# Copyright 2025 Alibaba Z-Image Team and The HuggingFace Team. All rights reserved. # # 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 inspect from typing import Any, Callable, Dict, List, Optional, Union import torch from transformers import AutoTokenizer, PreTrainedModel from ...image_processor import PipelineImageInput, VaeImageProcessor from ...loaders import FromSingleFileMixin, ZImageLoraLoaderMixin from ...models.autoencoders import AutoencoderKL from ...models.transformers import ZImageTransformer2DModel from ...pipelines.pipeline_utils import DiffusionPipeline from ...schedulers import FlowMatchEulerDiscreteScheduler from ...utils import is_torch_xla_available, logging, replace_example_docstring from ...utils.torch_utils import randn_tensor from .pipeline_output import ZImagePipelineOutput 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 ZImageInpaintPipeline >>> from diffusers.utils import load_image >>> pipe = ZImageInpaintPipeline.from_pretrained("Tongyi-MAI/Z-Image-Turbo", torch_dtype=torch.bfloat16) >>> 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)) >>> # Create a mask (white = inpaint, black = preserve) >>> import numpy as np >>> from PIL import Image >>> mask = np.zeros((1024, 1024), dtype=np.uint8) >>> mask[256:768, 256:768] = 255 # Inpaint center region >>> mask_image = Image.fromarray(mask) >>> prompt = "A beautiful lake with mountains in the background" >>> image = pipe( ... prompt, ... image=init_image, ... mask_image=mask_image, ... strength=1.0, ... num_inference_steps=9, ... guidance_scale=0.0, ... generator=torch.Generator("cuda").manual_seed(42), ... ).images[0] >>> image.save("zimage_inpaint.png") ``` """ # Copied from diffusers.pipelines.flux.pipeline_flux.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_img2img.retrieve_latents def retrieve_latents( encoder_output: torch.Tensor, generator: torch.Generator | None = 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.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps def retrieve_timesteps( scheduler, num_inference_steps: int | None = None, device: str | torch.device | None = None, timesteps: list[int] | None = None, sigmas: list[float] | None = 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 ZImageInpaintPipeline(DiffusionPipeline, ZImageLoraLoaderMixin, FromSingleFileMixin): r""" The ZImage pipeline for inpainting. Args: 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 ([`PreTrainedModel`]): A text encoder model to encode text prompts. tokenizer ([`AutoTokenizer`]): A tokenizer to tokenize text prompts. transformer ([`ZImageTransformer2DModel`]): A ZImage transformer model to denoise the encoded image latents. """ model_cpu_offload_seq = "text_encoder->transformer->vae" _optional_components = [] _callback_tensor_inputs = ["latents", "prompt_embeds", "mask", "masked_image_latents"] def __init__( self, scheduler: FlowMatchEulerDiscreteScheduler, vae: AutoencoderKL, text_encoder: PreTrainedModel, tokenizer: AutoTokenizer, transformer: ZImageTransformer2DModel, ): super().__init__() self.register_modules( vae=vae, text_encoder=text_encoder, tokenizer=tokenizer, scheduler=scheduler, transformer=transformer, ) self.vae_scale_factor = ( 2 ** (len(self.vae.config.block_out_channels) - 1) if hasattr(self, "vae") and self.vae is not None else 8 ) self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor * 2) self.mask_processor = VaeImageProcessor( vae_scale_factor=self.vae_scale_factor * 2, do_normalize=False, do_binarize=True, do_convert_grayscale=True, ) # Copied from diffusers.pipelines.z_image.pipeline_z_image.ZImagePipeline.encode_prompt def encode_prompt( self, prompt: str | list[str], device: torch.device | None = None, do_classifier_free_guidance: bool = True, negative_prompt: str | list[str] | None = None, prompt_embeds: list[torch.FloatTensor] | None = None, negative_prompt_embeds: torch.FloatTensor | None = None, max_sequence_length: int = 512, ): prompt = [prompt] if isinstance(prompt, str) else prompt prompt_embeds = self._encode_prompt( prompt=prompt, device=device, prompt_embeds=prompt_embeds, max_sequence_length=max_sequence_length, ) if do_classifier_free_guidance: if negative_prompt is None: negative_prompt = ["" for _ in prompt] else: negative_prompt = [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt assert len(prompt) == len(negative_prompt) negative_prompt_embeds = self._encode_prompt( prompt=negative_prompt, device=device, prompt_embeds=negative_prompt_embeds, max_sequence_length=max_sequence_length, ) else: negative_prompt_embeds = [] return prompt_embeds, negative_prompt_embeds # Copied from diffusers.pipelines.z_image.pipeline_z_image.ZImagePipeline._encode_prompt def _encode_prompt( self, prompt: str | list[str], device: torch.device | None = None, prompt_embeds: list[torch.FloatTensor] | None = None, max_sequence_length: int = 512, ) -> list[torch.FloatTensor]: device = device or self._execution_device if prompt_embeds is not None: return prompt_embeds if isinstance(prompt, str): prompt = [prompt] for i, prompt_item in enumerate(prompt): messages = [ {"role": "user", "content": prompt_item}, ] prompt_item = self.tokenizer.apply_chat_template( messages, tokenize=False, add_generation_prompt=True, enable_thinking=True, ) prompt[i] = prompt_item text_inputs = self.tokenizer( prompt, padding="max_length", max_length=max_sequence_length, truncation=True, return_tensors="pt", ) text_input_ids = text_inputs.input_ids.to(device) prompt_masks = text_inputs.attention_mask.to(device).bool() prompt_embeds = self.text_encoder( input_ids=text_input_ids, attention_mask=prompt_masks, output_hidden_states=True, ).hidden_states[-2] embeddings_list = [] for i in range(len(prompt_embeds)): embeddings_list.append(prompt_embeds[i][prompt_masks[i]]) return embeddings_list # 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 def prepare_mask_latents( self, mask, masked_image, batch_size, height, width, dtype, device, generator, ): """Prepare mask and masked image latents for inpainting. Args: mask: Binary mask tensor where 1 = inpaint region, 0 = preserve region. masked_image: Original image with masked regions zeroed out. batch_size: Number of images to generate. height: Output image height. width: Output image width. dtype: Data type for the tensors. device: Device to place tensors on. generator: Random generator for reproducibility. Returns: Tuple of (mask, masked_image_latents) prepared for the denoising loop. """ # Calculate latent dimensions latent_height = 2 * (int(height) // (self.vae_scale_factor * 2)) latent_width = 2 * (int(width) // (self.vae_scale_factor * 2)) # Resize mask to latent dimensions mask = torch.nn.functional.interpolate(mask, size=(latent_height, latent_width), mode="nearest") mask = mask.to(device=device, dtype=dtype) # Encode masked image to latents masked_image = masked_image.to(device=device, dtype=dtype) if isinstance(generator, list): masked_image_latents = [ retrieve_latents(self.vae.encode(masked_image[i : i + 1]), generator=generator[i]) for i in range(masked_image.shape[0]) ] masked_image_latents = torch.cat(masked_image_latents, dim=0) else: masked_image_latents = retrieve_latents(self.vae.encode(masked_image), generator=generator) # Apply VAE scaling masked_image_latents = (masked_image_latents - self.vae.config.shift_factor) * self.vae.config.scaling_factor # Expand for batch size 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) return mask, masked_image_latents def prepare_latents( self, image, timestep, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None, ): """Prepare latents for inpainting, returning noise and image_latents for blending. Returns: Tuple of (latents, noise, image_latents) where: - latents: Noised image latents for denoising - noise: The noise tensor used for blending - image_latents: Clean image latents for blending """ height = 2 * (int(height) // (self.vae_scale_factor * 2)) width = 2 * (int(width) // (self.vae_scale_factor * 2)) shape = (batch_size, num_channels_latents, height, width) if latents is not None: # Generate noise for blending even if latents are provided noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) # Encode image for blending image = image.to(device=device, dtype=dtype) 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) image_latents = (image_latents - self.vae.config.shift_factor) * self.vae.config.scaling_factor if batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] == 0: image_latents = torch.cat([image_latents] * (batch_size // image_latents.shape[0]), dim=0) return latents.to(device=device, dtype=dtype), noise, image_latents # Encode the input image image = image.to(device=device, dtype=dtype) if image.shape[1] != num_channels_latents: 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) # Apply scaling (inverse of decoding: decode does latents/scaling_factor + shift_factor) image_latents = (image_latents - self.vae.config.shift_factor) * self.vae.config.scaling_factor else: image_latents = image # Handle batch size expansion if batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] == 0: 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." ) # Generate noise for both initial noising and later blending noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) # Add noise using flow matching scale_noise latents = self.scheduler.scale_noise(image_latents, timestep, noise) return latents, noise, image_latents @property def guidance_scale(self): return self._guidance_scale @property def do_classifier_free_guidance(self): return self._guidance_scale > 0 @property def joint_attention_kwargs(self): return self._joint_attention_kwargs @property def num_timesteps(self): return self._num_timesteps @property def interrupt(self): return self._interrupt def check_inputs( self, prompt, image, mask_image, strength, height, width, output_type, negative_prompt=None, prompt_embeds=None, negative_prompt_embeds=None, callback_on_step_end_tensor_inputs=None, ): if strength < 0 or strength > 1: raise ValueError(f"The value of strength should be in [0.0, 1.0] but is {strength}") 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 image is None: raise ValueError("`image` input cannot be undefined for inpainting.") if mask_image is None: raise ValueError("`mask_image` input cannot be undefined for inpainting.") if output_type not in ["latent", "pil", "np", "pt"]: raise ValueError(f"`output_type` must be one of 'latent', 'pil', 'np', or 'pt', but got {output_type}") @torch.no_grad() @replace_example_docstring(EXAMPLE_DOC_STRING) def __call__( self, prompt: Union[str, List[str]] = None, image: PipelineImageInput = None, mask_image: PipelineImageInput = None, masked_image_latents: Optional[torch.FloatTensor] = None, strength: float = 1.0, height: int | None = None, width: int | None = None, num_inference_steps: int = 50, sigmas: list[float] | None = None, guidance_scale: float = 5.0, cfg_normalization: bool = False, cfg_truncation: float = 1.0, negative_prompt: Optional[Union[str, List[str]]] = None, num_images_per_prompt: Optional[int] = 1, generator: torch.Generator | list[torch.Generator] | None = None, latents: Optional[torch.FloatTensor] = None, prompt_embeds: Optional[List[torch.FloatTensor]] = None, negative_prompt_embeds: Optional[List[torch.FloatTensor]] = None, output_type: str = "pil", return_dict: bool = True, joint_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 inpainting. 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. 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 of 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)`. 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 a mask image for inpainting. White pixels (value 1) in the mask will be inpainted, black pixels (value 0) will be preserved from the original image. masked_image_latents (`torch.FloatTensor`, *optional*): Pre-encoded masked image latents. If provided, the masked image encoding step will be skipped. 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` in the masked region. height (`int`, *optional*, defaults to 1024): The height in pixels of the generated image. If not provided, uses the input image height. width (`int`, *optional*, defaults to 1024): The width in pixels of the generated image. If not provided, uses the input image width. 