from diffusers import QwenImageEditPipeline import inspect import math from typing import Any, Callable, Dict, List, Optional, Union from PIL import Image import numpy as np import torch from transformers import Qwen2_5_VLForConditionalGeneration, Qwen2Tokenizer, Qwen2VLProcessor from diffusers.utils import is_torch_xla_available, logging, replace_example_docstring from diffusers.image_processor import PipelineImageInput, VaeImageProcessor from diffusers.loaders import QwenImageLoraLoaderMixin from diffusers.models import AutoencoderKLQwenImage, QwenImageTransformer2DModel from diffusers.schedulers import FlowMatchEulerDiscreteScheduler from diffusers.utils import is_torch_xla_available, logging, replace_example_docstring from diffusers.utils.torch_utils import randn_tensor from diffusers.pipelines.pipeline_utils import DiffusionPipeline from diffusers.pipelines.qwenimage.pipeline_output import QwenImagePipelineOutput from diffusers.pipelines.qwenimage.pipeline_qwenimage_edit import * from models.model import Qwen2_5_VLForConditionalGeneration_Quant, Qwen2_5_VL_Quant from transformers.modeling_utils import PreTrainedModel from typing import Union, List, Optional, Any import torch from typing import Any, Callable, Dict, List, Optional, Union from functools import partial import torch import numpy as np from PIL import Image from .utils import retrieve_raw_timesteps from .lakonlab.pipelines.piflow_loader import PiFlowLoaderMixin from .lakonlab.models.diffusions.piflow_policies.dx import DXPolicy from .lakonlab.models.diffusions.piflow_policies.gmflow import GMFlowPolicy POLICY_CLASSES = dict( DX=DXPolicy, GMFlow=GMFlowPolicy ) class CoTylePipeline(QwenImageEditPipeline): def _get_qwen_prompt_embeds( self, prompt: Union[str, List[str]] = None, image: Optional[torch.Tensor] = None, device: Optional[torch.device] = None, dtype: Optional[torch.dtype] = None, codebook_id: Any = None, ): device = device or self._execution_device dtype = dtype or self.text_encoder.dtype prompt = [prompt] if isinstance(prompt, str) else prompt template = self.prompt_template_encode drop_idx = self.prompt_template_encode_start_idx txt = [template.format(e) for e in prompt] model_inputs = self.processor( text=txt, images=image, padding=True, return_tensors="pt", ).to(device) outputs = self.text_encoder( input_ids=model_inputs.input_ids, attention_mask=model_inputs.attention_mask, pixel_values=model_inputs.pixel_values, image_grid_thw=model_inputs.image_grid_thw, output_hidden_states=True, codebook_id=codebook_id, ) hidden_states = outputs.hidden_states[-1] split_hidden_states = self._extract_masked_hidden(hidden_states, model_inputs.attention_mask) split_hidden_states = [e[drop_idx:] for e in split_hidden_states] attn_mask_list = [torch.ones(e.size(0), dtype=torch.long, device=e.device) for e in split_hidden_states] max_seq_len = max([e.size(0) for e in split_hidden_states]) prompt_embeds = torch.stack( [torch.cat([u, u.new_zeros(max_seq_len - u.size(0), u.size(1))]) for u in split_hidden_states] ) encoder_attention_mask = torch.stack( [torch.cat([u, u.new_zeros(max_seq_len - u.size(0))]) for u in attn_mask_list] ) prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) return prompt_embeds, encoder_attention_mask def encode_prompt( self, prompt: Union[str, List[str]], image: Optional[torch.Tensor] = None, device: Optional[torch.device] = None, num_images_per_prompt: int = 1, prompt_embeds: Optional[torch.Tensor] = None, prompt_embeds_mask: Optional[torch.Tensor] = None, max_sequence_length: int = 1024, codebook_id: Any = None, ): r""" Args: prompt (`str` or `List[str]`, *optional*): prompt to be encoded image (`torch.Tensor`, *optional*): image to be encoded device: (`torch.device`): torch device num_images_per_prompt (`int`): number of images that should be generated per prompt prompt_embeds (`torch.Tensor`, *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. """ device = device or self._execution_device prompt = [prompt] if isinstance(prompt, str) else prompt batch_size = len(prompt) if prompt_embeds is None else prompt_embeds.shape[0] if prompt_embeds is None: prompt_embeds, prompt_embeds_mask = self._get_qwen_prompt_embeds(prompt, image, device, codebook_id=codebook_id) _, seq_len, _ = prompt_embeds.shape prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) prompt_embeds_mask = prompt_embeds_mask.repeat(1, num_images_per_prompt, 1) prompt_embeds_mask = prompt_embeds_mask.view(batch_size * num_images_per_prompt, seq_len) return prompt_embeds, prompt_embeds_mask @torch.no_grad() def __call__( self, image: Optional[PipelineImageInput] = None, prompt: Union[str, List[str]] = None, negative_prompt: Union[str, List[str]] = None, true_cfg_scale: float = 4.0, height: Optional[int] = None, width: Optional[int] = None, num_inference_steps: int = 50, sigmas: Optional[List[float]] = None, guidance_scale: float = 1.0, num_images_per_prompt: int = 1, generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, latents: Optional[torch.Tensor] = None, prompt_embeds: Optional[torch.Tensor] = None, prompt_embeds_mask: Optional[torch.Tensor] = None, negative_prompt_embeds: Optional[torch.Tensor] = None, negative_prompt_embeds_mask: Optional[torch.Tensor] = None, output_type: Optional[str] = "pil", return_dict: bool = True, attention_kwargs: Optional[Dict[str, Any]] = None, callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None, callback_on_step_end_tensor_inputs: List[str] = ["latents"], max_sequence_length: int = 512, **kwargs, ): r""" 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 `true_cfg_scale` is not greater than `1`). true_cfg_scale (`float`, *optional*, defaults to 1.0): When > 1.0 and a provided `negative_prompt`, enables true classifier-free guidance. height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): The height in pixels of the generated image. This is set to 1024 by default for the best results. width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): The width in pixels of the generated image. This is set to 1024 by default for the best results. 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. sigmas (`List[float]`, *optional*): Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed will be used. guidance_scale (`float`, *optional*, defaults to 3.5): Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. of [Imagen Paper](https://huggingface.co/papers/2205.11487). 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. This parameter in the pipeline is there to support future guidance-distilled models when they come up. Note that passing `guidance_scale` to the pipeline is ineffective. To enable classifier-free guidance, please pass `true_cfg_scale` and `negative_prompt` (even an empty negative prompt like " ") should enable classifier-free guidance computations. num_images_per_prompt (`int`, *optional*, defaults to 1): The number of images to generate per prompt. 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.Tensor`, *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 be generated by sampling using the supplied random `generator`. prompt_embeds (`torch.Tensor`, *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. negative_prompt_embeds (`torch.Tensor`, *optional*): Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input argument. 