From da096a4999a02a34c14579383f0a8abb18fba2dd Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Nguy=E1=BB=85n=20Tr=E1=BB=8Dng=20Tu=E1=BA=A5n?= <119487916+Trgtuan10@users.noreply.github.com> Date: Wed, 13 Aug 2025 11:11:50 +0700 Subject: [PATCH 1/2] Add QwenImage Inpainting and Img2Img pipeline (#12117) * feat/qwenimage-img2img-inpaint * Update qwenimage.md to reflect new pipelines and add # Copied from convention * tiny fix for passing ruff check * reformat code * fix copied from statement * fix copied from statement * copy and style fix * fix dummies --------- Co-authored-by: TuanNT-ZenAI Co-authored-by: DN6 --- docs/source/en/api/pipelines/qwenimage.md | 12 + src/diffusers/__init__.py | 4 + src/diffusers/pipelines/__init__.py | 8 +- src/diffusers/pipelines/qwenimage/__init__.py | 4 + .../qwenimage/pipeline_qwenimage_img2img.py | 839 ++++++++++++++ .../qwenimage/pipeline_qwenimage_inpaint.py | 1025 +++++++++++++++++ .../dummy_torch_and_transformers_objects.py | 30 + .../qwenimage/test_qwenimage_img2img.py | 218 ++++ .../qwenimage/test_qwenimage_inpaint.py | 233 ++++ 9 files changed, 2371 insertions(+), 2 deletions(-) create mode 100644 src/diffusers/pipelines/qwenimage/pipeline_qwenimage_img2img.py create mode 100644 src/diffusers/pipelines/qwenimage/pipeline_qwenimage_inpaint.py create mode 100644 tests/pipelines/qwenimage/test_qwenimage_img2img.py create mode 100644 tests/pipelines/qwenimage/test_qwenimage_inpaint.py diff --git a/docs/source/en/api/pipelines/qwenimage.md b/docs/source/en/api/pipelines/qwenimage.md index 872e721049..557249f7a3 100644 --- a/docs/source/en/api/pipelines/qwenimage.md +++ b/docs/source/en/api/pipelines/qwenimage.md @@ -90,3 +90,15 @@ image.save("qwen_fewsteps.png") ## QwenImagePipelineOutput [[autodoc]] pipelines.qwenimage.pipeline_output.QwenImagePipelineOutput + +## QwenImageImg2ImgPipeline + +[[autodoc]] QwenImageImg2ImgPipeline + - all + - __call__ + +## QwenImageInpaintPipeline + +[[autodoc]] QwenImageInpaintPipeline + - all + - __call__ diff --git a/src/diffusers/__init__.py b/src/diffusers/__init__.py index 9e399f9d38..0053074bad 100644 --- a/src/diffusers/__init__.py +++ b/src/diffusers/__init__.py @@ -489,6 +489,8 @@ else: "PixArtAlphaPipeline", "PixArtSigmaPAGPipeline", "PixArtSigmaPipeline", + "QwenImageImg2ImgPipeline", + "QwenImageInpaintPipeline", "QwenImagePipeline", "ReduxImageEncoder", "SanaControlNetPipeline", @@ -1121,6 +1123,8 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: PixArtAlphaPipeline, PixArtSigmaPAGPipeline, PixArtSigmaPipeline, + QwenImageImg2ImgPipeline, + QwenImageInpaintPipeline, QwenImagePipeline, ReduxImageEncoder, SanaControlNetPipeline, diff --git a/src/diffusers/pipelines/__init__.py b/src/diffusers/pipelines/__init__.py index aab7664fd2..535b23dbb4 100644 --- a/src/diffusers/pipelines/__init__.py +++ b/src/diffusers/pipelines/__init__.py @@ -387,7 +387,11 @@ else: "SkyReelsV2ImageToVideoPipeline", "SkyReelsV2Pipeline", ] - _import_structure["qwenimage"] = ["QwenImagePipeline"] + _import_structure["qwenimage"] = [ + "QwenImagePipeline", + "QwenImageImg2ImgPipeline", + "QwenImageInpaintPipeline", + ] try: if not is_onnx_available(): raise OptionalDependencyNotAvailable() @@ -704,7 +708,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: from .paint_by_example import PaintByExamplePipeline from .pia import PIAPipeline from .pixart_alpha import PixArtAlphaPipeline, PixArtSigmaPipeline - from .qwenimage import QwenImagePipeline + from .qwenimage import QwenImageImg2ImgPipeline, QwenImageInpaintPipeline, QwenImagePipeline from .sana import SanaControlNetPipeline, SanaPipeline, SanaSprintImg2ImgPipeline, SanaSprintPipeline from .semantic_stable_diffusion import SemanticStableDiffusionPipeline from .shap_e import ShapEImg2ImgPipeline, ShapEPipeline diff --git a/src/diffusers/pipelines/qwenimage/__init__.py b/src/diffusers/pipelines/qwenimage/__init__.py index 963732ded0..64265880e7 100644 --- a/src/diffusers/pipelines/qwenimage/__init__.py +++ b/src/diffusers/pipelines/qwenimage/__init__.py @@ -24,6 +24,8 @@ except OptionalDependencyNotAvailable: else: _import_structure["modeling_qwenimage"] = ["ReduxImageEncoder"] _import_structure["pipeline_qwenimage"] = ["QwenImagePipeline"] + _import_structure["pipeline_qwenimage_img2img"] = ["QwenImageImg2ImgPipeline"] + _import_structure["pipeline_qwenimage_inpaint"] = ["QwenImageInpaintPipeline"] if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: try: @@ -33,6 +35,8 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: from ...utils.dummy_torch_and_transformers_objects import * # noqa F403 else: from .pipeline_qwenimage import QwenImagePipeline + from .pipeline_qwenimage_img2img import QwenImageImg2ImgPipeline + from .pipeline_qwenimage_inpaint import QwenImageInpaintPipeline else: import sys diff --git a/src/diffusers/pipelines/qwenimage/pipeline_qwenimage_img2img.py b/src/diffusers/pipelines/qwenimage/pipeline_qwenimage_img2img.py new file mode 100644 index 0000000000..4fc84a31cc --- /dev/null +++ b/src/diffusers/pipelines/qwenimage/pipeline_qwenimage_img2img.py @@ -0,0 +1,839 @@ +import inspect +from typing import Any, Callable, Dict, List, Optional, Union + +import numpy as np +import torch +from transformers import Qwen2_5_VLForConditionalGeneration, Qwen2Tokenizer + +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import QwenImageLoraLoaderMixin +from ...models import AutoencoderKLQwenImage, QwenImageTransformer2DModel +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import is_torch_xla_available, logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline +from .pipeline_output import QwenImagePipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import QwenImageImg2ImgPipeline + >>> from diffusers.utils import load_image + + >>> pipe = QwenImageImg2ImgPipeline.from_pretrained("Qwen/Qwen-Image", torch_dtype=torch.bfloat16) + >>> pipe = pipe.to("cuda") + >>> url = "https://raw.githubusercontent.com/CompVis/stable-diffusion/main/assets/stable-samples/img2img/sketch-mountains-input.jpg" + >>> init_image = load_image(url).resize((1024, 1024)) + >>> prompt = "cat wizard, gandalf, lord of the rings, detailed, fantasy, cute, adorable, Pixar, Disney" + >>> images = pipe(prompt=prompt, negative_prompt=" ", image=init_image, strength=0.95).images[0] + >>> images.save("qwenimage_img2img.png") + ``` +""" + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: Optional[torch.Generator] = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +# Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage.calculate_shift +def calculate_shift( + image_seq_len, + base_seq_len: int = 256, + max_seq_len: int = 4096, + base_shift: float = 0.5, + max_shift: float = 1.15, +): + m = (max_shift - base_shift) / (max_seq_len - base_seq_len) + b = base_shift - m * base_seq_len + mu = image_seq_len * m + b + return mu + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: Optional[int] = None, + device: Optional[Union[str, torch.device]] = None, + timesteps: Optional[List[int]] = None, + sigmas: Optional[List[float]] = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`List[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`List[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class QwenImageImg2ImgPipeline(DiffusionPipeline, QwenImageLoraLoaderMixin): + r""" + The QwenImage pipeline for text-to-image generation. + + 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`]): + [Qwen2.5-VL-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct), specifically the + [Qwen2.5-VL-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct) variant. + tokenizer (`QwenTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/en/model_doc/clip#transformers.CLIPTokenizer). + """ + + model_cpu_offload_seq = "text_encoder->transformer->vae" + _callback_tensor_inputs = ["latents", "prompt_embeds"] + + def __init__( + self, + scheduler: FlowMatchEulerDiscreteScheduler, + vae: AutoencoderKLQwenImage, + text_encoder: Qwen2_5_VLForConditionalGeneration, + tokenizer: Qwen2Tokenizer, + transformer: QwenImageTransformer2DModel, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + transformer=transformer, + scheduler=scheduler, + ) + self.vae_scale_factor = 2 ** len(self.vae.temperal_downsample) if getattr(self, "vae", None) else 8 + # QwenImage latents are turned into 2x2 patches and packed. This means the latent width and height has to be divisible + # by the patch size. So the vae scale factor is multiplied by the patch size to account for this + self.latent_channels = self.vae.config.z_dim if getattr(self, "vae", None) else 16 + self.image_processor = VaeImageProcessor( + vae_scale_factor=self.vae_scale_factor * 2, vae_latent_channels=self.latent_channels + ) + self.tokenizer_max_length = 1024 + self.prompt_template_encode = "<|im_start|>system\nDescribe the image by detailing the color, shape, size, texture, quantity, text, spatial relationships of the objects and background:<|im_end|>\n<|im_start|>user\n{}<|im_end|>\n<|im_start|>assistant\n" + self.prompt_template_encode_start_idx = 34 + self.default_sample_size = 128 + + # Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage.QwenImagePipeline._extract_masked_hidden + def _extract_masked_hidden(self, hidden_states: torch.Tensor, mask: torch.Tensor): + bool_mask = mask.bool() + valid_lengths = bool_mask.sum(dim=1) + selected = hidden_states[bool_mask] + split_result = torch.split(selected, valid_lengths.tolist(), dim=0) + + return split_result + + # Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage.QwenImagePipeline._get_qwen_prompt_embeds + def _get_qwen_prompt_embeds( + self, + prompt: Union[str, List[str]] = None, + device: Optional[torch.device] = None, + dtype: Optional[torch.dtype] = 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] + txt_tokens = self.tokenizer( + txt, max_length=self.tokenizer_max_length + drop_idx, padding=True, truncation=True, return_tensors="pt" + ).to(device) + encoder_hidden_states = self.text_encoder( + input_ids=txt_tokens.input_ids, + attention_mask=txt_tokens.attention_mask, + output_hidden_states=True, + ) + hidden_states = encoder_hidden_states.hidden_states[-1] + split_hidden_states = self._extract_masked_hidden(hidden_states, txt_tokens.