diffusion/clip_guided_stable_diffusion.py
@@ -0,0 +1,457 @@
+import inspect
+import warnings
+from typing import List, Optional, Union
+
+import torch
+from torch import nn
+from torch.nn import functional as F
+
+from diffusers.configuration_utils import FrozenDict
+from diffusers import AutoencoderKL, DiffusionPipeline, DDIMScheduler, LMSDiscreteScheduler, PNDMScheduler, UNet2DConditionModel
+from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion import StableDiffusionPipelineOutput
+from diffusers.utils import logging
+from torchvision import transforms
+from transformers import CLIPFeatureExtractor, CLIPModel, CLIPTextModel, CLIPTokenizer
+from schedulers.scheduling_euler_a import EulerAScheduler, CFGDenoiserForward
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+class MakeCutouts(nn.Module):
+    def __init__(self, cut_size, cut_power=1.0):
+        super().__init__()
+
+        self.cut_size = cut_size
+        self.cut_power = cut_power
+
+    def forward(self, pixel_values, num_cutouts):
+        sideY, sideX = pixel_values.shape[2:4]
+        max_size = min(sideX, sideY)
+        min_size = min(sideX, sideY, self.cut_size)
+        cutouts = []
+        for _ in range(num_cutouts):
+            size = int(torch.rand([]) ** self.cut_power * (max_size - min_size) + min_size)
+            offsetx = torch.randint(0, sideX - size + 1, ())
+            offsety = torch.randint(0, sideY - size + 1, ())
+            cutout = pixel_values[:, :, offsety: offsety + size, offsetx: offsetx + size]
+            cutouts.append(F.adaptive_avg_pool2d(cutout, self.cut_size))
+        return torch.cat(cutouts)
+
+
+def spherical_dist_loss(x, y):
+    x = F.normalize(x, dim=-1)
+    y = F.normalize(y, dim=-1)
+    return (x - y).norm(dim=-1).div(2).arcsin().pow(2).mul(2)
+
+
+def set_requires_grad(model, value):
+    for param in model.parameters():
+        param.requires_grad = value
+
+
+class CLIPGuidedStableDiffusion(DiffusionPipeline):
+    """CLIP guided stable diffusion based on the amazing repo by @crowsonkb and @Jack000
+    - https://github.com/Jack000/glid-3-xl
+    - https://github.dev/crowsonkb/k-diffusion
+    """
+
+    def __init__(
+        self,
+        vae: AutoencoderKL,
+        text_encoder: CLIPTextModel,
+        clip_model: CLIPModel,
+        tokenizer: CLIPTokenizer,
+        unet: UNet2DConditionModel,
+        scheduler: Union[DDIMScheduler, PNDMScheduler, LMSDiscreteScheduler],
+        feature_extractor: CLIPFeatureExtractor,
+        **kwargs,
+    ):
+        super().__init__()
+
+        if hasattr(scheduler.config, "steps_offset") and scheduler.config.steps_offset != 1:
+            warnings.warn(
+                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",
+                DeprecationWarning,
+            )
+            new_config = dict(scheduler.config)
+            new_config["steps_offset"] = 1
+            scheduler._internal_dict = FrozenDict(new_config)
+
+        self.register_modules(
+            vae=vae,
+            text_encoder=text_encoder,
+            clip_model=clip_model,
+            tokenizer=tokenizer,
+            unet=unet,
+            scheduler=scheduler,
+            feature_extractor=feature_extractor,
+        )
+
+        self.normalize = transforms.Normalize(mean=feature_extractor.image_mean, std=feature_extractor.image_std)
+        self.make_cutouts = MakeCutouts(feature_extractor.size)
+
+        set_requires_grad(self.text_encoder, False)
+        set_requires_grad(self.clip_model, False)
+
+    def enable_attention_slicing(self, slice_size: Optional[Union[str, int]] = "auto"):
+        r"""
+        Enable sliced attention computation.
+
+        When this option is enabled, the attention module will split the input tensor in slices, to compute attention
+        in several steps. This is useful to save some memory in exchange for a small speed decrease.