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.qwenimage.QwenImagePipelineOutput`] instead of a plain tuple. attention_kwargs (`dict`, *optional*): A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under `self.processor` in [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). callback_on_step_end (`Callable`, *optional*): A function that calls at the end of each denoising steps during the inference. The function is called with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by `callback_on_step_end_tensor_inputs`. callback_on_step_end_tensor_inputs (`List`, *optional*): The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the `._callback_tensor_inputs` attribute of your pipeline class. max_sequence_length (`int` defaults to 512): Maximum sequence length to use with the `prompt`. Examples: Returns: [`~pipelines.qwenimage.QwenImagePipelineOutput`] or `tuple`: [`~pipelines.qwenimage.QwenImagePipelineOutput`] if `return_dict` is True, otherwise a `tuple`. When returning a tuple, the first element is a list with the generated images. """ image_size = image[0].size if isinstance(image, list) else image.size calculated_width, calculated_height, _ = calculate_dimensions(1024 * 1024, image_size[0] / image_size[1]) height = height or calculated_height width = width or calculated_width multiple_of = self.vae_scale_factor * 2 width = width // multiple_of * multiple_of height = height // multiple_of * multiple_of # 1. Check inputs. Raise error if not correct self.check_inputs( prompt, height, width, negative_prompt=negative_prompt, prompt_embeds=prompt_embeds, negative_prompt_embeds=negative_prompt_embeds, prompt_embeds_mask=prompt_embeds_mask, negative_prompt_embeds_mask=negative_prompt_embeds_mask, callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, max_sequence_length=max_sequence_length, ) self._guidance_scale = guidance_scale self._attention_kwargs = attention_kwargs self._current_timestep = None self._interrupt = False # 2. Define call parameters 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.transformer.device # 3. Preprocess image if image is not None and not (isinstance(image, torch.Tensor) and image.size(1) == self.latent_channels): image = self.image_processor.resize(image, calculated_height, calculated_width) prompt_image = image image = self.image_processor.preprocess(image, calculated_height, calculated_width) image = image.unsqueeze(2) has_neg_prompt = negative_prompt is not None or ( negative_prompt_embeds is not None and negative_prompt_embeds_mask is not None ) # do_true_cfg = true_cfg_scale > 1 and has_neg_prompt prompt_embeds, prompt_embeds_mask = self.encode_prompt( image=prompt_image, prompt=prompt, prompt_embeds=prompt_embeds, prompt_embeds_mask=prompt_embeds_mask, device=device, num_images_per_prompt=num_images_per_prompt, max_sequence_length=max_sequence_length, codebook_id=kwargs.get('codebook_id'), ) # if do_true_cfg: cfg_img = Image.new("RGB", (392, 392), (0,0,0)) negative_prompt_embeds, negative_prompt_embeds_mask = self.encode_prompt( image=cfg_img, prompt=negative_prompt, prompt_embeds=negative_prompt_embeds, prompt_embeds_mask=negative_prompt_embeds_mask, device=device, num_images_per_prompt=num_images_per_prompt, max_sequence_length=max_sequence_length, codebook_id=None, ) # 4. Prepare latent variables num_channels_latents = self.transformer.config.in_channels // 4 latents, image_latents = self.prepare_latents( None, batch_size * num_images_per_prompt, num_channels_latents, height, width, prompt_embeds.dtype, device, generator, latents, ) # img_shapes = [ [ (1, height // self.vae_scale_factor // 2, width // self.vae_scale_factor // 2), # (1, calculated_height // self.vae_scale_factor // 2, calculated_width // self.vae_scale_factor // 2), ] ] * batch_size # 5. Prepare timesteps sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps) if sigmas is None else sigmas image_seq_len = latents.shape[1] mu = calculate_shift( image_seq_len, self.scheduler.config.get("base_image_seq_len", 256), self.scheduler.config.get("max_image_seq_len", 4096), self.scheduler.config.get("base_shift", 0.5), self.scheduler.config.get("max_shift", 1.15), ) timesteps, num_inference_steps = retrieve_timesteps( self.scheduler, num_inference_steps, device, sigmas=sigmas, mu=mu, ) num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) self._num_timesteps = len(timesteps) # handle guidance if self.transformer.config.guidance_embeds: guidance = torch.full([1], guidance_scale, device=device, dtype=torch.float32) guidance = guidance.expand(latents.shape[0]) else: guidance = None if self.attention_kwargs is None: self._attention_kwargs = {} txt_seq_lens = prompt_embeds_mask.sum(dim=1).tolist() if prompt_embeds_mask is not None else None negative_txt_seq_lens = ( negative_prompt_embeds_mask.sum(dim=1).tolist() if negative_prompt_embeds_mask is not None else None ) # 6. Denoising loop self.scheduler.set_begin_index(0) with self.progress_bar(total=num_inference_steps) as progress_bar: for i, t in enumerate(timesteps): if self.interrupt: continue self._current_timestep = t latent_model_input = latents if image_latents is not None: latent_model_input = torch.cat([latents, image_latents], dim=1) # broadcast to batch dimension in a way that's compatible with ONNX/Core ML timestep = t.expand(latents.shape[0]).to(latents.dtype) with self.transformer.cache_context("cond"): noise_pred = self.transformer( hidden_states=latent_model_input.to(dtype=self.transformer.dtype), timestep=timestep / 1000, guidance=guidance, encoder_hidden_states_mask=prompt_embeds_mask, encoder_hidden_states=prompt_embeds, img_shapes=img_shapes, txt_seq_lens=txt_seq_lens, attention_kwargs=self.attention_kwargs, return_dict=False, )[0] noise_pred = noise_pred[:, : latents.size(1)] if do_true_cfg: with self.transformer.cache_context("uncond"): neg_noise_pred = self.transformer( hidden_states=latent_model_input, timestep=timestep / 1000, guidance=guidance, encoder_hidden_states_mask=negative_prompt_embeds_mask, encoder_hidden_states=negative_prompt_embeds, img_shapes=img_shapes, txt_seq_lens=negative_txt_seq_lens, attention_kwargs=self.attention_kwargs, return_dict=False, )[0] neg_noise_pred = neg_noise_pred[:, : latents.size(1)] comb_pred = neg_noise_pred + true_cfg_scale * (noise_pred - neg_noise_pred) cond_norm = torch.norm(noise_pred, dim=-1, keepdim=True) noise_norm = torch.norm(comb_pred, dim=-1, keepdim=True) noise_pred = comb_pred * (cond_norm / noise_norm) # compute the previous noisy sample x_t -> x_t-1 latents_dtype = latents.dtype latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0] if latents.dtype != latents_dtype: if torch.backends.mps.is_available(): # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 latents = latents.to(latents_dtype) if callback_on_step_end is not None: callback_kwargs = {} for k in callback_on_step_end_tensor_inputs: callback_kwargs[k] = locals()[k] callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) latents = callback_outputs.pop("latents", latents) prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) # 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 XLA_AVAILABLE: xm.mark_step() self._current_timestep = None if output_type == "latent": image = latents else: latents = self._unpack_latents(latents, height, width, self.vae_scale_factor) latents = latents.to(self.vae.dtype) latents_mean = ( torch.tensor(self.vae.config.latents_mean) .view(1, self.vae.config.z_dim, 1, 1, 1) .to(latents.device, latents.dtype) ) latents_std = 1.0 / torch.tensor(self.vae.config.latents_std).view(1, self.vae.config.z_dim, 1, 1, 1).to( latents.device, latents.dtype ) latents = latents / latents_std + latents_mean image = self.vae.decode(latents, return_dict=False)[0][:, :, 0] 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 QwenImagePipelineOutput(images=image) class PiCoTylePipeline(QwenImageEditPipeline, PiFlowLoaderMixin): r""" Args: transformer ([`QwenImageTransformer2DModel`]): Conditional Transformer (MMDiT) architecture to denoise the encoded image latents. scheduler ([`FlowMatchEulerDiscreteScheduler`]): A scheduler to be used in combination with `transformer` to denoise the encoded image latents. vae ([`AutoencoderKL`]): Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. text_encoder ([`Qwen2.5-VL-7B-Instruct`]): Text encoder for processing prompts and images. tokenizer (`QwenTokenizer`): Tokenizer for text processing. processor (`Qwen2VLProcessor`): Processor for handling vision-language inputs. policy_type (`str`, *optional*, defaults to `"GMFlow"`): The type of flow policy to use. Currently supports `"GMFlow"` and `"DX"`. policy_kwargs (`Dict`, *optional*): Additional keyword arguments to pass to the policy class. """ def __init__( self, scheduler: FlowMatchEulerDiscreteScheduler, vae: AutoencoderKLQwenImage, text_encoder: Qwen2_5_VLForConditionalGeneration, tokenizer: Qwen2Tokenizer, processor: Qwen2VLProcessor, transformer: QwenImageTransformer2DModel, policy_type: str = 'GMFlow', policy_kwargs: Optional[Dict[str, Any]] = None, ): super().__init__( scheduler, vae, text_encoder, tokenizer, processor, transformer, ) assert policy_type in POLICY_CLASSES, f'Invalid policy: {policy_type}. Supported policies are {list(POLICY_CLASSES.keys())}.' self.policy_type = policy_type self.policy_class = partial( POLICY_CLASSES[policy_type], **policy_kwargs ) if policy_kwargs else POLICY_CLASSES[policy_type] def _unpack_gm(self, gm, height, width, num_channels_latents, patch_size=2, gm_patch_size=1): """Unpack Gaussian Mixture output for GMFlow policy.""" c = num_channels_latents * patch_size * patch_size h = (int(height) // (self.vae_scale_factor * patch_size)) w = (int(width) // (self.vae_scale_factor * patch_size)) bs = gm['means'].size(0) k = self.transformer.num_gaussians scale = patch_size // gm_patch_size gm['means'] = gm['means'].reshape( bs, h, w, k, c // (scale * scale), scale, scale ).permute( 0, 3, 4, 1, 5, 2, 6 ).reshape( bs, k, c // (scale * scale), h * scale, w * scale) gm['logweights'] = gm['logweights'].reshape( bs, h, w, k, 1, scale, scale ).permute( 0, 3, 4, 1, 5, 2, 6 ).reshape( bs, k, 1, h * scale, w * scale) gm['logstds'] = gm['logstds'].reshape(bs, 1, 1, 1, 1) return gm @staticmethod def _pack_latents(latents, batch_size, num_channels_latents, height, width, patch_size=1, target_patch_size=2): """Pack latents with configurable patch sizes.""" scale = target_patch_size // patch_size latents = latents.view( batch_size, num_channels_latents * patch_size * patch_size, height // target_patch_size, scale, width // target_patch_size, scale) latents = latents.permute(0, 2, 4, 1, 3, 5) latents = latents.reshape( batch_size, (height // target_patch_size) * (width // target_patch_size), num_channels_latents * target_patch_size * target_patch_size) return latents @staticmethod def _unpack_latents(latents, height, width, vae_scale_factor, patch_size=2, target_patch_size=1): """Unpack latents with configurable patch sizes.""" batch_size, num_patches, channels = latents.shape scale = patch_size // target_patch_size # VAE applies 8x compression on images but we must also account for packing which requires # latent height and width to be divisible by 2. height = (int(height) // (vae_scale_factor * patch_size)) width = (int(width) // (vae_scale_factor * patch_size)) latents = latents.view( batch_size, height, width, channels // (scale * scale), scale, scale) latents = latents.permute(0, 3, 1, 4, 2, 5) latents = latents.reshape(batch_size, channels // (scale * scale), height * scale, width * scale) return latents def _get_qwen_prompt_embeds( self, prompt: Union[str, List[str]] = None, image: Optional[torch.Tensor] = None, device: Optional[torch.device] = None, dtype: Optional[torch.dtype] = None, codebook_id: Any = None, ): """Override to support codebook_id parameter.""" device = device or self._execution_device dtype = dtype or self.text_encoder.dtype prompt = [prompt] if isinstance(prompt, str) else prompt template = self.prompt_template_encode drop_idx = self.prompt_template_encode_start_idx txt = [template.format(e) for e in prompt] model_inputs = self.processor( text=txt, images=image, padding=True, return_tensors="pt", ).to(device) outputs = self.text_encoder( input_ids=model_inputs.input_ids, attention_mask=model_inputs.attention_mask, pixel_values=model_inputs.pixel_values, image_grid_thw=model_inputs.image_grid_thw, output_hidden_states=True, codebook_id=codebook_id, ) hidden_states = outputs.hidden_states[-1] split_hidden_states = self._extract_masked_hidden(hidden_states, model_inputs.attention_mask) split_hidden_states = [e[drop_idx:] for e in split_hidden_states] attn_mask_list = [torch.ones(e.size(0), dtype=torch.long, device=e.device) for e in split_hidden_states] max_seq_len = max([e.size(0) for e in split_hidden_states]) prompt_embeds = torch.stack( [torch.cat([u, u.new_zeros(max_seq_len - u.size(0), u.size(1))]) for u in split_hidden_states] ) encoder_attention_mask = torch.stack( [torch.cat([u, u.new_zeros(max_seq_len - u.size(0))]) for u in attn_mask_list] ) prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) return prompt_embeds, encoder_attention_mask def encode_prompt( self, prompt: Union[str, List[str]], image: Optional[torch.Tensor] = None, device: Optional[torch.device] = None, num_images_per_prompt: int = 1, prompt_embeds: Optional[torch.Tensor] = None, prompt_embeds_mask: Optional[torch.Tensor] = None, max_sequence_length: int = 1024, codebook_id: Any = None, ): """Override to support codebook_id parameter.""" device = device or self._execution_device prompt = [prompt] if isinstance(prompt, str) else prompt batch_size = len(prompt) if prompt_embeds is None else prompt_embeds.shape[0] if prompt_embeds is None: prompt_embeds, prompt_embeds_mask = self._get_qwen_prompt_embeds( prompt, image, device, codebook_id=codebook_id ) _, seq_len, _ = prompt_embeds.shape prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) prompt_embeds_mask = prompt_embeds_mask.repeat(1, num_images_per_prompt, 1) prompt_embeds_mask = prompt_embeds_mask.view(batch_size * num_images_per_prompt, seq_len) return prompt_embeds, prompt_embeds_mask @torch.inference_mode() def __call__( self, image: Optional[PipelineImageInput] = None, prompt: Union[str, List[str]] = None, height: Optional[int] = None, width: Optional[int] = None, num_inference_steps: int = 4, total_substeps: int = 128, final_step_size_scale: float = 0.5, temperature: Union[float, str] = 'auto', num_images_per_prompt: Optional[int] = 1, generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, latents: Optional[torch.FloatTensor] = None, prompt_embeds: Optional[torch.FloatTensor] = None, prompt_embeds_mask: Optional[torch.Tensor] = None, output_type: Optional[str] = "pil", return_dict: bool = True, attention_kwargs: Optional[Dict[str, Any]] = None, callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None, callback_on_step_end_tensor_inputs: List[str] = ["latents"], max_sequence_length: int = 512, **kwargs, ): r""" Function invoked when calling the pipeline for policy-based image editing. Args: image (`PipelineImageInput`, *optional*): The input image to be edited. prompt (`str` or `List[str]`, *optional*): The prompt or prompts to guide the image editing. height (`int`, *optional*): The height in pixels of the generated image. width (`int`, *optional*): The width in pixels of the generated image. num_inference_steps (`int`, *optional*, defaults to 4): The number of denoising steps. total_substeps (`int`, *optional*, defaults to 128): The total number of substeps for policy-based flow integration. final_step_size_scale (`float`, *optional*, defaults to 0.5): The scale for the final step size. temperature (`float` or `"auto"`, *optional*, defaults to `"auto"`): The temperature parameter for the flow policy. num_images_per_prompt (`int`, *optional*, defaults to 1): The number of images to generate per prompt. generator (`torch.Generator` or `List[torch.Generator]`, *optional*): One or a list of torch generators to make generation deterministic. latents (`torch.FloatTensor`, *optional*): Pre-generated noisy latents. prompt_embeds (`torch.FloatTensor`, *optional*): Pre-generated text embeddings. prompt_embeds_mask (`torch.Tensor`, *optional*): Mask for prompt embeddings. output_type (`str`, *optional*, defaults to `"pil"`): The output format of the generated image. return_dict (`bool`, *optional*, defaults to `True`): Whether or not to return a QwenImagePipelineOutput instead of a plain tuple. attention_kwargs (`dict`, *optional*): Additional kwargs for attention processors. callback_on_step_end (`Callable`, *optional*): A function that calls at the end of each denoising step. callback_on_step_end_tensor_inputs (`List`, *optional*): The list of tensor inputs for the callback function. max_sequence_length (`int`, defaults to 512): Maximum sequence length to use with the prompt. Returns: [`~pipelines.qwenimage.QwenImagePipelineOutput`] or `tuple`: QwenImagePipelineOutput if `return_dict` is True, otherwise a tuple. """ # Calculate dimensions based on input image image_size = image[0].size if isinstance(image, list) else image.size calculated_width, calculated_height, _ = calculate_dimensions(1024 * 1024, image_size[0] / image_size[1]) height = height or calculated_height width = width or calculated_width multiple_of = self.vae_scale_factor * 2 width = width // multiple_of * multiple_of height = height // multiple_of * multiple_of # 1. Check inputs self.check_inputs( prompt, height, width, prompt_embeds=prompt_embeds, prompt_embeds_mask=prompt_embeds_mask, callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, max_sequence_length=max_sequence_length, ) self._attention_kwargs = attention_kwargs self._current_timestep = None self._interrupt = False # 2. Define call parameters 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.transformer.device # 3. Preprocess image if image is not None and not (isinstance(image, torch.Tensor) and image.size(1) == self.latent_channels): image = self.image_processor.resize(image, calculated_height, calculated_width) prompt_image = image image = self.image_processor.preprocess(image, calculated_height, calculated_width) image = image.unsqueeze(2) # 4. Encode prompt with image prompt_embeds, prompt_embeds_mask = self.encode_prompt( image=prompt_image, prompt=prompt, prompt_embeds=prompt_embeds, prompt_embeds_mask=prompt_embeds_mask, device=device, num_images_per_prompt=num_images_per_prompt, max_sequence_length=max_sequence_length, codebook_id=kwargs.get('codebook_id'), ) # 5. Prepare latent variables num_channels_latents = self.transformer.config.in_channels // 4 latents, image_latents = self.prepare_latents( None, batch_size * num_images_per_prompt, num_channels_latents, height, width, torch.bfloat16, device, generator, latents, ) img_shapes = [ [ (1, height // self.vae_scale_factor // 2, width // self.vae_scale_factor // 2), ] ] * batch_size # 6. Prepare timesteps for policy-based flow raw_timesteps, num_inference_substeps, total_substeps = retrieve_raw_timesteps( num_inference_steps, total_substeps, final_step_size_scale ) image_seq_len = latents.shape[1] mu = calculate_shift( image_seq_len, self.scheduler.config.get("base_image_seq_len", 256), self.scheduler.config.get("max_image_seq_len", 4096), self.scheduler.config.get("base_shift", 0.5), self.scheduler.config.get("max_shift", 1.15), ) timesteps, _ = retrieve_timesteps( self.scheduler, num_inference_steps, device, sigmas=raw_timesteps, mu=mu, ) assert len(timesteps) == total_substeps self._num_timesteps = total_substeps if self.attention_kwargs is None: self._attention_kwargs = {} txt_seq_lens = prompt_embeds_mask.sum(dim=1).tolist() if prompt_embeds_mask is not None else None # 7. Policy-based denoising loop self.scheduler.set_begin_index(0) timestep_id = 0 with