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_vae_image(self, image: torch.Tensor, generator: torch.Generator): + if isinstance(generator, list): + image_latents = [ + retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i]) + for i in range(image.shape[0]) + ] + image_latents = torch.cat(image_latents, dim=0) + else: + image_latents = retrieve_latents(self.vae.encode(image), generator=generator) + + latents_mean = ( + torch.tensor(self.vae.config.latents_mean) + .view(1, self.vae.config.z_dim, 1, 1, 1) + .to(image_latents.device, image_latents.dtype) + ) + latents_std = 1.0 / torch.tensor(self.vae.config.latents_std).view(1, self.vae.config.z_dim, 1, 1, 1).to( + image_latents.device, image_latents.dtype + ) + + image_latents = (image_latents - latents_mean) * latents_std + + return image_latents + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3_img2img.StableDiffusion3Img2ImgPipeline.get_timesteps + def get_timesteps(self, num_inference_steps, strength, device): + # get the original timestep using init_timestep + init_timestep = min(num_inference_steps * strength, num_inference_steps) + + t_start = int(max(num_inference_steps - init_timestep, 0)) + timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :] + if hasattr(self.scheduler, "set_begin_index"): + self.scheduler.set_begin_index(t_start * self.scheduler.order) + + return timesteps, num_inference_steps - t_start + + # Copied fromCopied from diffusers.pipelines.qwenimage.pipeline_qwenimage.QwenImagePipeline.encode_prompt + def encode_prompt( + self, + prompt: Union[str, List[str]], + 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, + ): + r""" + + Args: + prompt (`str` or `List[str]`, *optional*): + prompt 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, device) + + _, 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 + + def check_inputs( + self, + prompt, + strength, + height, + width, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + prompt_embeds_mask=None, + negative_prompt_embeds_mask=None, + callback_on_step_end_tensor_inputs=None, + max_sequence_length=None, + ): + if strength < 0 or strength > 1: + raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}") + + if height % (self.vae_scale_factor * 2) != 0 or width % (self.vae_scale_factor * 2) != 0: + logger.warning( + f"`height` and `width` have to be divisible by {self.vae_scale_factor * 2} but are {height} and {width}. Dimensions will be resized accordingly" + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and prompt_embeds_mask is None: + raise ValueError( + "If `prompt_embeds` are provided, `prompt_embeds_mask` also have to be passed. Make sure to generate `prompt_embeds_mask` from the same text encoder that was used to generate `prompt_embeds`." + ) + if negative_prompt_embeds is not None and negative_prompt_embeds_mask is None: + raise ValueError( + "If `negative_prompt_embeds` are provided, `negative_prompt_embeds_mask` also have to be passed. Make sure to generate `negative_prompt_embeds_mask` from the same text encoder that was used to generate `negative_prompt_embeds`." + ) + + if max_sequence_length is not None and max_sequence_length > 1024: + raise ValueError(f"`max_sequence_length` cannot be greater than 1024 but is {max_sequence_length}") + + @staticmethod + # Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage.QwenImagePipeline._prepare_latent_image_ids + def _prepare_latent_image_ids(batch_size, height, width, device, dtype): + latent_image_ids = torch.zeros(height, width, 3) + latent_image_ids[..., 1] = latent_image_ids[..., 1] + torch.arange(height)[:, None] + latent_image_ids[..., 2] = latent_image_ids[..., 2] + torch.arange(width)[None, :] + + latent_image_id_height, latent_image_id_width, latent_image_id_channels = latent_image_ids.shape + + latent_image_ids = latent_image_ids.reshape( + latent_image_id_height * latent_image_id_width, latent_image_id_channels + ) + + return latent_image_ids.to(device=device, dtype=dtype) + + @staticmethod + # Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage.QwenImagePipeline._pack_latents + def _pack_latents(latents, batch_size, num_channels_latents, height, width): + latents = latents.view(batch_size, num_channels_latents, height // 2, 2, width // 2, 2) + latents = latents.permute(0, 2, 4, 1, 3, 5) + latents = latents.reshape(batch_size, (height // 2) * (width // 2), num_channels_latents * 4) + + return latents + + @staticmethod + # Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage.QwenImagePipeline._unpack_latents + def _unpack_latents(latents, height, width, vae_scale_factor): + batch_size, num_patches, channels = latents.shape + + # 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 = 2 * (int(height) // (vae_scale_factor * 2)) + width = 2 * (int(width) // (vae_scale_factor * 2)) + + latents = latents.view(batch_size, height // 2, width // 2, channels // 4, 2, 2) + latents = latents.permute(0, 3, 1, 4, 2, 5) + + latents = latents.reshape(batch_size, channels // (2 * 2), 1, height, width) + + return latents + + def enable_vae_slicing(self): + r""" + Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to + compute decoding in several steps. This is useful to save some memory and allow larger batch sizes. + """ + self.vae.enable_slicing() + + def disable_vae_slicing(self): + r""" + Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to + computing decoding in one step. + """ + self.vae.disable_slicing() + + def enable_vae_tiling(self): + r""" + Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to + compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow + processing larger images. + """ + self.vae.enable_tiling() + + def disable_vae_tiling(self): + r""" + Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to + computing decoding in one step. + """ + self.vae.disable_tiling() + + def prepare_latents( + self, + image, + timestep, + batch_size, + num_channels_latents, + height, + width, + dtype, + device, + generator, + latents=None, + ): + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + # 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 = 2 * (int(height) // (self.vae_scale_factor * 2)) + width = 2 * (int(width) // (self.vae_scale_factor * 2)) + + shape = (batch_size, 1, num_channels_latents, height, width) + + # If image is [B,C,H,W] -> add T=1. If it's already [B,C,T,H,W], leave it. + if image.dim() == 4: + image = image.unsqueeze(2) + elif image.dim() != 5: + raise ValueError(f"Expected image dims 4 or 5, got {image.dim()}.") + + if latents is not None: + latent_image_ids = self._prepare_latent_image_ids(batch_size, height // 2, width // 2, device, dtype) + return latents.to(device=device, dtype=dtype), latent_image_ids + + image = image.to(device=device, dtype=dtype) + if image.shape[1] != self.latent_channels: + image_latents = self._encode_vae_image(image=image, generator=generator) # [B,z,1,H',W'] + else: + image_latents = image + if batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] == 0: + # expand init_latents for batch_size + additional_image_per_prompt = batch_size // image_latents.shape[0] + image_latents = torch.cat([image_latents] * additional_image_per_prompt, dim=0) + elif batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image` of batch size {image_latents.shape[0]} to {batch_size} text prompts." + ) + else: + image_latents = torch.cat([image_latents], dim=0) + + image_latents = image_latents.transpose(1, 2) # [B,1,z,H',W'] + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + latents = self.scheduler.scale_noise(image_latents, timestep, noise) + latents = self._pack_latents(latents, batch_size, num_channels_latents, height, width) + + latent_image_ids = self._prepare_latent_image_ids(batch_size, height // 2, width // 2, device, dtype) + + return latents, latent_image_ids + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def attention_kwargs(self): + return self._attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def current_timestep(self): + return self._current_timestep + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: Union[str, List[str]] = None, + negative_prompt: Union[str, List[str]] = None, + true_cfg_scale: float = 4.0, + image: PipelineImageInput = None, + height: Optional[int] = None, + width: Optional[int] = None, + strength: float = 0.6, + 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, + ): + 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`). + image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.Tensor]`, `List[PIL.Image.Image]`, or `List[np.ndarray]`): + `Image`, numpy array or tensor representing an image batch to be used as the starting point. For both + numpy array and pytorch tensor, the expected value range is between `[0, 1]` If it's a tensor or a list + or tensors, the expected shape should be `(B, C, H, W)` or `(C, H, W)`. If it is a numpy array or a + list of arrays, the expected shape should be `(B, H, W, C)` or `(H, W, C)` It can also accept image + latents as `image`, but if passing latents directly it is not encoded again. + 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. + strength (`float`, *optional*, defaults to 1.0): + Indicates extent to transform the reference `image`. Must be between 0 and 1. `image` is used as a + starting point and more noise is added the higher the `strength`. The number of denoising steps depends + on the amount of noise initially added. When `strength` is 1, added noise is maximum and the denoising + process runs for the full number of iterations specified in `num_inference_steps`. A value of 1 + essentially ignores `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. + 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. + 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. + """ + + height = height or self.default_sample_size * self.vae_scale_factor + width = width or self.default_sample_size * self.vae_scale_factor + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + strength, + 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. Preprocess image + init_image = self.image_processor.preprocess(image, height=height, width=width) + init_image = init_image.to(dtype=torch.float32) + + # 3. 