+
+        Args:
+            slice_size (`str` or `int`, *optional*, defaults to `"auto"`):
+                When `"auto"`, halves the input to the attention heads, so attention will be computed in two steps. If
+                a number is provided, uses as many slices as `attention_head_dim // slice_size`. In this case,
+                `attention_head_dim` must be a multiple of `slice_size`.
+        """
+        if slice_size == "auto":
+            # half the attention head size is usually a good trade-off between
+            # speed and memory
+            slice_size = self.unet.config.attention_head_dim // 2
+        self.unet.set_attention_slice(slice_size)
+
+    def disable_attention_slicing(self):
+        r"""
+        Disable sliced attention computation. If `enable_attention_slicing` was previously invoked, this method will go
+        back to computing attention in one step.
+        """
+        # set slice_size = `None` to disable `attention slicing`
+        self.enable_attention_slicing(None)
+
+    def freeze_vae(self):
+        set_requires_grad(self.vae, False)
+
+    def unfreeze_vae(self):
+        set_requires_grad(self.vae, True)
+
+    def freeze_unet(self):
+        set_requires_grad(self.unet, False)
+
+    def unfreeze_unet(self):
+        set_requires_grad(self.unet, True)
+
+    @torch.enable_grad()
+    def cond_fn(
+        self,
+        latents,
+        timestep,
+        index,
+        text_embeddings,
+        noise_pred_original,
+        text_embeddings_clip,
+        clip_guidance_scale,
+        num_cutouts,
+        use_cutouts=True,
+    ):
+        latents = latents.detach().requires_grad_()
+
+        if isinstance(self.scheduler, LMSDiscreteScheduler):
+            sigma = self.scheduler.sigmas[index]
+            # the model input needs to be scaled to match the continuous ODE formulation in K-LMS
+            latent_model_input = latents / ((sigma**2 + 1) ** 0.5)
+        else:
+            latent_model_input = latents
+
+        # predict the noise residual
+        noise_pred = self.unet(latent_model_input, timestep, encoder_hidden_states=text_embeddings).sample
+
+        if isinstance(self.scheduler, PNDMScheduler):
+            alpha_prod_t = self.scheduler.alphas_cumprod[timestep]
+            beta_prod_t = 1 - alpha_prod_t
+            # 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)
+
+            fac = torch.sqrt(beta_prod_t)
+            sample = pred_original_sample * (fac) + latents * (1 - fac)
+        elif isinstance(self.scheduler, LMSDiscreteScheduler):
+            sigma = self.scheduler.sigmas[index]
+            sample = latents - sigma * noise_pred
+        else:
+            raise ValueError(f"scheduler type {type(self.scheduler)} not supported")
+
+        sample = 1 / 0.18215 * sample
+        image = self.vae.decode(sample).sample
+        image = (image / 2 + 0.5).clamp(0, 1)
+
+        if use_cutouts:
+            image = self.make_cutouts(image, num_cutouts)
+        else:
+            image = transforms.Resize(self.feature_extractor.size)(image)
+        image = self.normalize(image)
+
+        image_embeddings_clip = self.clip_model.get_image_features(image).float()
+        image_embeddings_clip = image_embeddings_clip / image_embeddings_clip.norm(p=2, dim=-1, keepdim=True)
+
+        if use_cutouts:
+            dists = spherical_dist_loss(image_embeddings_clip, text_embeddings_clip)
+            dists = dists.view([num_cutouts, sample.shape[0], -1])
+            loss = dists.sum(2).mean(0).sum() * clip_guidance_scale
+        else:
+            loss = spherical_dist_loss(image_embeddings_clip, text_embeddings_clip).mean() * clip_guidance_scale
+
+        grads = -torch.autograd.grad(loss, latents)[0]
+
+        if isinstance(self.scheduler, LMSDiscreteScheduler):
+            latents = latents.detach() + grads * (sigma**2)
+            noise_pred = noise_pred_original
+        else:
+            noise_pred = noise_pred_original - torch.sqrt(beta_prod_t) * grads
+        return noise_pred, latents
+
+    @torch.no_grad()
+    def __call__(
+        self,
+        prompt: Union[str, List[str]],
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        height: Optional[int] = 512,
+        width: Optional[int] = 512,
+        num_inference_steps: Optional[int] = 50,
+        guidance_scale: Optional[float] = 7.5,
+        eta: Optional[float] = 0.0,
+        clip_guidance_scale: Optional[float] = 100,
+        clip_prompt: Optional[Union[str, List[str]]] = None,
+        num_cutouts: Optional[int] = 4,
+        use_cutouts: Optional[bool] = True,
+        generator: Optional[torch.Generator] = None,
+        latents: Optional[torch.FloatTensor] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+    ):
+        r"""
+        Function invoked when calling the pipeline for generation.