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 5.0): Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598). `guidance_scale` is defined as `w` of equation 2. of [Imagen Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`, usually at the expense of lower image quality. cfg_normalization (`bool`, *optional*, defaults to False): Whether to apply configuration normalization. cfg_truncation (`float`, *optional*, defaults to 1.0): The truncation value for configuration. negative_prompt (`str` or `List[str]`, *optional*): The prompt or prompts not to guide the image generation. If not defined, one has to pass `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is less than `1`). num_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.FloatTensor`, *optional*): Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image generation. Can be used to tweak the same generation with different prompts. If not provided, a latents tensor will be generated by sampling using the supplied random `generator`. prompt_embeds (`List[torch.FloatTensor]`, *optional*): Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not provided, text embeddings will be generated from `prompt` input argument. negative_prompt_embeds (`List[torch.FloatTensor]`, *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.stable_diffusion.ZImagePipelineOutput`] instead of a plain tuple. joint_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`, *optional*, defaults to 512): Maximum sequence length to use with the `prompt`. Examples: Returns: [`~pipelines.z_image.ZImagePipelineOutput`] or `tuple`: [`~pipelines.z_image.ZImagePipelineOutput`] if `return_dict` is True, otherwise a `tuple`. When returning a tuple, the first element is a list with the generated images. """ # 1. Check inputs self.check_inputs( prompt=prompt, image=image, mask_image=mask_image, strength=strength, height=height, width=width, output_type=output_type, negative_prompt=negative_prompt, prompt_embeds=prompt_embeds, negative_prompt_embeds=negative_prompt_embeds, callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, ) # 2. Preprocess image and mask init_image = self.image_processor.preprocess(image) init_image = init_image.to(dtype=torch.float32) # Get dimensions from the preprocessed image if not specified if height is None: height = init_image.shape[-2] if width is None: width = init_image.shape[-1] vae_scale = self.vae_scale_factor * 2 if height % vae_scale != 0: raise ValueError( f"Height must be divisible by {vae_scale} (got {height}). " f"Please adjust the height to a multiple of {vae_scale}." ) if width % vae_scale != 0: raise ValueError( f"Width must be divisible by {vae_scale} (got {width}). " f"Please adjust the width to a multiple of {vae_scale}." ) # Preprocess mask mask = self.mask_processor.preprocess(mask_image, height=height, width=width) device = self._execution_device self._guidance_scale = guidance_scale self._joint_attention_kwargs = joint_attention_kwargs self._interrupt = False self._cfg_normalization = cfg_normalization self._cfg_truncation = cfg_truncation # 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 = len(prompt_embeds) # If prompt_embeds is provided and prompt is None, skip encoding if prompt_embeds is not None and prompt is None: if self.do_classifier_free_guidance and negative_prompt_embeds is None: raise ValueError( "When `prompt_embeds` is provided without `prompt`, " "`negative_prompt_embeds` must also be provided for classifier-free guidance." ) else: ( prompt_embeds, negative_prompt_embeds, ) = self.encode_prompt( prompt=prompt, negative_prompt=negative_prompt, do_classifier_free_guidance=self.do_classifier_free_guidance, prompt_embeds=prompt_embeds, negative_prompt_embeds=negative_prompt_embeds, device=device, max_sequence_length=max_sequence_length, ) # 4. Prepare latent variables num_channels_latents = self.transformer.in_channels # Repeat prompt_embeds for num_images_per_prompt if num_images_per_prompt > 1: prompt_embeds = [pe for pe in prompt_embeds for _ in range(num_images_per_prompt)] if self.do_classifier_free_guidance and negative_prompt_embeds: negative_prompt_embeds = [npe for npe in negative_prompt_embeds for _ in range(num_images_per_prompt)] actual_batch_size = batch_size * num_images_per_prompt # Calculate latent dimensions for image_seq_len latent_height = 2 * (int(height) // (self.vae_scale_factor * 2)) latent_width = 2 * (int(width) // (self.vae_scale_factor * 2)) image_seq_len = (latent_height // 2) * (latent_width // 2) # 5. Prepare timesteps 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), ) self.scheduler.sigma_min = 0.0 scheduler_kwargs = {"mu": mu} timesteps, num_inference_steps = retrieve_timesteps( self.scheduler, num_inference_steps, device, sigmas=sigmas, **scheduler_kwargs, ) # 