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._execution_device + + 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( + 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, + ) + if do_true_cfg: + negative_prompt_embeds, negative_prompt_embeds_mask = self.encode_prompt( + 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, + ) + + # 4. Prepare timesteps + sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps) if sigmas is None else sigmas + image_seq_len = (int(height) // self.vae_scale_factor // 2) * (int(width) // self.vae_scale_factor // 2) + 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, + ) + timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength, device) + if num_inference_steps < 1: + raise ValueError( + f"After adjusting the num_inference_steps by strength parameter: {strength}, the number of pipeline" + f"steps is {num_inference_steps} which is < 1 and not appropriate for this pipeline." + ) + latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt) + + # 5. Prepare latent variables + num_channels_latents = self.transformer.config.in_channels // 4 + latents, latent_image_ids = self.prepare_latents( + init_image, + latent_timestep, + 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)] * batch_size + + 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 + 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 + # 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=latents, + 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] + + if do_true_cfg: + with self.transformer.cache_context("uncond"): + neg_noise_pred = self.transformer( + hidden_states=latents, + 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] + 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) diff --git a/src/diffusers/pipelines/qwenimage/pipeline_qwenimage_inpaint.py b/src/diffusers/pipelines/qwenimage/pipeline_qwenimage_inpaint.py new file mode 100644 index 0000000000..5ffec0c447 --- /dev/null +++ b/src/diffusers/pipelines/qwenimage/pipeline_qwenimage_inpaint.py @@ -0,0 +1,1025 @@ +import inspect +from typing import Any, Callable, Dict, List, Optional, Union + +import numpy as np +import PIL.Image +import torch +from transformers import Qwen2_5_VLForConditionalGeneration, Qwen2Tokenizer + +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import QwenImageLoraLoaderMixin +from ...models import AutoencoderKLQwenImage, QwenImageTransformer2DModel +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import is_torch_xla_available, logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline +from .pipeline_output import QwenImagePipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import QwenImageInpaintPipeline + >>> from diffusers.utils import load_image + + >>> pipe = QwenImageInpaintPipeline.from_pretrained("Qwen/Qwen-Image", torch_dtype=torch.bfloat16) + >>> pipe.to("cuda") + >>> prompt = "Face of a yellow cat, high resolution, sitting on a park bench" + >>> img_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo.png" + >>> mask_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo_mask.png" + >>> source = load_image(img_url) + >>> mask = load_image(mask_url) + >>> image = pipe(prompt=prompt, negative_prompt=" ", image=source, mask_image=mask, strength=0.85).images[0] + >>> image.save("qwenimage_inpainting.png") + ``` +""" + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: Optional[torch.Generator] = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +# Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage.calculate_shift +def calculate_shift( + image_seq_len, + base_seq_len: int = 256, + max_seq_len: int = 4096, + base_shift: float = 0.5, + max_shift: float = 1.15, +): + m = (max_shift - base_shift) / (max_seq_len - base_seq_len) + b = base_shift - m * base_seq_len + mu = image_seq_len * m + b + return mu + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: Optional[int] = None, + device: Optional[Union[str, torch.device]] = None, + timesteps: Optional[List[int]] = None, + sigmas: Optional[List[float]] = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`List[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`List[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class QwenImageInpaintPipeline(DiffusionPipeline, QwenImageLoraLoaderMixin): + r""" + The QwenImage pipeline for text-to-image generation. + + 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`]): + [Qwen2.5-VL-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct), specifically the + [Qwen2.5-VL-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct) variant. + tokenizer (`QwenTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/en/model_doc/clip#transformers.CLIPTokenizer). + """ + + model_cpu_offload_seq = "text_encoder->transformer->vae" + _callback_tensor_inputs = ["latents", "prompt_embeds"] + + def __init__( + self, + scheduler: FlowMatchEulerDiscreteScheduler, + vae: AutoencoderKLQwenImage, + text_encoder: Qwen2_5_VLForConditionalGeneration, + tokenizer: Qwen2Tokenizer, + transformer: QwenImageTransformer2DModel, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + transformer=transformer, + scheduler=scheduler, + ) + self.vae_scale_factor = 2 ** len(self.vae.temperal_downsample) if getattr(self, "vae", None) else 8 + # QwenImage latents are turned into 2x2 patches and packed. This means the latent width and height has to be divisible + # by the patch size. So the vae scale factor is multiplied by the patch size to account for this + self.latent_channels = self.vae.config.z_dim if getattr(self, "vae", None) else 16 + self.image_processor = VaeImageProcessor( + vae_scale_factor=self.vae_scale_factor * 2, vae_latent_channels=self.latent_channels + ) + self.mask_processor = VaeImageProcessor( + vae_scale_factor=self.vae_scale_factor * 2, + vae_latent_channels=self.latent_channels, + do_normalize=False, + do_binarize=True, + do_convert_grayscale=True, + ) + self.tokenizer_max_length = 1024 + self.prompt_template_encode = "<|im_start|>system\nDescribe the image by detailing the color, shape, size, texture, quantity, text, spatial relationships of the objects and background:<|im_end|>\n<|im_start|>user\n{}<|im_end|>\n<|im_start|>assistant\n" + self.prompt_template_encode_start_idx = 34 + self.default_sample_size = 128 + + # Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage.QwenImagePipeline._extract_masked_hidden + def _extract_masked_hidden(self, hidden_states: torch.Tensor, mask: torch.Tensor): + bool_mask = mask.bool() + valid_lengths = bool_mask.sum(dim=1) + selected = hidden_states[bool_mask] + split_result = torch.split(selected, valid_lengths.tolist(), dim=0) + + return split_result + + # Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage.QwenImagePipeline._get_qwen_prompt_embeds + def _get_qwen_prompt_embeds( + self, + prompt: Union[str, List[str]] = None, + device: Optional[torch.device] = None, + dtype: Optional[torch.dtype] = 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] + txt_tokens = self.tokenizer( + txt, max_length=self.tokenizer_max_length + drop_idx, padding=True, truncation=True, return_tensors="pt" + ).to(device) + encoder_hidden_states = self.text_encoder( + input_ids=txt_tokens.input_ids, + attention_mask=txt_tokens.attention_mask, + output_hidden_states=True, + ) + hidden_states = encoder_hidden_states.hidden_states[-1] + split_hidden_states = self._extract_masked_hidden(hidden_states, txt_tokens.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 + + # Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage_img2img.QwenImageImg2ImgPipeline._encode_vae_image + def _encode_vae_image(self, image: torch.Tensor, generator: torch.Generator): + if isinstance(generator, list): + image_latents = [ + retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i]) + for i in range(image.shape[0]) + ] + image_latents = torch.cat(image_latents, dim=0) + else: + image_latents = retrieve_latents(self.vae.encode(image), generator=generator) + + latents_mean = ( + torch.tensor(self.vae.config.latents_mean) + .view(1, self.vae.config.z_dim, 1, 1, 1) + .to(image_latents.device, image_latents.dtype) + ) + latents_std = 1.0 / torch.tensor(self.vae.config.latents_std).view(1, self.vae.config.z_dim, 1, 1, 1).to( + image_latents.device, image_latents.dtype + ) + + image_latents = (image_latents - latents_mean) * latents_std + + return image_latents + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3_img2img.StableDiffusion3Img2ImgPipeline.get_timesteps + def get_timesteps(self, num_inference_steps, strength, device): + # get the original timestep using init_timestep + init_timestep = min(num_inference_steps * strength, num_inference_steps) + + t_start = int(max(num_inference_steps - init_timestep, 0)) + timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :] + if hasattr(self.scheduler, "set_begin_index"): + self.scheduler.set_begin_index(t_start * self.scheduler.order) + + return timesteps, num_inference_steps - t_start + + # Copied fromCopied from diffusers.pipelines.qwenimage.pipeline_qwenimage.QwenImagePipeline.encode_prompt + def encode_prompt( + self, + prompt: Union[str, List[str]], + 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, + ): + r""" + + Args: + prompt (`str` or `List[str]`, *optional*): + prompt 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, device) + + _, 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 + + def check_inputs( + self, + prompt, + image, + mask_image, + strength, + height, + width, + output_type, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + prompt_embeds_mask=None, + negative_prompt_embeds_mask=None, + callback_on_step_end_tensor_inputs=None, + padding_mask_crop=None, + max_sequence_length=None, + ): + if