+
+        Args:
+            prompt (`str` or `List[str]`):
+                The prompt or prompts to guide the image generation.
+            height (`int`, *optional*, defaults to 512):
+                The height in pixels of the generated image.
+            width (`int`, *optional*, defaults to 512):
+                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.
+            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.
+            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`, *optional*):
+                A [torch generator](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.StableDiffusionPipelineOutput`] instead of a
+                plain tuple.
+
+        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`.
+        """
+
+        if isinstance(prompt, str):
+            batch_size = 1
+        elif isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+        if negative_prompt is None:
+            negative_prompt = [""] * batch_size
+        elif isinstance(negative_prompt, str):
+            negative_prompt = [negative_prompt] * batch_size
+        elif isinstance(negative_prompt, list):
+            if len(negative_prompt) != batch_size:
+                raise ValueError(
+                    f"`prompt` and `negative_prompt` have to be the same length, but are {len(prompt)} and {len(negative_prompt)}")
+        else:
+            raise ValueError(f"`negative_prompt` has to be of type `str` or `list` but is {type(negative_prompt)}")
+
+        if height % 8 != 0 or width % 8 != 0:
+            raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+        # get prompt text embeddings
+        text_inputs = self.tokenizer(
+            prompt,
+            padding="max_length",
+            max_length=self.tokenizer.model_max_length,
+            return_tensors="pt",
+        )
+        text_input_ids = text_inputs.input_ids
+
+        if text_input_ids.shape[-1] > self.tokenizer.model_max_length:
+            removed_text = self.tokenizer.batch_decode(text_input_ids[:, self.tokenizer.model_max_length:])
+            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}"
+            )
+            text_input_ids = text_input_ids[:, : self.tokenizer.model_max_length]
+        text_embeddings = self.text_encoder(text_input_ids.to(self.device))[0]
+
+        if clip_guidance_scale > 0:
+            if clip_prompt is not None:
+                clip_text_inputs = self.tokenizer(
+                    clip_prompt,
+                    padding="max_length",
+                    max_length=self.tokenizer.model_max_length,
+                    truncation=True,
+                    return_tensors="pt",
+                )
+                clip_text_input_ids = clip_text_inputs.input_ids
+            else:
+                clip_text_input_ids = text_input_ids
+            text_embeddings_clip = self.clip_model.get_text_features(clip_text_input_ids.to(self.device))
+            text_embeddings_clip = text_embeddings_clip / text_embeddings_clip.norm(p=2, dim=-1, keepdim=True)
+
+        # 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
+        # get unconditional embeddings for classifier free guidance
+        if do_classifier_free_guidance:
+            max_length = text_input_ids.shape[-1]
+            uncond_input = self.tokenizer(
+                negative_prompt, padding="max_length", max_length=max_length, return_tensors="pt"
+            )
+            uncond_embeddings = self.text_encoder(uncond_input.input_ids.to(self.device))[0]
+
+            # 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 = torch.cat([uncond_embeddings, text_embeddings])
+
+        # get the initial random noise unless the user supplied it
+
+        # Unlike in other pipelines, latents need to be generated in the target device
+        # for 1-to-1 results reproducibility with the CompVis implementation.
+        # However this currently doesn't work in `mps`.