6. Adjust timesteps based on strength 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(actual_batch_size) # 7. Prepare latents from image (returns noise and image_latents for blending) latents, noise, image_latents = self.prepare_latents( init_image, latent_timestep, actual_batch_size, num_channels_latents, height, width, prompt_embeds[0].dtype, device, generator, latents, ) # 8. Prepare mask and masked image latents # Create masked image: preserve only unmasked regions (mask=0) if masked_image_latents is None: masked_image = init_image * (mask < 0.5) else: masked_image = None # Will use provided masked_image_latents mask, masked_image_latents = self.prepare_mask_latents( mask, masked_image if masked_image is not None else init_image, actual_batch_size, height, width, prompt_embeds[0].dtype, device, generator, ) num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) self._num_timesteps = len(timesteps) # 9. Denoising loop with self.progress_bar(total=num_inference_steps) as progress_bar: for i, t in enumerate(timesteps): if self.interrupt: continue # broadcast to batch dimension in a way that's compatible with ONNX/Core ML timestep = t.expand(latents.shape[0]) timestep = (1000 - timestep) / 1000 # Normalized time for time-aware config (0 at start, 1 at end) t_norm = timestep[0].item() # Handle cfg truncation current_guidance_scale = self.guidance_scale if ( self.do_classifier_free_guidance and self._cfg_truncation is not None and float(self._cfg_truncation) <= 1 ): if t_norm > self._cfg_truncation: current_guidance_scale = 0.0 # Run CFG only if configured AND scale is non-zero apply_cfg = self.do_classifier_free_guidance and current_guidance_scale > 0 if apply_cfg: latents_typed = latents.to(self.transformer.dtype) latent_model_input = latents_typed.repeat(2, 1, 1, 1) prompt_embeds_model_input = prompt_embeds + negative_prompt_embeds timestep_model_input = timestep.repeat(2) else: latent_model_input = latents.to(self.transformer.dtype) prompt_embeds_model_input = prompt_embeds timestep_model_input = timestep latent_model_input = latent_model_input.unsqueeze(2) latent_model_input_list = list(latent_model_input.unbind(dim=0)) model_out_list = self.transformer( latent_model_input_list, timestep_model_input, prompt_embeds_model_input, )[0] if apply_cfg: # Perform CFG pos_out = model_out_list[:actual_batch_size] neg_out = model_out_list[actual_batch_size:] noise_pred = [] for j in range(actual_batch_size): pos = pos_out[j].float() neg = neg_out[j].float() pred = pos + current_guidance_scale * (pos - neg) # Renormalization if self._cfg_normalization and float(self._cfg_normalization) > 0.0: ori_pos_norm = torch.linalg.vector_norm(pos) new_pos_norm = torch.linalg.vector_norm(pred) max_new_norm = ori_pos_norm * float(self._cfg_normalization) if new_pos_norm > max_new_norm: pred = pred * (max_new_norm / new_pos_norm) noise_pred.append(pred) noise_pred = torch.stack(noise_pred, dim=0) else: noise_pred = torch.stack([t.float() for t in model_out_list], dim=0) noise_pred = noise_pred.squeeze(2) noise_pred = -noise_pred # compute the previous noisy sample x_t -> x_t-1 latents = self.scheduler.step(noise_pred.to(torch.float32), t, latents, return_dict=False)[0] assert latents.dtype == torch.float32 # Inpainting blend: combine denoised latents with original image latents init_latents_proper = image_latents # Re-scale original latents to current noise level for proper blending 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 ) # Blend: mask=1 for inpaint region (use denoised), mask=0 for preserve region (use original) latents = (1 - mask) * init_latents_proper + mask * latents 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) negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) mask = callback_outputs.pop("mask", mask) masked_image_latents = callback_outputs.pop("masked_image_latents", masked_image_latents) # 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() if output_type == "latent": image = latents else: latents = latents.to(self.vae.dtype) latents = (latents / self.vae.config.scaling_factor) + self.vae.config.shift_factor image = self.vae.decode(latents, return_dict=False)[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 ZImagePipelineOutput(images=image)