strength < 0 or strength > 1: + raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}") + + if height % (self.vae_scale_factor * 2) != 0 or width % (self.vae_scale_factor * 2) != 0: + logger.warning( + f"`height` and `width` have to be divisible by {self.vae_scale_factor * 2} but are {height} and {width}. Dimensions will be resized accordingly" + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and prompt_embeds_mask is None: + raise ValueError( + "If `prompt_embeds` are provided, `prompt_embeds_mask` also have to be passed. Make sure to generate `prompt_embeds_mask` from the same text encoder that was used to generate `prompt_embeds`." + ) + if negative_prompt_embeds is not None and negative_prompt_embeds_mask is None: + raise ValueError( + "If `negative_prompt_embeds` are provided, `negative_prompt_embeds_mask` also have to be passed. Make sure to generate `negative_prompt_embeds_mask` from the same text encoder that was used to generate `negative_prompt_embeds`." + ) + if padding_mask_crop is not None: + if not isinstance(image, PIL.Image.Image): + raise ValueError( + f"The image should be a PIL image when inpainting mask crop, but is of type {type(image)}." + ) + if not isinstance(mask_image, PIL.Image.Image): + raise ValueError( + f"The mask image should be a PIL image when inpainting mask crop, but is of type" + f" {type(mask_image)}." + ) + if output_type != "pil": + raise ValueError(f"The output type should be PIL when inpainting mask crop, but is {output_type}.") + + if max_sequence_length is not None and max_sequence_length > 1024: + raise ValueError(f"`max_sequence_length` cannot be greater than 1024 but is {max_sequence_length}") + + @staticmethod + # Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage.QwenImagePipeline._prepare_latent_image_ids + def _prepare_latent_image_ids(batch_size, height, width, device, dtype): + latent_image_ids = torch.zeros(height, width, 3) + latent_image_ids[..., 1] = latent_image_ids[..., 1] + torch.arange(height)[:, None] + latent_image_ids[..., 2] = latent_image_ids[..., 2] + torch.arange(width)[None, :] + + latent_image_id_height, latent_image_id_width, latent_image_id_channels = latent_image_ids.shape + + latent_image_ids = latent_image_ids.reshape( + latent_image_id_height * latent_image_id_width, latent_image_id_channels + ) + + return latent_image_ids.to(device=device, dtype=dtype) + + @staticmethod + # Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage.QwenImagePipeline._pack_latents + def _pack_latents(latents, batch_size, num_channels_latents, height, width): + latents = latents.view(batch_size, num_channels_latents, height // 2, 2, width // 2, 2) + latents = latents.permute(0, 2, 4, 1, 3, 5) + latents = latents.reshape(batch_size, (height // 2) * (width // 2), num_channels_latents * 4) + + return latents + + @staticmethod + # Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage.QwenImagePipeline._unpack_latents + def _unpack_latents(latents, height, width, vae_scale_factor): + batch_size, num_patches, channels = latents.shape + + # 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 = 2 * (int(height) // (vae_scale_factor * 2)) + width = 2 * (int(width) // (vae_scale_factor * 2)) + + latents = latents.view(batch_size, height // 2, width // 2, channels // 4, 2, 2) + latents = latents.permute(0, 3, 1, 4, 2, 5) + + latents = latents.reshape(batch_size, channels // (2 * 2), 1, height, width) + + return latents + + def enable_vae_slicing(self): + r""" + Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to + compute decoding in several steps. This is useful to save some memory and allow larger batch sizes. + """ + self.vae.enable_slicing() + + def disable_vae_slicing(self): + r""" + Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to + computing decoding in one step. + """ + self.vae.disable_slicing() + + def enable_vae_tiling(self): + r""" + Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to + compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow + processing larger images. + """ + self.vae.enable_tiling() + + def disable_vae_tiling(self): + r""" + Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to + computing decoding in one step. + """ + self.vae.disable_tiling() + + def prepare_latents( + self, + image, + timestep, + batch_size, + num_channels_latents, + height, + width, + dtype, + device, + generator, + latents=None, + ): + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + # 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 = 2 * (int(height) // (self.vae_scale_factor * 2)) + width = 2 * (int(width) // (self.vae_scale_factor * 2)) + + shape = (batch_size, 1, num_channels_latents, height, width) + + # If image is [B,C,H,W] -> add T=1. If it's already [B,C,T,H,W], leave it. + if image.dim() == 4: + image = image.unsqueeze(2) + elif image.dim() != 5: + raise ValueError(f"Expected image dims 4 or 5, got {image.dim()}.") + + if latents is not None: + latent_image_ids = self._prepare_latent_image_ids(batch_size, height // 2, width // 2, device, dtype) + return latents.to(device=device, dtype=dtype), latent_image_ids + + image = image.to(device=device, dtype=dtype) + if image.shape[1] != self.latent_channels: + image_latents = self._encode_vae_image(image=image, generator=generator) # [B,z,1,H',W'] + else: + image_latents = image + if batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] == 0: + # expand init_latents for batch_size + additional_image_per_prompt = batch_size // image_latents.shape[0] + image_latents = torch.cat([image_latents] * additional_image_per_prompt, dim=0) + elif batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image` of batch size {image_latents.shape[0]} to {batch_size} text prompts." + ) + else: + image_latents = torch.cat([image_latents], dim=0) + + image_latents = image_latents.transpose(1, 2) # [B,1,z,H',W'] + + if latents is None: + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + latents = self.scheduler.scale_noise(image_latents, timestep, noise) + else: + noise = latents.to(device) + latents = noise + + noise = self._pack_latents(noise, batch_size, num_channels_latents, height, width) + image_latents = self._pack_latents(image_latents, batch_size, num_channels_latents, height, width) + latents = self._pack_latents(latents, batch_size, num_channels_latents, height, width) + + latent_image_ids = self._prepare_latent_image_ids(batch_size, height // 2, width // 2, device, dtype) + + return latents, noise, image_latents, latent_image_ids + + def prepare_mask_latents( + self, + mask, + masked_image, + batch_size, + num_channels_latents, + num_images_per_prompt, + height, + width, + dtype, + device, + generator, + ): + # 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 = 2 * (int(height) // (self.vae_scale_factor * 2)) + width = 2 * (int(width) // (self.vae_scale_factor * 2)) + # resize the mask to latents shape as we concatenate the mask to the latents + # we do that before converting to dtype to avoid breaking in case we're using cpu_offload + # and half precision + mask = torch.nn.functional.interpolate(mask, size=(height, width)) + mask = mask.to(device=device, dtype=dtype) + + batch_size = batch_size * num_images_per_prompt + + if masked_image.dim() == 4: + masked_image = masked_image.unsqueeze(2) + elif masked_image.dim() != 5: + raise ValueError(f"Expected image dims 4 or 5, got {masked_image.dim()}.") + + masked_image = masked_image.to(device=device, dtype=dtype) + + if masked_image.shape[1] == self.latent_channels: + masked_image_latents = masked_image + else: + masked_image_latents = self._encode_vae_image(image=masked_image, generator=generator) + + # duplicate mask and masked_image_latents for each generation per prompt, using mps friendly method + if mask.shape[0] < batch_size: + if not batch_size % mask.shape[0] == 0: + raise ValueError( + "The passed mask and the required batch size don't match. Masks are supposed to be duplicated to" + f" a total batch size of {batch_size}, but {mask.shape[0]} masks were passed. Make sure the number" + " of masks that you pass is divisible by the total requested batch size." + ) + mask = mask.repeat(batch_size // mask.shape[0], 1, 1, 1) + if masked_image_latents.shape[0] < batch_size: + if not batch_size % masked_image_latents.shape[0] == 0: + raise ValueError( + "The passed images and the required batch size don't match. Images are supposed to be duplicated" + f" to a total batch size of {batch_size}, but {masked_image_latents.shape[0]} images were passed." + " Make sure the number of images that you pass is divisible by the total requested batch size." + ) + masked_image_latents = masked_image_latents.repeat(batch_size // masked_image_latents.shape[0], 1, 1, 1, 1) + + # aligning device to prevent device errors when concating it with the latent model input + masked_image_latents = masked_image_latents.to(device=device, dtype=dtype) + + masked_image_latents = self._pack_latents( + masked_image_latents, + batch_size, + num_channels_latents, + height, + width, + ) + mask = self._pack_latents( + mask.repeat(1, num_channels_latents, 1, 1), + batch_size, + num_channels_latents, + height, + width, + ) + + return mask, masked_image_latents + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def attention_kwargs(self): + return self._attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def current_timestep(self): + return self._current_timestep + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: Union[str, List[str]] = None, + negative_prompt: Union[str, List[str]] = None, + true_cfg_scale: float = 4.0, + image: PipelineImageInput = None, + mask_image: PipelineImageInput = None, + masked_image_latents: PipelineImageInput = None, + height: Optional[int] = None, + width: Optional[int] = None, + padding_mask_crop: Optional[int] = None, + strength: float = 0.6, + 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, + ): + 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`). + image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.Tensor]`, `List[PIL.Image.Image]`, or `List[np.ndarray]`): + `Image`, numpy array or tensor representing an image batch to be used as the starting point. For both + numpy array and pytorch tensor, the expected value range is between `[0, 1]` If it's a tensor or a list + or tensors, the expected shape should be `(B, C, H, W)` or `(C, H, W)`. If it is a numpy array or a + list of arrays, the expected shape should be `(B, H, W, C)` or `(H, W, C)` It can also accept image + latents as `image`, but if passing latents directly it is not encoded again. + true_cfg_scale (`float`, *optional*, defaults to 1.0): + When > 1.0 and a provided `negative_prompt`, enables true classifier-free guidance. + mask_image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.Tensor]`, `List[PIL.Image.Image]`, or `List[np.ndarray]`): + `Image`, numpy array or tensor representing an image batch to mask `image`. White pixels in the mask + are repainted while black pixels are preserved. If `mask_image` is a PIL image, it is converted to a + single channel (luminance) before use. If it's a numpy array or pytorch tensor, it should contain one + color channel (L) instead of 3, so the expected shape for pytorch tensor would be `(B, 1, H, W)`, `(B, + H, W)`, `(1, H, W)`, `(H, W)`. And for numpy array would be for `(B, H, W, 1)`, `(B, H, W)`, `(H, W, + 1)`, or `(H, W)`. + mask_image_latent (`torch.Tensor`, `List[torch.Tensor]`): + `Tensor` representing an image batch to mask `image` generated by VAE. If not provided, the mask + latents tensor will ge generated by `mask_image`. + 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. + padding_mask_crop (`int`, *optional*, defaults to `None`): + The size of margin in the crop to be applied to the image and masking. If `None`, no crop is applied to + image and mask_image. If `padding_mask_crop` is not `None`, it will first find a rectangular region + with the same aspect ration of the image and contains all masked area, and then expand that area based + on `padding_mask_crop`. The image and mask_image will then be cropped based on the expanded area before + resizing to the original image size for inpainting. This is useful when the masked area is small while + the image is large and contain information irrelevant for inpainting, such as background. + strength (`float`, *optional*, defaults to 1.0): + Indicates extent to transform the reference `image`. Must be between 0 and 1. `image` is used as a + starting point and more noise is added the higher the `strength`. The number of denoising steps depends + on the amount of noise initially added. When `strength` is 1, added noise is maximum and the denoising + process runs for the full number of iterations specified in `num_inference_steps`. A value of 1 + essentially ignores `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. + 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. + 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. + """ + + height = height or self.default_sample_size * self.vae_scale_factor + width = width or self.default_sample_size * self.vae_scale_factor + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + image, + mask_image, + strength, + height, + width, + output_type=output_type, + 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, + padding_mask_crop=padding_mask_crop, + max_sequence_length=max_sequence_length, + ) + + self._guidance_scale = guidance_scale + self._attention_kwargs = attention_kwargs + self._current_timestep = None + self._interrupt = False + + # 2. Preprocess image + if padding_mask_crop is not None: + crops_coords = self.mask_processor.get_crop_region(mask_image, width, height, pad=padding_mask_crop) + resize_mode = "fill" + else: + crops_coords = None + resize_mode = "default" + + original_image = image + init_image = self.image_processor.preprocess( + image, height=height, width=width, crops_coords=crops_coords, resize_mode=resize_mode + ) + init_image = init_image.to(dtype=torch.float32) + + # 3. 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._execution_device + + 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( + 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, + ) + if do_true_cfg: + negative_prompt_embeds, negative_prompt_embeds_mask = self.encode_prompt( + 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, + ) + + # 4. Prepare timesteps + sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps) if sigmas is None else sigmas + image_seq_len = (int(height) // self.vae_scale_factor // 2) * (int(width) // self.vae_scale_factor // 2) + 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, + ) + timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength, device) + + if num_inference_steps < 1: + raise ValueError( + f"After adjusting the num_inference_steps by strength parameter: {strength}, the number of pipeline" + f"steps is {num_inference_steps} which is < 1 and not appropriate for this pipeline." + ) + latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt) + + # 5. Prepare latent variables + num_channels_latents = self.transformer.config.in_channels // 4 + + latents, noise, image_latents, latent_image_ids = self.prepare_latents( + init_image, + latent_timestep, + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + mask_condition = self.mask_processor.preprocess( + mask_image, height=height, width=width, resize_mode=resize_mode, crops_coords=crops_coords + ) + + if masked_image_latents is None: + masked_image = init_image * (mask_condition < 0.5) + else: + masked_image = masked_image_latents + + mask, masked_image_latents = self.prepare_mask_latents( + mask_condition, + masked_image, + batch_size, + num_channels_latents, + num_images_per_prompt, + height, + width, + prompt_embeds.dtype, + device, + generator, + ) + + img_shapes = [(1, height // self.vae_scale_factor // 2, width // self.vae_scale_factor // 2)] * batch_size + + 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 + 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 + # 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=latents, + 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] + + if do_true_cfg: + with self.transformer.cache_context("uncond"): + neg_noise_pred = self.transformer( + hidden_states=latents, + 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] + 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] + + # for 64 channel transformer only. + init_latents_proper = image_latents + init_mask = mask + + if i < len(timesteps) - 1: + noise_timestep = timesteps[i + 1] + init_latents_proper = self.scheduler.scale_noise( + init_latents_proper, torch.tensor([noise_timestep]), noise + ) + + latents = (1 - init_mask) * init_latents_proper + init_mask * latents + + 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) + + if padding_mask_crop is not None: + image = [ + self.image_processor.apply_overlay(mask_image, original_image, i, crops_coords) for i in image + ] + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return QwenImagePipelineOutput(images=image) diff --git a/src/diffusers/utils/dummy_torch_and_transformers_objects.py b/src/diffusers/utils/dummy_torch_and_transformers_objects.py index 293086631f..e02457bf8d 100644 --- a/src/diffusers/utils/dummy_torch_and_transformers_objects.py +++ b/src/diffusers/utils/dummy_torch_and_transformers_objects.py @@ -1742,6 +1742,36 @@ class PixArtSigmaPipeline(metaclass=DummyObject): requires_backends(cls, ["torch", "transformers"]) +class QwenImageImg2ImgPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class QwenImageInpaintPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + class QwenImagePipeline(metaclass=DummyObject): _backends = ["torch", "transformers"] diff --git a/tests/pipelines/qwenimage/test_qwenimage_img2img.py b/tests/pipelines/qwenimage/test_qwenimage_img2img.py new file mode 100644 index 0000000000..9f21257299 --- /dev/null +++ b/tests/pipelines/qwenimage/test_qwenimage_img2img.py @@ -0,0 +1,218 @@ +import random +import unittest + +import numpy as np +import torch +from transformers import Qwen2_5_VLConfig, Qwen2_5_VLForConditionalGeneration, Qwen2Tokenizer + +from diffusers import ( + AutoencoderKLQwenImage, + FlowMatchEulerDiscreteScheduler, + QwenImageImg2ImgPipeline, + QwenImageTransformer2DModel, +) +from diffusers.utils.testing_utils import ( + enable_full_determinism, + floats_tensor, + torch_device, +) + +from ..test_pipelines_common import PipelineTesterMixin, to_np + + +enable_full_determinism() + + +class QwenImageImg2ImgPipelineFastTests(unittest.TestCase, PipelineTesterMixin): + pipeline_class = QwenImageImg2ImgPipeline + params = frozenset(["prompt", "image", "height", "width", "guidance_scale", "true_cfg_scale", "strength"]) + batch_params = frozenset(["prompt", "image"]) + image_params = frozenset(["image"]) + image_latents_params = frozenset(["latents"]) + required_optional_params = frozenset( + [ + "num_inference_steps", + "generator", + "latents", + "return_dict", + "callback_on_step_end", + "callback_on_step_end_tensor_inputs", + ] + ) + supports_dduf = False + test_xformers_attention = False + test_attention_slicing = True + test_layerwise_casting = True + test_group_offloading = True + + def get_dummy_components(self): + torch.manual_seed(0) + transformer = QwenImageTransformer2DModel( + patch_size=2, + in_channels=16, + out_channels=4, + num_layers=2, + attention_head_dim=16, + num_attention_heads=3, + joint_attention_dim=16, + guidance_embeds=False, + axes_dims_rope=(8, 4, 4), + ) + + torch.manual_seed(0) + z_dim = 4 + vae = AutoencoderKLQwenImage( + base_dim=z_dim * 6, + z_dim=z_dim, + dim_mult=[1, 2, 4], + num_res_blocks=1, + temperal_downsample=[False, True], + latents_mean=[0.0] * 4, + latents_std=[1.0] * 4, + ) + + torch.manual_seed(0) + scheduler = FlowMatchEulerDiscreteScheduler() + + torch.manual_seed(0) + config = Qwen2_5_VLConfig( + text_config={ + "hidden_size": 16, + "intermediate_size": 16, + "num_hidden_layers": 2, + "num_attention_heads": 2, + "num_key_value_heads": 