+        latents_device = "cpu" if self.device.type == "mps" else self.device
+        latents_shape = (batch_size, self.unet.in_channels, height // 8, width // 8)
+        if latents is None:
+            latents = torch.randn(
+                latents_shape,
+                generator=generator,
+                device=latents_device,
+                dtype=text_embeddings.dtype,
+            )
+        else:
+            if latents.shape != latents_shape:
+                raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {latents_shape}")
+        latents = latents.to(self.device)
+
+        # set timesteps
+        self.scheduler.set_timesteps(num_inference_steps)
+
+        # Some schedulers like PNDM have timesteps as arrays
+        # It's more optimzed to move all timesteps to correct device beforehand
+        if torch.is_tensor(self.scheduler.timesteps):
+            timesteps_tensor = self.scheduler.timesteps.to(self.device)
+        else:
+            timesteps_tensor = torch.tensor(self.scheduler.timesteps.copy(), device=self.device)
+
+        # if we use LMSDiscreteScheduler, let's make sure latents are multiplied by sigmas
+        if isinstance(self.scheduler, LMSDiscreteScheduler):
+            latents = latents * self.scheduler.sigmas[0]
+        elif isinstance(self.scheduler, EulerAScheduler):
+            sigma = self.scheduler.timesteps[0]
+            latents = latents * sigma
+
+        # 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]
+        scheduler_step_args = set(inspect.signature(self.scheduler.step).parameters.keys())
+        accepts_eta = "eta" in scheduler_step_args
+        extra_step_kwargs = {}
+        if accepts_eta:
+            extra_step_kwargs["eta"] = eta
+        accepts_generator = "generator" in scheduler_step_args
+        if generator is not None and accepts_generator:
+            extra_step_kwargs["generator"] = generator
+
+        for i, t in enumerate(self.progress_bar(timesteps_tensor)):
+            # expand the latents if we are doing classifier free guidance
+            latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+            if isinstance(self.scheduler, LMSDiscreteScheduler):
+                sigma = self.scheduler.sigmas[i]
+                # the model input needs to be scaled to match the continuous ODE formulation in K-LMS
+                latent_model_input = latent_model_input / ((sigma**2 + 1) ** 0.5)
+
+            noise_pred = None
+            if isinstance(self.scheduler, EulerAScheduler):
+                sigma = t.reshape(1)
+                sigma_in = torch.cat([sigma] * 2)
+                # noise_pred = model(latent_model_input,sigma_in,uncond_embeddings, text_embeddings,guidance_scale)
+                noise_pred = CFGDenoiserForward(self.unet, latent_model_input, sigma_in,
+                                                text_embeddings, guidance_scale, DSsigmas=self.scheduler.DSsigmas)
+                # noise_pred = self.unet(latent_model_input, sigma_in, encoder_hidden_states=text_embeddings).sample
+            else:
+                # predict the noise residual
+                noise_pred = self.unet(latent_model_input, t, encoder_hidden_states=text_embeddings).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)
+
+            # perform clip guidance
+            if clip_guidance_scale > 0:
+                text_embeddings_for_guidance = (
+                    text_embeddings.chunk(2)[1] if do_classifier_free_guidance else text_embeddings
+                )
+                noise_pred, latents = self.cond_fn(
+                    latents,
+                    t,
+                    i,
+                    text_embeddings_for_guidance,
+                    noise_pred,
+                    text_embeddings_clip,
+                    clip_guidance_scale,
+                    num_cutouts,
+                    use_cutouts,
+                )
+
+            # compute the previous noisy sample x_t -> x_t-1
+            if isinstance(self.scheduler, LMSDiscreteScheduler):
+                latents = self.scheduler.step(noise_pred, i, latents, **extra_step_kwargs).prev_sample
+            elif isinstance(self.scheduler, EulerAScheduler):
+                if i < self.scheduler.timesteps.shape[0] - 1:  # avoid out of bound error
+                    t_prev = self.scheduler.timesteps[i+1]
+                    latents = self.scheduler.step(noise_pred, t, t_prev, latents, **extra_step_kwargs).prev_sample
+            else:
+                latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
+
+        # scale and decode the image latents with vae
+        latents = 1 / 0.18215 * latents
+        image = self.vae.decode(latents).sample
+
+        image = (image / 2 + 0.5).clamp(0, 1)
+        image = image.cpu().permute(0, 2, 3, 1).numpy()
+
+        if output_type == "pil":
+            image = self.numpy_to_pil(image)
+
+        if not return_dict:
+            return (image, None)
+
+        return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=None)