Update composable_stable_diffusion_pipeline.py

composable_stable_diffusion_pipeline.py

@@ -15,8 +15,60 @@ from diffusers.pipeline_utils import DiffusionPipeline
 from diffusers.schedulers import DDIMScheduler, LMSDiscreteScheduler, PNDMScheduler
 from safety_checker import StableDiffusionSafetyChecker
 
+from dataclasses import dataclass
+from typing import List, Union
+
+import numpy as np
+
+import PIL
+
+from diffusers.utils import BaseOutput
+
+
+@dataclass
+class StableDiffusionPipelineOutput(BaseOutput):
+    """
+    Output class for Stable Diffusion pipelines.
+    Args:
+        images (`List[PIL.Image.Image]` or `np.ndarray`)
+            List of denoised PIL images of length `batch_size` or numpy array of shape `(batch_size, height, width,
+            num_channels)`. PIL images or numpy array present the denoised images of the diffusion pipeline.
+        nsfw_content_detected (`List[bool]`)
+            List of flags denoting whether the corresponding generated image likely represents "not-safe-for-work"
+            (nsfw) content.
+    """
+
+    images: Union[List[PIL.Image.Image], np.ndarray]
+    nsfw_content_detected: List[bool]
 
 class ComposableStableDiffusionPipeline(DiffusionPipeline):
+    r"""
+    Pipeline for text-to-image generation using Stable Diffusion.
+
+    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 latens. Can be one of
+            [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+        safety_checker ([`StableDiffusionSafetyChecker`]):
+            Classification module that estimates whether generated images could be considered offsensive or harmful.
+            Please, refer to the [model card](https://huggingface.co/CompVis/stable-diffusion-v1-4) for details.
+        feature_extractor ([`CLIPFeatureExtractor`]):
+            Model that extracts features from generated images to be used as inputs for the `safety_checker`.
+    """
+
     def __init__(
         self,
         vae: AutoencoderKL,
@@ -39,6 +91,33 @@ class ComposableStableDiffusionPipeline(DiffusionPipeline):
             feature_extractor=feature_extractor,
         )
 
+    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)
+
     @torch.no_grad()
     def __call__(
         self,
@@ -49,9 +128,56 @@ class ComposableStableDiffusionPipeline(DiffusionPipeline):
         guidance_scale: Optional[float] = 7.5,
         eta: Optional[float] = 0.0,
         generator: Optional[torch.Generator] = None,
+        latents: Optional[torch.FloatTensor] = None,
         output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        weights: Optional[str] = "",
         **kwargs,
     ):
+        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 `nd.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 "torch_device" in kwargs:
             device = kwargs.pop("torch_device")
             warnings.warn(
@@ -76,7 +202,7 @@ class ComposableStableDiffusionPipeline(DiffusionPipeline):
 
         if '|' in prompt:
             prompt = [x.strip() for x in prompt.split('|')]
-            print(prompt)
+            print(f"composing {prompt}...")
 
         # get prompt text embeddings
         text_input = self.tokenizer(
@@ -88,6 +214,38 @@ class ComposableStableDiffusionPipeline(DiffusionPipeline):
         )
         text_embeddings = self.text_encoder(text_input.input_ids.to(self.device))[0]
 