2, + "rope_scaling": { + "mrope_section": [1, 1, 2], + "rope_type": "default", + "type": "default", + }, + "rope_theta": 1000000.0, + }, + vision_config={ + "depth": 2, + "hidden_size": 16, + "intermediate_size": 16, + "num_heads": 2, + "out_hidden_size": 16, + }, + hidden_size=16, + vocab_size=152064, + vision_end_token_id=151653, + vision_start_token_id=151652, + vision_token_id=151654, + ) + text_encoder = Qwen2_5_VLForConditionalGeneration(config) + tokenizer = Qwen2Tokenizer.from_pretrained("hf-internal-testing/tiny-random-Qwen2VLForConditionalGeneration") + + return { + "transformer": transformer, + "vae": vae, + "scheduler": scheduler, + "text_encoder": text_encoder, + "tokenizer": tokenizer, + } + + def get_dummy_inputs(self, device, seed=0): + image = floats_tensor((1, 3, 32, 32), rng=random.Random(seed)).to(device) + if str(device).startswith("mps"): + generator = torch.manual_seed(seed) + else: + generator = torch.Generator(device="cpu").manual_seed(seed) + + inputs = { + "image": image, + "prompt": "dance monkey", + "negative_prompt": "bad quality", + "generator": generator, + "num_inference_steps": 2, + "guidance_scale": 3.0, + "true_cfg_scale": 1.0, + "height": 32, + "width": 32, + "max_sequence_length": 16, + "output_type": "pt", + } + + return inputs + + def test_inference(self): + device = "cpu" + + components = self.get_dummy_components() + pipe = self.pipeline_class(**components) + pipe.to(device) + pipe.set_progress_bar_config(disable=None) + + inputs = self.get_dummy_inputs(device) + image = pipe(**inputs).images + generated_image = image[0] + self.assertEqual(generated_image.shape, (3, 32, 32)) + + def test_inference_batch_single_identical(self): + self._test_inference_batch_single_identical(batch_size=3, expected_max_diff=1e-1) + + def test_attention_slicing_forward_pass( + self, test_max_difference=True, test_mean_pixel_difference=True, expected_max_diff=1e-3 + ): + if not self.test_attention_slicing: + return + + components = self.get_dummy_components() + pipe = self.pipeline_class(**components) + for component in pipe.components.values(): + if hasattr(component, "set_default_attn_processor"): + component.set_default_attn_processor() + pipe.to(torch_device) + pipe.set_progress_bar_config(disable=None) + + generator_device = "cpu" + inputs = self.get_dummy_inputs(generator_device) + output_without_slicing = pipe(**inputs).images[0] + + pipe.enable_attention_slicing(slice_size=1) + inputs = self.get_dummy_inputs(generator_device) + output_with_slicing1 = pipe(**inputs).images[0] + + pipe.enable_attention_slicing(slice_size=2) + inputs = self.get_dummy_inputs(generator_device) + output_with_slicing2 = pipe(**inputs).images[0] + + if test_max_difference: + max_diff1 = np.abs(to_np(output_with_slicing1) - to_np(output_without_slicing)).max() + max_diff2 = np.abs(to_np(output_with_slicing2) - to_np(output_without_slicing)).max() + self.assertLess( + max(max_diff1, max_diff2), + expected_max_diff, + "Attention slicing should not affect the inference results", + ) + + def test_vae_tiling(self, expected_diff_max: float = 0.2): + generator_device = "cpu" + components = self.get_dummy_components() + + pipe = self.pipeline_class(**components) + pipe.to("cpu") + pipe.set_progress_bar_config(disable=None) + + # Without tiling + inputs = self.get_dummy_inputs(generator_device) + inputs["height"] = inputs["width"] = 128 + output_without_tiling = pipe(**inputs)[0] + + # With tiling + pipe.vae.enable_tiling( + tile_sample_min_height=96, + tile_sample_min_width=96, + tile_sample_stride_height=64, + tile_sample_stride_width=64, + ) + inputs = self.get_dummy_inputs(generator_device) + inputs["height"] = inputs["width"] = 128 + output_with_tiling = pipe(**inputs)[0] + + self.assertLess( + (to_np(output_without_tiling) - to_np(output_with_tiling)).max(), + expected_diff_max, + "VAE tiling should not affect the inference results", + ) diff --git a/tests/pipelines/qwenimage/test_qwenimage_inpaint.py b/tests/pipelines/qwenimage/test_qwenimage_inpaint.py new file mode 100644 index 0000000000..1a40630a2d --- /dev/null +++ b/tests/pipelines/qwenimage/test_qwenimage_inpaint.py @@ -0,0 +1,233 @@ +# Copyright 2025 The HuggingFace Team. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import random +import unittest + +import numpy as np +import torch +from transformers import Qwen2_5_VLConfig, Qwen2_5_VLForConditionalGeneration, Qwen2Tokenizer + +from diffusers import ( + AutoencoderKLQwenImage, + FlowMatchEulerDiscreteScheduler, + QwenImageInpaintPipeline, + QwenImageTransformer2DModel, +) +from diffusers.utils.testing_utils import enable_full_determinism, floats_tensor, torch_device + +from ..pipeline_params import TEXT_TO_IMAGE_BATCH_PARAMS, TEXT_TO_IMAGE_IMAGE_PARAMS, TEXT_TO_IMAGE_PARAMS +from ..test_pipelines_common import PipelineTesterMixin, to_np + + +enable_full_determinism() + + +class QwenImageInpaintPipelineFastTests(PipelineTesterMixin, unittest.TestCase): + pipeline_class = QwenImageInpaintPipeline + params = TEXT_TO_IMAGE_PARAMS - {"cross_attention_kwargs"} + batch_params = TEXT_TO_IMAGE_BATCH_PARAMS + image_params = TEXT_TO_IMAGE_IMAGE_PARAMS + image_latents_params = TEXT_TO_IMAGE_IMAGE_PARAMS + required_optional_params = frozenset( + [ + "num_inference_steps", + "generator", + "latents", + "return_dict", + "callback_on_step_end", + "callback_on_step_end_tensor_inputs", + ] + ) + supports_dduf = False + test_xformers_attention = False + test_layerwise_casting = True + test_group_offloading = True + + def get_dummy_components(self): + torch.manual_seed(0) + transformer = QwenImageTransformer2DModel( + patch_size=2, + in_channels=16, + out_channels=4, + num_layers=2, + attention_head_dim=16, + num_attention_heads=3, + joint_attention_dim=16, + guidance_embeds=False, + axes_dims_rope=(8, 4, 4), + ) + + torch.manual_seed(0) + z_dim = 4 + vae = AutoencoderKLQwenImage( + base_dim=z_dim * 6, + z_dim=z_dim, + dim_mult=[1, 2, 4], + num_res_blocks=1, + temperal_downsample=[False, True], + # fmt: off + latents_mean=[0.0] * 4, + latents_std=[1.0] * 4, + # fmt: on + ) + + torch.manual_seed(0) + scheduler = FlowMatchEulerDiscreteScheduler() + + torch.manual_seed(0) + config = Qwen2_5_VLConfig( + text_config={ + "hidden_size": 16, + "intermediate_size": 16, + "num_hidden_layers": 2, + "num_attention_heads": 2, + "num_key_value_heads": 2, + "rope_scaling": { + "mrope_section": [1, 1, 2], + "rope_type": "default", + "type": "default", + }, + "rope_theta": 1000000.0, + }, + vision_config={ + "depth": 2, + "hidden_size": 16, + "intermediate_size": 16, + "num_heads": 2, + "out_hidden_size": 16, + }, + hidden_size=16, + vocab_size=152064, + vision_end_token_id=151653, + vision_start_token_id=151652, + vision_token_id=151654, + ) + text_encoder = Qwen2_5_VLForConditionalGeneration(config) + tokenizer = Qwen2Tokenizer.from_pretrained("hf-internal-testing/tiny-random-Qwen2VLForConditionalGeneration") + + components = { + "transformer": transformer, + "vae": vae, + "scheduler": scheduler, + "text_encoder": text_encoder, + "tokenizer": tokenizer, + } + return components + + def get_dummy_inputs(self, device, seed=0): + image = floats_tensor((1, 3, 32, 32), rng=random.Random(seed)).to(device) + mask_image = torch.ones((1, 1, 32, 32)).to(device) + if str(device).startswith("mps"): + generator = torch.manual_seed(seed) + else: + generator = torch.Generator(device=device).manual_seed(seed) + + inputs = { + "prompt": "dance monkey", + "negative_prompt": "bad quality", + "image": image, + "mask_image": mask_image, + "generator": generator, + "num_inference_steps": 2, + "guidance_scale": 3.0, + "true_cfg_scale": 1.0, + "height": 32, + "width": 32, + "max_sequence_length": 16, + "output_type": "pt", + } + + return inputs + + def test_inference(self): + device = "cpu" + + components = self.get_dummy_components() + pipe = self.pipeline_class(**components) + pipe.to(device) + pipe.set_progress_bar_config(disable=None) + + inputs = self.get_dummy_inputs(device) + image = pipe(**inputs).images + generated_image = image[0] + self.assertEqual(generated_image.shape, (3, 32, 32)) + + def test_inference_batch_single_identical(self): + self._test_inference_batch_single_identical(batch_size=3, expected_max_diff=1e-1) + + def test_attention_slicing_forward_pass( + self, test_max_difference=True, test_mean_pixel_difference=True, expected_max_diff=1e-3 + ): + if not self.test_attention_slicing: + return + + components = self.get_dummy_components() + pipe = self.pipeline_class(**components) + for component in pipe.components.values(): + if hasattr(component, "set_default_attn_processor"): + component.set_default_attn_processor() + pipe.to(torch_device) + pipe.set_progress_bar_config(disable=None) + + generator_device = "cpu" + inputs = self.get_dummy_inputs(generator_device) + output_without_slicing = pipe(**inputs)[0] + + pipe.enable_attention_slicing(slice_size=1) + inputs = self.get_dummy_inputs(generator_device) + output_with_slicing1 = pipe(**inputs)[0] + + pipe.enable_attention_slicing(slice_size=2) + inputs = self.get_dummy_inputs(generator_device) + output_with_slicing2 = pipe(**inputs)[0] + + if test_max_difference: + max_diff1 = np.abs(to_np(output_with_slicing1) - to_np(output_without_slicing)).max() + max_diff2 = np.abs(to_np(output_with_slicing2) - to_np(output_without_slicing)).max() + self.assertLess( + max(max_diff1, max_diff2), + expected_max_diff, + "Attention slicing should not affect the inference results", + ) + + def test_vae_tiling(self, expected_diff_max: float = 0.2): + generator_device = "cpu" + components = self.get_dummy_components() + + pipe = self.pipeline_class(**components) + pipe.to("cpu") + pipe.set_progress_bar_config(disable=None) + + # Without tiling + inputs = self.get_dummy_inputs(generator_device) + inputs["height"] = inputs["width"] = 128 + output_without_tiling = pipe(**inputs)[0] + + # With tiling + pipe.vae.enable_tiling( + tile_sample_min_height=96, + tile_sample_min_width=96, + tile_sample_stride_height=64, + tile_sample_stride_width=64, + ) + inputs = self.get_dummy_inputs(generator_device) + inputs["height"] = inputs["width"] = 128 + output_with_tiling = pipe(**inputs)[0] + + self.assertLess( + (to_np(output_without_tiling) - to_np(output_with_tiling)).max(), + expected_diff_max, + "VAE tiling should not affect the inference results", + ) From baa9b582f348e52aa2fc245e366611f454e1082b Mon Sep 17 00:00:00 2001 From: Sayak Paul Date: Wed, 13 Aug 2025 10:33:20 +0530 Subject: [PATCH 2/2] [core] parallel loading of shards (#12028) * checking. * checking * checking * up * up * up * Apply suggestions from code review Co-authored-by: Dhruv Nair * up * up * fix * review feedback. --------- Co-authored-by: Dhruv Nair --- src/diffusers/loaders/single_file_model.py | 2 +- src/diffusers/loaders/single_file_utils.py | 2 +- src/diffusers/loaders/transformer_flux.py | 3 +- src/diffusers/loaders/transformer_sd3.py | 3 +- src/diffusers/loaders/unet.py | 3 +- src/diffusers/models/model_loading_utils.py | 158 ++++++++++++++++++++ src/diffusers/models/modeling_utils.py | 105 ++++++------- src/diffusers/utils/__init__.py | 2 + src/diffusers/utils/constants.py | 2 + tests/models/test_modeling_common.py | 35 +++++ 10 files changed, 250 insertions(+), 65 deletions(-) diff --git a/src/diffusers/loaders/single_file_model.py b/src/diffusers/loaders/single_file_model.py index ecccf3c113..16bd044107 100644 --- a/src/diffusers/loaders/single_file_model.py +++ b/src/diffusers/loaders/single_file_model.py @@ -62,7 +62,7 @@ logger = logging.get_logger(__name__) if is_accelerate_available(): from accelerate import dispatch_model, init_empty_weights - from ..models.modeling_utils import load_model_dict_into_meta + from ..models.model_loading_utils import load_model_dict_into_meta if is_torch_version(">=", "1.9.0") and is_accelerate_available(): _LOW_CPU_MEM_USAGE_DEFAULT = True diff --git a/src/diffusers/loaders/single_file_utils.py b/src/diffusers/loaders/single_file_utils.py index 723f0c136f..ef6c41e3ce 100644 --- a/src/diffusers/loaders/single_file_utils.py +++ b/src/diffusers/loaders/single_file_utils.py @@ -55,7 +55,7 @@ if is_transformers_available(): if is_accelerate_available(): from accelerate import init_empty_weights - from ..models.modeling_utils import load_model_dict_into_meta + from ..models.model_loading_utils import load_model_dict_into_meta logger = logging.get_logger(__name__) # pylint: disable=invalid-name diff --git a/src/diffusers/loaders/transformer_flux.py b/src/diffusers/loaders/transformer_flux.py index ced81960fa..ef7b921b7d 100644 --- a/src/diffusers/loaders/transformer_flux.py +++ b/src/diffusers/loaders/transformer_flux.py @@ -17,7 +17,8 @@ from ..models.embeddings import ( ImageProjection, MultiIPAdapterImageProjection, ) -from ..models.modeling_utils import _LOW_CPU_MEM_USAGE_DEFAULT, load_model_dict_into_meta +from ..models.model_loading_utils import load_model_dict_into_meta +from ..models.modeling_utils import _LOW_CPU_MEM_USAGE_DEFAULT from ..utils import is_accelerate_available, is_torch_version, logging from ..utils.torch_utils import empty_device_cache diff --git a/src/diffusers/loaders/transformer_sd3.py b/src/diffusers/loaders/transformer_sd3.py index 1bc3a9c7a8..e3728082ef 100644 --- a/src/diffusers/loaders/transformer_sd3.py +++ b/src/diffusers/loaders/transformer_sd3.py @@ -16,7 +16,8 @@ from typing import Dict from ..models.attention_processor import SD3IPAdapterJointAttnProcessor2_0 from ..models.embeddings import IPAdapterTimeImageProjection -from ..models.modeling_utils import _LOW_CPU_MEM_USAGE_DEFAULT, load_model_dict_into_meta +from ..models.model_loading_utils import load_model_dict_into_meta +from ..models.modeling_utils import _LOW_CPU_MEM_USAGE_DEFAULT from ..utils import is_accelerate_available, is_torch_version, logging from ..utils.torch_utils import empty_device_cache diff --git a/src/diffusers/loaders/unet.py b/src/diffusers/loaders/unet.py index 1d698e5a8b..c5e56af156 100644 --- a/src/diffusers/loaders/unet.py +++ b/src/diffusers/loaders/unet.py @@ -30,7 +30,8 @@ from ..models.embeddings import ( IPAdapterPlusImageProjection, MultiIPAdapterImageProjection, ) -from ..models.modeling_utils import _LOW_CPU_MEM_USAGE_DEFAULT, load_model_dict_into_meta, load_state_dict +from ..models.model_loading_utils import load_model_dict_into_meta +from ..models.modeling_utils import _LOW_CPU_MEM_USAGE_DEFAULT, load_state_dict from ..utils import ( USE_PEFT_BACKEND, _get_model_file, diff --git a/src/diffusers/models/model_loading_utils.py b/src/diffusers/models/model_loading_utils.py index 4e2d24b750..1fcaedcb87 100644 --- a/src/diffusers/models/model_loading_utils.py +++ b/src/diffusers/models/model_loading_utils.py @@ -14,12 +14,14 @@ # See the License for the specific language governing permissions and # limitations under the License. +import functools import importlib import inspect import math import os from array import array from collections import OrderedDict, defaultdict +from concurrent.futures import ThreadPoolExecutor, as_completed from pathlib import Path from typing import Dict, List, Optional, Union from zipfile import is_zipfile @@ -31,6 +33,7 @@ from huggingface_hub.utils import EntryNotFoundError from ..quantizers import DiffusersQuantizer from ..utils import ( + DEFAULT_HF_PARALLEL_LOADING_WORKERS, GGUF_FILE_EXTENSION, SAFE_WEIGHTS_INDEX_NAME, SAFETENSORS_FILE_EXTENSION, @@ -310,6 +313,161 @@ def load_model_dict_into_meta( return offload_index, state_dict_index +def check_support_param_buffer_assignment(model_to_load, state_dict, start_prefix=""): + """ + Checks if `model_to_load` supports param buffer assignment (such as when loading in empty weights) by first + checking if the model explicitly disables it, then by ensuring that the state dict keys are a subset of the model's + parameters. + + """ + if model_to_load.device.type == "meta": + return False + + if len([key for key in state_dict if key.startswith(start_prefix)]) == 0: + return False + + # Some models explicitly do not support param buffer assignment + if not getattr(model_to_load, "_supports_param_buffer_assignment", True): + logger.debug( + f"{model_to_load.__class__.__name__} does not support param buffer assignment, loading will be slower" + ) + return False + + # If the model does, the incoming `state_dict` and the `model_to_load` must be the same dtype + first_key = next(iter(model_to_load.state_dict().keys())) + if start_prefix + first_key in state_dict: + return state_dict[start_prefix + first_key].dtype == model_to_load.state_dict()[first_key].dtype + + return False + + +def _load_shard_file( + shard_file, + model, + model_state_dict, + device_map=None, + dtype=None, + hf_quantizer=None, + keep_in_fp32_modules=None, + dduf_entries=None, + loaded_keys=None, + unexpected_keys=None, + offload_index=None, + offload_folder=None, + state_dict_index=None, + state_dict_folder=None, + ignore_mismatched_sizes=False, + low_cpu_mem_usage=False, +): + state_dict = load_state_dict(shard_file, dduf_entries=dduf_entries) + mismatched_keys = _find_mismatched_keys( + state_dict, + model_state_dict, + loaded_keys, + ignore_mismatched_sizes, + ) + error_msgs = [] + if low_cpu_mem_usage: + offload_index, state_dict_index = load_model_dict_into_meta( + model, + state_dict, + device_map=device_map, + dtype=dtype, + hf_quantizer=hf_quantizer, + keep_in_fp32_modules=keep_in_fp32_modules, + unexpected_keys=unexpected_keys, + offload_folder=offload_folder, + offload_index=offload_index, + state_dict_index=state_dict_index, + state_dict_folder=state_dict_folder, + ) + else: + assign_to_params_buffers = check_support_param_buffer_assignment(model, state_dict) + + error_msgs += _load_state_dict_into_model(model, state_dict, assign_to_params_buffers) + return offload_index, state_dict_index, mismatched_keys, error_msgs + + +def _load_shard_files_with_threadpool( + shard_files, + model, + model_state_dict, + device_map=None, + dtype=None, + hf_quantizer=None, + keep_in_fp32_modules=None, + dduf_entries=None, + loaded_keys=None, + unexpected_keys=None, + offload_index=None, + offload_folder=None, + state_dict_index=None, + state_dict_folder=None, + ignore_mismatched_sizes=False, + low_cpu_mem_usage=False, +): + # Do not spawn anymore workers than you need + num_workers = min(len(shard_files), DEFAULT_HF_PARALLEL_LOADING_WORKERS) + + logger.info(f"Loading model weights in parallel with {num_workers} workers...") + + error_msgs = [] + mismatched_keys = [] + + load_one = functools.partial( + _load_shard_file, + model=model, + model_state_dict=model_state_dict, + device_map=device_map, + dtype=dtype, + hf_quantizer=hf_quantizer, + keep_in_fp32_modules=keep_in_fp32_modules, + dduf_entries=dduf_entries, + loaded_keys=loaded_keys, + unexpected_keys=unexpected_keys, + offload_index=offload_index, + offload_folder=offload_folder, + state_dict_index=state_dict_index, + state_dict_folder=state_dict_folder, + ignore_mismatched_sizes=ignore_mismatched_sizes, + low_cpu_mem_usage=low_cpu_mem_usage, + ) + + with ThreadPoolExecutor(max_workers=num_workers) as executor: + with logging.tqdm(total=len(shard_files), desc="Loading checkpoint shards") as pbar: + futures = [executor.submit(load_one, shard_file) for shard_file in shard_files] + for future in as_completed(futures): + result = future.result() + offload_index, state_dict_index, _mismatched_keys, _error_msgs = result + error_msgs += _error_msgs + mismatched_keys += _mismatched_keys + pbar.update(1) + + return offload_index, state_dict_index, mismatched_keys, error_msgs + + +def _find_mismatched_keys( + state_dict, + model_state_dict, + loaded_keys, + ignore_mismatched_sizes, +): + mismatched_keys = [] + if ignore_mismatched_sizes: + for checkpoint_key in loaded_keys: + model_key = checkpoint_key + # If the checkpoint is sharded, we may not have the key here. + if checkpoint_key not in state_dict: + continue + + if model_key in model_state_dict and state_dict[checkpoint_key].shape != model_state_dict[model_key].shape: + mismatched_keys.append( + (checkpoint_key, state_dict[checkpoint_key].shape, model_state_dict[model_key].shape) + ) + del state_dict[checkpoint_key] + return mismatched_keys + + def _load_state_dict_into_model( model_to_load, state_dict: OrderedDict, assign_to_params_buffers: bool = False ) -> List[str]: diff --git a/src/diffusers/models/modeling_utils.py b/src/diffusers/models/modeling_utils.py index 815f12a707..8ab3014262 100644 --- a/src/diffusers/models/modeling_utils.py +++ b/src/diffusers/models/modeling_utils.py @@ -15,6 +15,7 @@ # limitations under the License. import copy +import functools import inspect import itertools import json @@ -41,7 +42,9 @@ from ..quantizers import DiffusersAutoQuantizer, DiffusersQuantizer from ..quantizers.quantization_config import QuantizationMethod from ..utils import ( CONFIG_NAME, + ENV_VARS_TRUE_VALUES, FLAX_WEIGHTS_NAME, + HF_PARALLEL_LOADING_FLAG, SAFE_WEIGHTS_INDEX_NAME, SAFETENSORS_WEIGHTS_NAME, WEIGHTS_INDEX_NAME, @@ -69,9 +72,8 @@ from .model_loading_utils import ( _expand_device_map, _fetch_index_file, _fetch_index_file_legacy, - _find_mismatched_keys, - _load_state_dict_into_model, - load_model_dict_into_meta, + _load_shard_file, + _load_shard_files_with_threadpool, load_state_dict, ) @@ -208,34 +210,6 @@ def get_parameter_dtype(parameter: torch.nn.Module) -> torch.dtype: return last_tuple[1].dtype -def check_support_param_buffer_assignment(model_to_load, state_dict, start_prefix=""): - """ - Checks if `model_to_load` supports param buffer assignment (such as when loading in empty weights) by first - checking if the model explicitly disables it, then by ensuring that the state dict keys are a subset of the model's - parameters. - - """ - if model_to_load.device.type == "meta": - return False - - if len([key for key in state_dict if key.startswith(start_prefix)]) == 0: - return False - - # Some models explicitly do not support param buffer assignment - if not getattr(model_to_load, "_supports_param_buffer_assignment", True): - logger.debug( - f"{model_to_load.__class__.__name__} does not support param buffer assignment, loading will be slower" - ) - return False - - # If the model does, the incoming `state_dict` and the `model_to_load` must be the same dtype - first_key = next(iter(model_to_load.state_dict().keys())) - if start_prefix + first_key in state_dict: - return state_dict[start_prefix + first_key].dtype == model_to_load.state_dict()[first_key].dtype - - return False - - @contextmanager def no_init_weights(): """ @@ -988,6 +962,10 @@ class ModelMixin(torch.nn.Module, PushToHubMixin): dduf_entries: Optional[Dict[str, DDUFEntry]] = kwargs.pop("dduf_entries", None) disable_mmap = kwargs.pop("disable_mmap", False) + is_parallel_loading_enabled = os.environ.get(HF_PARALLEL_LOADING_FLAG, "").upper() in ENV_VARS_TRUE_VALUES + if is_parallel_loading_enabled and not low_cpu_mem_usage: + raise NotImplementedError("Parallel loading is not supported when not using `low_cpu_mem_usage`.") + if torch_dtype is not None and not isinstance(torch_dtype, torch.dtype): torch_dtype = torch.float32 logger.warning( @@ -1323,6 +1301,7 @@ class ModelMixin(torch.nn.Module, PushToHubMixin): hf_quantizer=hf_quantizer, keep_in_fp32_modules=keep_in_fp32_modules, dduf_entries=dduf_entries, + is_parallel_loading_enabled=is_parallel_loading_enabled, ) loading_info = { "missing_keys": missing_keys, @@ -1518,6 +1497,7 @@ class ModelMixin(torch.nn.Module, PushToHubMixin): offload_state_dict: Optional[bool] = None, offload_folder: Optional[Union[str, os.PathLike]] = None, dduf_entries: Optional[Dict[str, DDUFEntry]] = None, + is_parallel_loading_enabled: Optional[bool] = False, ): model_state_dict = model.state_dict() expected_keys = list(model_state_dict.keys()) @@ -1531,6 +1511,9 @@ class ModelMixin(torch.nn.Module, PushToHubMixin): for pat in cls._keys_to_ignore_on_load_unexpected: unexpected_keys = [k for k in unexpected_keys if re.search(pat, k) is None] + mismatched_keys = [] + error_msgs = [] + # Deal with offload if device_map is not None and "disk" in device_map.values(): if offload_folder is None: @@ -1566,37 +1549,39 @@ class ModelMixin(torch.nn.Module, PushToHubMixin): # if state dict is not None, it means that we don't need to read the files from resolved_model_file also resolved_model_file = [state_dict] - if len(resolved_model_file) > 1: - resolved_model_file = logging.tqdm(resolved_model_file, desc="Loading checkpoint shards") + # Prepare the loading function sharing the attributes shared between them. + load_fn = functools.partial( + _load_shard_files_with_threadpool if is_parallel_loading_enabled else _load_shard_file, + model=model, + model_state_dict=model_state_dict, + device_map=device_map, + dtype=dtype, + hf_quantizer=hf_quantizer, + keep_in_fp32_modules=keep_in_fp32_modules, + dduf_entries=dduf_entries, + loaded_keys=loaded_keys, + unexpected_keys=unexpected_keys, + offload_index=offload_index, + offload_folder=offload_folder, + state_dict_index=state_dict_index, + state_dict_folder=state_dict_folder, + ignore_mismatched_sizes=ignore_mismatched_sizes, + low_cpu_mem_usage=low_cpu_mem_usage, + ) - mismatched_keys = [] - assign_to_params_buffers = None - error_msgs = [] + if is_parallel_loading_enabled: + offload_index, state_dict_index, _mismatched_keys, _error_msgs = load_fn(resolved_model_file) + error_msgs += _error_msgs + mismatched_keys += _mismatched_keys + else: + shard_files = resolved_model_file + if len(resolved_model_file) > 1: + shard_files = logging.tqdm(resolved_model_file, desc="Loading checkpoint shards") - for shard_file in resolved_model_file: - state_dict = load_state_dict(shard_file, dduf_entries=dduf_entries) - mismatched_keys += _find_mismatched_keys( - state_dict, model_state_dict, loaded_keys, ignore_mismatched_sizes - ) - - if low_cpu_mem_usage: - offload_index, state_dict_index = load_model_dict_into_meta( - model, - state_dict, - device_map=device_map, - dtype=dtype, - hf_quantizer=hf_quantizer, - keep_in_fp32_modules=keep_in_fp32_modules, - unexpected_keys=unexpected_keys, - offload_folder=offload_folder, - offload_index=offload_index, - state_dict_index=state_dict_index, - state_dict_folder=state_dict_folder, - ) - else: - if assign_to_params_buffers is None: - assign_to_params_buffers = check_support_param_buffer_assignment(model, state_dict) - error_msgs += _load_state_dict_into_model(model, state_dict, assign_to_params_buffers) + for shard_file in shard_files: + offload_index, state_dict_index, _mismatched_keys, _error_msgs = load_fn(shard_file) + error_msgs += _error_msgs + mismatched_keys += _mismatched_keys empty_device_cache() diff --git a/src/diffusers/utils/__init__.py b/src/diffusers/utils/__init__.py index 5f49f5e757..32bae015e3 100644 --- a/src/diffusers/utils/__init__.py +++ b/src/diffusers/utils/__init__.py @@ -20,11 +20,13 @@ from packaging import version from .. import __version__ from .constants import ( CONFIG_NAME, + DEFAULT_HF_PARALLEL_LOADING_WORKERS, DEPRECATED_REVISION_ARGS, DIFFUSERS_DYNAMIC_MODULE_NAME, FLAX_WEIGHTS_NAME, GGUF_FILE_EXTENSION, HF_MODULES_CACHE, + HF_PARALLEL_LOADING_FLAG, HUGGINGFACE_CO_RESOLVE_ENDPOINT, MIN_PEFT_VERSION, ONNX_EXTERNAL_WEIGHTS_NAME, diff --git a/src/diffusers/utils/constants.py b/src/diffusers/utils/constants.py index f8f04cc03a..6313d33ddd 100644 --- a/src/diffusers/utils/constants.py +++ b/src/diffusers/utils/constants.py @@ -43,6 +43,8 @@ DEPRECATED_REVISION_ARGS = ["fp16", "non-ema"] DIFFUSERS_REQUEST_TIMEOUT = 60 DIFFUSERS_ATTN_BACKEND = os.getenv("DIFFUSERS_ATTN_BACKEND", "native") DIFFUSERS_ATTN_CHECKS = os.getenv("DIFFUSERS_ATTN_CHECKS", "0") in ENV_VARS_TRUE_VALUES +DEFAULT_HF_PARALLEL_LOADING_WORKERS = 8 +HF_PARALLEL_LOADING_FLAG = "HF_ENABLE_PARALLEL_LOADING" # Below should be `True` if the current version of `peft` and `transformers` are compatible with # PEFT backend. Will automatically fall back to PEFT backend if the correct versions of the libraries are diff --git a/tests/models/test_modeling_common.py b/tests/models/test_modeling_common.py index 0254e7e8c8..0e16f95a42 100644 --- a/tests/models/test_modeling_common.py +++ b/tests/models/test_modeling_common.py @@ -1428,6 +1428,41 @@ class ModelTesterMixin: self.assertTrue(torch.allclose(base_output[0], new_output[0], atol=1e-5)) + @require_torch_accelerator + def test_sharded_checkpoints_with_parallel_loading(self): + torch.manual_seed(0) + config, inputs_dict = self.prepare_init_args_and_inputs_for_common() + model = self.model_class(**config).eval() + model = model.to(torch_device) + + base_output = model(**inputs_dict) + + model_size = compute_module_persistent_sizes(model)[""] + max_shard_size = int((model_size * 0.75) / (2**10)) # Convert to KB as these test models are small. + with tempfile.TemporaryDirectory() as tmp_dir: + model.cpu().save_pretrained(tmp_dir, max_shard_size=f"{max_shard_size}KB") + self.assertTrue(os.path.exists(os.path.join(tmp_dir, SAFE_WEIGHTS_INDEX_NAME))) + + # Now check if the right number of shards exists. First, let's get the number of shards. + # Since this number can be dependent on the model being tested, it's important that we calculate it + # instead of hardcoding it. + expected_num_shards = caculate_expected_num_shards(os.path.join(tmp_dir, SAFE_WEIGHTS_INDEX_NAME)) + actual_num_shards = len([file for file in os.listdir(tmp_dir) if file.endswith(".safetensors")]) + self.assertTrue(actual_num_shards == expected_num_shards) + + # Load with parallel loading + os.environ["HF_ENABLE_PARALLEL_LOADING"] = "yes" + new_model = self.model_class.from_pretrained(tmp_dir).eval() + new_model = new_model.to(torch_device) + + torch.manual_seed(0) + if "generator" in inputs_dict: + _, inputs_dict = self.prepare_init_args_and_inputs_for_common() + new_output = new_model(**inputs_dict) + self.assertTrue(torch.allclose(base_output[0], new_output[0], atol=1e-5)) + # set to no. + os.environ["HF_ENABLE_PARALLEL_LOADING"] = "no" + @require_torch_accelerator def test_sharded_checkpoints_device_map(self): if self.model_class._no_split_modules is None: