import torch from diffusers import ImagePipelineOutput, PixArtAlphaPipeline, AutoencoderKL, Transformer2DModel, \ DPMSolverMultistepScheduler from diffusers.image_processor import VaeImageProcessor from diffusers.models.attention import BasicTransformerBlock from diffusers.models.embeddings import PixArtAlphaTextProjection, PatchEmbed from diffusers.models.normalization import AdaLayerNormSingle from diffusers.pipelines.pixart_alpha.pipeline_pixart_alpha import retrieve_timesteps from typing import Callable, List, Optional, Tuple, Union from diffusers.utils import deprecate from torch import nn from transformers import T5Tokenizer, T5EncoderModel ASPECT_RATIO_2048_BIN = { "0.25": [1024.0, 4096.0], "0.26": [1024.0, 3968.0], "0.27": [1024.0, 3840.0], "0.28": [1024.0, 3712.0], "0.32": [1152.0, 3584.0], "0.33": [1152.0, 3456.0], "0.35": [1152.0, 3328.0], "0.4": [1280.0, 3200.0], "0.42": [1280.0, 3072.0], "0.48": [1408.0, 2944.0], "0.5": [1408.0, 2816.0], "0.52": [1408.0, 2688.0], "0.57": [1536.0, 2688.0], "0.6": [1536.0, 2560.0], "0.68": [1664.0, 2432.0], "0.72": [1664.0, 2304.0], "0.78": [1792.0, 2304.0], "0.82": [1792.0, 2176.0], "0.88": [1920.0, 2176.0], "0.94": [1920.0, 2048.0], "1.0": [2048.0, 2048.0], "1.07": [2048.0, 1920.0], "1.13": [2176.0, 1920.0], "1.21": [2176.0, 1792.0], "1.29": [2304.0, 1792.0], "1.38": [2304.0, 1664.0], "1.46": [2432.0, 1664.0], "1.67": [2560.0, 1536.0], "1.75": [2688.0, 1536.0], "2.0": [2816.0, 1408.0], "2.09": [2944.0, 1408.0], "2.4": [3072.0, 1280.0], "2.5": [3200.0, 1280.0], "2.89": [3328.0, 1152.0], "3.0": [3456.0, 1152.0], "3.11": [3584.0, 1152.0], "3.62": [3712.0, 1024.0], "3.75": [3840.0, 1024.0], "3.88": [3968.0, 1024.0], "4.0": [4096.0, 1024.0] } ASPECT_RATIO_256_BIN = { "0.25": [128.0, 512.0], "0.28": [128.0, 464.0], "0.32": [144.0, 448.0], "0.33": [144.0, 432.0], "0.35": [144.0, 416.0], "0.4": [160.0, 400.0], "0.42": [160.0, 384.0], "0.48": [176.0, 368.0], "0.5": [176.0, 352.0], "0.52": [176.0, 336.0], "0.57": [192.0, 336.0], "0.6": [192.0, 320.0], "0.68": [208.0, 304.0], "0.72": [208.0, 288.0], "0.78": [224.0, 288.0], "0.82": [224.0, 272.0], "0.88": [240.0, 272.0], "0.94": [240.0, 256.0], "1.0": [256.0, 256.0], "1.07": [256.0, 240.0], "1.13": [272.0, 240.0], "1.21": [272.0, 224.0], "1.29": [288.0, 224.0], "1.38": [288.0, 208.0], "1.46": [304.0, 208.0], "1.67": [320.0, 192.0], "1.75": [336.0, 192.0], "2.0": [352.0, 176.0], "2.09": [368.0, 176.0], "2.4": [384.0, 160.0], "2.5": [400.0, 160.0], "3.0": [432.0, 144.0], "4.0": [512.0, 128.0] } ASPECT_RATIO_1024_BIN = { "0.25": [512.0, 2048.0], "0.28": [512.0, 1856.0], "0.32": [576.0, 1792.0], "0.33": [576.0, 1728.0], "0.35": [576.0, 1664.0], "0.4": [640.0, 1600.0], "0.42": [640.0, 1536.0], "0.48": [704.0, 1472.0], "0.5": [704.0, 1408.0], "0.52": [704.0, 1344.0], "0.57": [768.0, 1344.0], "0.6": [768.0, 1280.0], "0.68": [832.0, 1216.0], "0.72": [832.0, 1152.0], "0.78": [896.0, 1152.0], "0.82": [896.0, 1088.0], "0.88": [960.0, 1088.0], "0.94": [960.0, 1024.0], "1.0": [1024.0, 1024.0], "1.07": [1024.0, 960.0], "1.13": [1088.0, 960.0], "1.21": [1088.0, 896.0], "1.29": [1152.0, 896.0], "1.38": [1152.0, 832.0], "1.46": [1216.0, 832.0], "1.67": [1280.0, 768.0], "1.75": [1344.0, 768.0], "2.0": [1408.0, 704.0], "2.09": [1472.0, 704.0], "2.4": [1536.0, 640.0], "2.5": [1600.0, 640.0], "3.0": [1728.0, 576.0], "4.0": [2048.0, 512.0], } ASPECT_RATIO_512_BIN = { "0.25": [256.0, 1024.0], "0.28": [256.0, 928.0], "0.32": [288.0, 896.0], "0.33": [288.0, 864.0], "0.35": [288.0, 832.0], "0.4": [320.0, 800.0], "0.42": [320.0, 768.0], "0.48": [352.0, 736.0], "0.5": [352.0, 704.0], "0.52": [352.0, 672.0], "0.57": [384.0, 672.0], "0.6": [384.0, 640.0], "0.68": [416.0, 608.0], "0.72": [416.0, 576.0], "0.78": [448.0, 576.0], "0.82": [448.0, 544.0], "0.88": [480.0, 544.0], "0.94": [480.0, 512.0], "1.0": [512.0, 512.0], "1.07": [512.0, 480.0], "1.13": [544.0, 480.0], "1.21": [544.0, 448.0], "1.29": [576.0, 448.0], "1.38": [576.0, 416.0], "1.46": [608.0, 416.0], "1.67": [640.0, 384.0], "1.75": [672.0, 384.0], "2.0": [704.0, 352.0], "2.09": [736.0, 352.0], "2.4": [768.0, 320.0], "2.5": [800.0, 320.0], "3.0": [864.0, 288.0], "4.0": [1024.0, 256.0], } def pipeline_pixart_alpha_call( self, prompt: Union[str, List[str]] = None, negative_prompt: str = "", num_inference_steps: int = 20, timesteps: List[int] = None, guidance_scale: float = 4.5, num_images_per_prompt: Optional[int] = 1, height: Optional[int] = None, width: Optional[int] = None, eta: float = 0.0, generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, latents: Optional[torch.FloatTensor] = None, prompt_embeds: Optional[torch.FloatTensor] = None, prompt_attention_mask: Optional[torch.FloatTensor] = None, negative_prompt_embeds: Optional[torch.FloatTensor] = None, negative_prompt_attention_mask: Optional[torch.FloatTensor] = None, output_type: Optional[str] = "pil", return_dict: bool = True, callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None, callback_steps: int = 1, clean_caption: bool = True, use_resolution_binning: bool = True, max_sequence_length: int = 120, **kwargs, ) -> Union[ImagePipelineOutput, Tuple]: """ 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. 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`). num_inference_steps (`int`, *optional*, defaults to 100): The number of denoising steps. More denoising steps usually lead to a higher quality image at the expense of slower inference. timesteps (`List[int]`, *optional*): Custom timesteps to use for the denoising process. If not defined, equal spaced `num_inference_steps` timesteps are used. Must be in descending order. guidance_scale (`float`, *optional*, defaults to 4.5): 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. num_images_per_prompt (`int`, *optional*, defaults to 1): The number of images to generate per prompt. height (`int`, *optional*, defaults to self.unet.config.sample_size): The height in pixels of the generated image. width (`int`, *optional*, defaults to self.unet.config.sample_size): The width in pixels of the generated image. 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`. prompt_embeds (`torch.FloatTensor`, *optional*): Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not provided, text embeddings will be generated from `prompt` input argument. prompt_attention_mask (`torch.FloatTensor`, *optional*): Pre-generated attention mask for text embeddings. negative_prompt_embeds (`torch.FloatTensor`, *optional*): Pre-generated negative text embeddings. For PixArt-Alpha this negative prompt should be "". If not provided, negative_prompt_embeds will be generated from `negative_prompt` input argument. negative_prompt_attention_mask (`torch.FloatTensor`, *optional*): Pre-generated attention mask for negative text embeddings. 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.IFPipelineOutput`] 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. clean_caption (`bool`, *optional*, defaults to `True`): Whether or not to clean the caption before creating embeddings. Requires `beautifulsoup4` and `ftfy` to be installed. If the dependencies are not installed, the embeddings will be created from the raw prompt. use_resolution_binning (`bool` defaults to `True`): If set to `True`, the requested height and width are first mapped to the closest resolutions using `ASPECT_RATIO_1024_BIN`. After the produced latents are decoded into images, they are resized back to the requested resolution. Useful for generating non-square images. Examples: Returns: [`~pipelines.ImagePipelineOutput`] or `tuple`: If `return_dict` is `True`, [`~pipelines.ImagePipelineOutput`] is returned, otherwise a `tuple` is returned where the first element is a list with the generated images """ if "mask_feature" in kwargs: deprecation_message = "The use of `mask_feature` is deprecated. It is no longer used in any computation and that doesn't affect the end results. It will be removed in a future version." deprecate("mask_feature", "1.0.0", deprecation_message, standard_warn=False) # 1. Check inputs. Raise error if not correct height = height or self.transformer.config.sample_size * self.vae_scale_factor width = width or self.transformer.config.sample_size * self.vae_scale_factor if use_resolution_binning: if self.transformer.config.sample_size == 32: aspect_ratio_bin = ASPECT_RATIO_256_BIN elif self.transformer.config.sample_size == 64: aspect_ratio_bin = ASPECT_RATIO_512_BIN elif self.transformer.config.sample_size == 128: aspect_ratio_bin = ASPECT_RATIO_1024_BIN elif self.transformer.config.sample_size == 256: aspect_ratio_bin = ASPECT_RATIO_2048_BIN else: raise ValueError("Invalid sample size") orig_height, orig_width = height, width height, width = self.classify_height_width_bin(height, width, ratios=aspect_ratio_bin) self.check_inputs( prompt, height, width, negative_prompt, callback_steps, prompt_embeds, negative_prompt_embeds, prompt_attention_mask, negative_prompt_attention_mask, ) # 2. Default height and width to transformer if prompt is not None and isinstance(prompt, str): batch_size = 1 elif prompt is not None and isinstance(prompt, list): batch_size = len(prompt) else: batch_size = prompt_embeds.shape[0] 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 ( prompt_embeds, prompt_attention_mask, negative_prompt_embeds, negative_prompt_attention_mask, ) = self.encode_prompt( prompt, do_classifier_free_guidance, negative_prompt=negative_prompt, num_images_per_prompt=num_images_per_prompt, device=device, prompt_embeds=prompt_embeds, negative_prompt_embeds=negative_prompt_embeds, prompt_attention_mask=prompt_attention_mask, negative_prompt_attention_mask=negative_prompt_attention_mask, clean_caption=clean_caption, max_sequence_length=max_sequence_length, ) if do_classifier_free_guidance: prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0) prompt_attention_mask = torch.cat([negative_prompt_attention_mask, prompt_attention_mask], dim=0) # 4. Prepare timesteps timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, timesteps) # 5. Prepare latents. latent_channels = self.transformer.config.in_channels latents = self.prepare_latents( batch_size * num_images_per_prompt, latent_channels, height, width, prompt_embeds.dtype, device, generator, latents, ) # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) # 6.1 Prepare micro-conditions. added_cond_kwargs = {"resolution": None, "aspect_ratio": None} if self.transformer.config.sample_size == 128: resolution = torch.tensor([height, width]).repeat(batch_size * num_images_per_prompt, 1) aspect_ratio = torch.tensor([float(height / width)]).repeat(batch_size * num_images_per_prompt, 1) resolution = resolution.to(dtype=prompt_embeds.dtype, device=device) aspect_ratio = aspect_ratio.to(dtype=prompt_embeds.dtype, device=device) if do_classifier_free_guidance: resolution = torch.cat([resolution, resolution], dim=0) aspect_ratio = torch.cat([aspect_ratio, aspect_ratio], dim=0) added_cond_kwargs = {"resolution": resolution, "aspect_ratio": aspect_ratio} # 7. Denoising loop num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) with self.progress_bar(total=num_inference_steps) as progress_bar: for i, t in enumerate(timesteps): 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) current_timestep = t if not torch.is_tensor(current_timestep): # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can # This would be a good case for the `match` statement (Python 3.10+) is_mps = latent_model_input.device.type == "mps" if isinstance(current_timestep, float): dtype = torch.float32 if is_mps else torch.float64 else: dtype = torch.int32 if is_mps else torch.int64 current_timestep = torch.tensor([current_timestep], dtype=dtype, device=latent_model_input.device) elif len(current_timestep.shape) == 0: current_timestep = current_timestep[None].to(latent_model_input.device) # broadcast to batch dimension in a way that's compatible with ONNX/Core ML current_timestep = current_timestep.expand(latent_model_input.shape[0]) # predict noise model_output noise_pred = self.transformer( latent_model_input, encoder_hidden_states=prompt_embeds, encoder_attention_mask=prompt_attention_mask, timestep=current_timestep, added_cond_kwargs=added_cond_kwargs, return_dict=False, )[0] # 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) # learned sigma if self.transformer.config.out_channels // 2 == latent_channels: noise_pred = noise_pred.chunk(2, dim=1)[0] else: noise_pred = noise_pred # compute previous image: x_t -> x_t-1 if num_inference_steps == 1: latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).pred_original_sample else: latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] # 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: step_idx = i // getattr(self.scheduler, "order", 1) callback(step_idx, t, latents) if not output_type == "latent": image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0] if use_resolution_binning: image = self.resize_and_crop_tensor(image, orig_width, orig_height) else: image = latents if not output_type == "latent": image = self.image_processor.postprocess(image, output_type=output_type) # Offload all models self.maybe_free_model_hooks() if not return_dict: return (image,) return ImagePipelineOutput(images=image) class PixArtSigmaPipeline(PixArtAlphaPipeline): r""" tmp Pipeline for text-to-image generation using PixArt-Sigma. """ def __init__( self, tokenizer: T5Tokenizer, text_encoder: T5EncoderModel, vae: AutoencoderKL, transformer: Transformer2DModel, scheduler: DPMSolverMultistepScheduler, ): super().__init__(tokenizer, text_encoder, vae, transformer, scheduler) self.register_modules( tokenizer=tokenizer, text_encoder=text_encoder, vae=vae, transformer=transformer, scheduler=scheduler ) self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) def pixart_sigma_init_patched_inputs(self, norm_type): assert self.config.sample_size is not None, "Transformer2DModel over patched input must provide sample_size" self.height = self.config.sample_size self.width = self.config.sample_size self.patch_size = self.config.patch_size interpolation_scale = ( self.config.interpolation_scale if self.config.interpolation_scale is not None else max(self.config.sample_size // 64, 1) ) self.pos_embed = PatchEmbed( height=self.config.sample_size, width=self.config.sample_size, patch_size=self.config.patch_size, in_channels=self.in_channels, embed_dim=self.inner_dim, interpolation_scale=interpolation_scale, ) self.transformer_blocks = nn.ModuleList( [ BasicTransformerBlock( self.inner_dim, self.config.num_attention_heads, self.config.attention_head_dim, dropout=self.config.dropout, cross_attention_dim=self.config.cross_attention_dim, activation_fn=self.config.activation_fn, num_embeds_ada_norm=self.config.num_embeds_ada_norm, attention_bias=self.config.attention_bias, only_cross_attention=self.config.only_cross_attention, double_self_attention=self.config.double_self_attention, upcast_attention=self.config.upcast_attention, norm_type=norm_type, norm_elementwise_affine=self.config.norm_elementwise_affine, norm_eps=self.config.norm_eps, attention_type=self.config.attention_type, ) for _ in range(self.config.num_layers) ] ) if self.config.norm_type != "ada_norm_single": self.norm_out = nn.LayerNorm(self.inner_dim, elementwise_affine=False, eps=1e-6) self.proj_out_1 = nn.Linear(self.inner_dim, 2 * self.inner_dim) self.proj_out_2 = nn.Linear( self.inner_dim, self.config.patch_size * self.config.patch_size * self.out_channels ) elif self.config.norm_type == "ada_norm_single": self.norm_out = nn.LayerNorm(self.inner_dim, elementwise_affine=False, eps=1e-6) self.scale_shift_table = nn.Parameter(torch.randn(2, self.inner_dim) / self.inner_dim ** 0.5) self.proj_out = nn.Linear( self.inner_dim, self.config.patch_size * self.config.patch_size * self.out_channels ) # PixArt-Sigma blocks. self.adaln_single = None self.use_additional_conditions = False if self.config.norm_type == "ada_norm_single": # TODO(Sayak, PVP) clean this, PixArt-Sigma doesn't use additional_conditions anymore # additional conditions until we find better name self.adaln_single = AdaLayerNormSingle( self.inner_dim, use_additional_conditions=self.use_additional_conditions ) self.caption_projection = None if self.caption_channels is not None: self.caption_projection = PixArtAlphaTextProjection( in_features=self.caption_channels, hidden_size=self.inner_dim )