+        if not weights:
+            # specify weights for prompts (excluding the unconditional score)
+            print('using equal weights for all prompts...')
+            pos_weights = torch.tensor([1 / (text_embeddings.shape[0] - 1)] * (text_embeddings.shape[0] - 1),
+                                       device=self.device).reshape(-1, 1, 1, 1)
+            neg_weights = torch.tensor([1.], device=self.device).reshape(-1, 1, 1, 1)
+            mask = torch.tensor([False] + [True] * pos_weights.shape[0], dtype=torch.bool)
+        else:
+            # set prompt weight for each
+            num_prompts = len(prompt) if isinstance(prompt, list) else 1
+            weights = [float(w.strip()) for w in weights.split("|")]
+            if len(weights) < num_prompts:
+                weights.append(1.)
+            weights = torch.tensor(weights, device=self.device)
+            assert len(weights) == text_embeddings.shape[0], "weights specified are not equal to the number of prompts"
+            pos_weights = []
+            neg_weights = []
+            mask = []  # first one is unconditional score
+            for w in weights:
+                if w > 0:
+                    pos_weights.append(w)
+                    mask.append(True)
+                else:
+                    neg_weights.append(abs(w))
+                    mask.append(False)
+            # normalize the weights
+            pos_weights = torch.tensor(pos_weights, device=self.device).reshape(-1, 1, 1, 1)
+            pos_weights = pos_weights / pos_weights.sum()
+            neg_weights = torch.tensor(neg_weights, device=self.device).reshape(-1, 1, 1, 1)
+            neg_weights = neg_weights / neg_weights.sum()
+            mask = torch.tensor(mask, device=self.device, dtype=torch.bool)
+
         # 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.
@@ -95,22 +253,40 @@ class ComposableStableDiffusionPipeline(DiffusionPipeline):
         # get unconditional embeddings for classifier free guidance
         if do_classifier_free_guidance:
             max_length = text_input.input_ids.shape[-1]
-            uncond_input = self.tokenizer(
-                [""] * batch_size, padding="max_length", max_length=max_length, return_tensors="pt"
+
+            if torch.all(mask):
+                # no negative prompts, so we use empty string as the negative prompt
+                uncond_input = self.tokenizer(
+                    [""] * batch_size, 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])
+
+                # update negative weights
+                neg_weights = torch.tensor([1.], device=self.device)
+                mask = torch.tensor([False] + mask.detach().tolist(), device=self.device, dtype=torch.bool)
+
+        # 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,
             )
-            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 intial random noise
-        latents = torch.randn(
-            (batch_size, self.unet.in_channels, height // 8, width // 8),
-            generator=generator,
-            device=self.device,
-        )
+        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
         accepts_offset = "offset" in set(inspect.signature(self.scheduler.set_timesteps).parameters.keys())
@@ -133,31 +309,38 @@ class ComposableStableDiffusionPipeline(DiffusionPipeline):
         if accepts_eta:
             extra_step_kwargs["eta"] = eta
 
-        for i, t in tqdm(enumerate(self.scheduler.timesteps)):
+        for i, t in enumerate(self.progress_bar(self.scheduler.timesteps)):
             # expand the latents if we are doing classifier free guidance
             latent_model_input = torch.cat([latents] * text_embeddings.shape[0]) 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)
 
+            # reduce memory by predicting each score sequentially
+            noise_preds = []
             # predict the noise residual
-            noise_pred = self.unet(latent_model_input, t, encoder_hidden_states=text_embeddings)["sample"]
+            for latent_in, text_embedding_in in zip(
+                    torch.chunk(latent_model_input, chunks=latent_model_input.shape[0], dim=0),
+                    torch.chunk(text_embeddings, chunks=text_embeddings.shape[0], dim=0)):
+                noise_preds.append(self.unet(latent_in, t, encoder_hidden_states=text_embedding_in).sample)
+            noise_preds = torch.cat(noise_preds, dim=0)
 
             # perform guidance
             if do_classifier_free_guidance:
-                pred_decomp = noise_pred.chunk(text_embeddings.shape[0])
-                noise_pred_uncond, noise_pred_text = pred_decomp[0], torch.cat(pred_decomp[1:], dim=0).mean(dim=0, keepdim=True)
+                noise_pred_uncond = noise_preds[~mask] * neg_weights
+                noise_pred_text = (noise_preds[mask] * pos_weights).sum(dim=0, keepdims=True)
                 noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
 
             # 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"]
+                latents = self.scheduler.step(noise_pred, i, latents, **extra_step_kwargs).prev_sample
             else:
-                latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs)["prev_sample"]
+                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)
+        image = self.vae.decode(latents).sample
 
         image = (image / 2 + 0.5).clamp(0, 1)
         image = image.cpu().permute(0, 2, 3, 1).numpy()
@@ -169,4 +352,7 @@ class ComposableStableDiffusionPipeline(DiffusionPipeline):
         if output_type == "pil":
             image = self.numpy_to_pil(image)
 
-        return {"sample": image, "nsfw_content_detected": has_nsfw_concept}
+        if not return_dict:
+            return (image, has_nsfw_concept)
+
+        return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)