self.progress_bar(total=num_inference_steps) as progress_bar: for i in range(num_inference_steps): if self.interrupt: continue t_src = timesteps[timestep_id] sigma_t_src = t_src / self.scheduler.config.num_train_timesteps is_final_step = i == (num_inference_steps - 1) self._current_timestep = t_src # Prepare model input with image latents latent_model_input = latents if image_latents is not None: latent_model_input = torch.cat([latents, image_latents], dim=1) with self.transformer.cache_context("cond"): denoising_output = self.transformer( hidden_states=latent_model_input.to(dtype=self.transformer.dtype), timestep=t_src.expand(latents.shape[0]) / 1000, encoder_hidden_states_mask=prompt_embeds_mask, encoder_hidden_states=prompt_embeds, img_shapes=img_shapes, txt_seq_lens=txt_seq_lens, attention_kwargs=self.attention_kwargs, ) # Extract only the latents part (not image_latents) if isinstance(denoising_output, tuple): denoising_output = denoising_output[0] if image_latents is not None: if isinstance(denoising_output, dict): # For GMFlow output for key in denoising_output: denoising_output[key] = denoising_output[key][:, :latents.size(1)] else: # For DX output denoising_output = denoising_output[:, :latents.size(1)] # Unpack and create policy latents = self._unpack_latents( latents, height, width, self.vae_scale_factor, target_patch_size=1 ) if self.policy_type == 'GMFlow': denoising_output = self._unpack_gm( denoising_output, height, width, num_channels_latents, gm_patch_size=1 ) denoising_output = {k: v.to(torch.float32) for k, v in denoising_output.items()} policy = self.policy_class( denoising_output, latents, sigma_t_src ) if not is_final_step: if temperature == 'auto': temperature = min(max(0.1 * (num_inference_steps - 1), 0), 1) else: assert isinstance(temperature, (float, int)) policy.temperature_(temperature) elif self.policy_type == 'DX': denoising_output = self._unpack_latents( denoising_output, height, width, self.vae_scale_factor, target_patch_size=1 ) denoising_output = denoising_output.reshape(latents.size(0), -1, *latents.shape[1:]) denoising_output = denoising_output.to(torch.float32) policy = self.policy_class( denoising_output, latents, sigma_t_src ) else: raise ValueError(f'Unknown policy type: {self.policy_type}.') # Compute the previous noisy sample x_t -> x_t-1 using policy for _ in range(num_inference_substeps[i]): t = timesteps[timestep_id] sigma_t = t / self.scheduler.config.num_train_timesteps u = policy.pi(latents, sigma_t) latents = self.scheduler.step(u, t, latents, return_dict=False)[0] timestep_id += 1 # Repack latents latents = self._pack_latents( latents, latents.size(0), num_channels_latents, 2 * (int(height) // (self.vae_scale_factor * 2)), 2 * (int(width) // (self.vae_scale_factor * 2)), patch_size=1 ) if callback_on_step_end is not None: callback_kwargs = {} for k in callback_on_step_end_tensor_inputs: callback_kwargs[k] = locals()[k] callback_outputs = callback_on_step_end(self, i, t_src, callback_kwargs) latents = callback_outputs.pop("latents", latents) prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) progress_bar.update() if XLA_AVAILABLE: xm.mark_step() self._current_timestep = None if output_type == "latent": image = latents else: latents = self._unpack_latents(latents, height, width, self.vae_scale_factor)[:, :, None] latents_mean = ( torch.tensor(self.vae.config.latents_mean) .view(1, self.vae.config.z_dim, 1, 1, 1) .to(latents.device, latents.dtype) ) latents_std = torch.tensor(self.vae.config.latents_std).view(1, self.vae.config.z_dim, 1, 1, 1).to( latents.device, latents.dtype ) # Note: multiply by std, not divide (key difference from original) latents = latents * latents_std + latents_mean image = self.vae.decode(latents.to(self.vae.dtype), return_dict=False)[0][:, :, 0] 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 QwenImagePipelineOutput(images=image)