diff --git a/src/diffusers/__init__.py b/src/diffusers/__init__.py index 02dd42e4a580..bf74c701b2d5 100644 --- a/src/diffusers/__init__.py +++ b/src/diffusers/__init__.py @@ -655,6 +655,7 @@ "WuerstchenCombinedPipeline", "WuerstchenDecoderPipeline", "WuerstchenPriorPipeline", + "ZImageImg2ImgPipeline", "ZImagePipeline", ] ) @@ -1346,6 +1347,7 @@ WuerstchenCombinedPipeline, WuerstchenDecoderPipeline, WuerstchenPriorPipeline, + ZImageImg2ImgPipeline, ZImagePipeline, ) diff --git a/src/diffusers/pipelines/__init__.py b/src/diffusers/pipelines/__init__.py index cf86456642eb..90404fb74f69 100644 --- a/src/diffusers/pipelines/__init__.py +++ b/src/diffusers/pipelines/__init__.py @@ -397,7 +397,7 @@ "WanVACEPipeline", "WanAnimatePipeline", ] - _import_structure["z_image"] = ["ZImagePipeline"] + _import_structure["z_image"] = ["ZImageImg2ImgPipeline", "ZImagePipeline"] _import_structure["kandinsky5"] = ["Kandinsky5T2VPipeline"] _import_structure["skyreels_v2"] = [ "SkyReelsV2DiffusionForcingPipeline", @@ -829,7 +829,7 @@ WuerstchenDecoderPipeline, WuerstchenPriorPipeline, ) - from .z_image import ZImagePipeline + from .z_image import ZImageImg2ImgPipeline, ZImagePipeline try: if not is_onnx_available(): diff --git a/src/diffusers/pipelines/auto_pipeline.py b/src/diffusers/pipelines/auto_pipeline.py index 044d854390e4..2a549a2d3807 100644 --- a/src/diffusers/pipelines/auto_pipeline.py +++ b/src/diffusers/pipelines/auto_pipeline.py @@ -119,6 +119,7 @@ ) from .wan import WanImageToVideoPipeline, WanPipeline, WanVideoToVideoPipeline from .wuerstchen import WuerstchenCombinedPipeline, WuerstchenDecoderPipeline +from .z_image import ZImageImg2ImgPipeline AUTO_TEXT2IMAGE_PIPELINES_MAPPING = OrderedDict( @@ -189,6 +190,7 @@ ("qwenimage", QwenImageImg2ImgPipeline), ("qwenimage-edit", QwenImageEditPipeline), ("qwenimage-edit-plus", QwenImageEditPlusPipeline), + ("z-image", ZImageImg2ImgPipeline), ] ) diff --git a/src/diffusers/pipelines/z_image/__init__.py b/src/diffusers/pipelines/z_image/__init__.py index f95b3e5a0bed..f4342713e3e9 100644 --- a/src/diffusers/pipelines/z_image/__init__.py +++ b/src/diffusers/pipelines/z_image/__init__.py @@ -23,6 +23,7 @@ else: _import_structure["pipeline_output"] = ["ZImagePipelineOutput"] _import_structure["pipeline_z_image"] = ["ZImagePipeline"] + _import_structure["pipeline_z_image_img2img"] = ["ZImageImg2ImgPipeline"] if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: @@ -35,6 +36,7 @@ else: from .pipeline_output import ZImagePipelineOutput from .pipeline_z_image import ZImagePipeline + from .pipeline_z_image_img2img import ZImageImg2ImgPipeline else: import sys diff --git a/src/diffusers/pipelines/z_image/pipeline_z_image_img2img.py b/src/diffusers/pipelines/z_image/pipeline_z_image_img2img.py new file mode 100644 index 000000000000..c2261004a264 --- /dev/null +++ b/src/diffusers/pipelines/z_image/pipeline_z_image_img2img.py @@ -0,0 +1,707 @@ +# 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 logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from .pipeline_output import ZImagePipelineOutput + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import ZImageImg2ImgPipeline + >>> from diffusers.utils import load_image + + >>> pipe = ZImageImg2ImgPipeline.from_pretrained("Z-a-o/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)) + + >>> prompt = "A fantasy landscape with mountains and a river, detailed, vibrant colors" + >>> image = pipe( + ... prompt, + ... image=init_image, + ... strength=0.6, + ... num_inference_steps=9, + ... guidance_scale=0.0, + ... generator=torch.Generator("cuda").manual_seed(42), + ... ).images[0] + >>> image.save("zimage_img2img.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: 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.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 ZImageImg2ImgPipeline(DiffusionPipeline, ZImageLoraLoaderMixin, FromSingleFileMixin): + r""" + The ZImage pipeline for image-to-image generation. + + 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"] + + 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) + + def encode_prompt( + self, + prompt: Union[str, List[str]], + device: Optional[torch.device] = None, + do_classifier_free_guidance: bool = True, + negative_prompt: Optional[Union[str, List[str]]] = None, + prompt_embeds: Optional[List[torch.FloatTensor]] = None, + negative_prompt_embeds: Optional[torch.FloatTensor] = 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 + + def _encode_prompt( + self, + prompt: Union[str, List[str]], + device: Optional[torch.device] = None, + prompt_embeds: Optional[List[torch.FloatTensor]] = 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_latents( + self, + image, + timestep, + batch_size, + num_channels_latents, + height, + width, + dtype, + device, + generator, + latents=None, + ): + 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: + return latents.to(device=device, dtype=dtype) + + # 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." + ) + + # Add noise using flow matching scale_noise + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + latents = self.scheduler.scale_noise(image_latents, timestep, noise) + + return latents + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 + + @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 + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: Union[str, List[str]] = None, + image: PipelineImageInput = None, + strength: float = 0.6, + height: Optional[int] = None, + width: Optional[int] = None, + num_inference_steps: int = 50, + sigmas: Optional[List[float]] = 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: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + latents: Optional[torch.FloatTensor] = None, + prompt_embeds: Optional[List[torch.FloatTensor]] = None, + negative_prompt_embeds: Optional[List[torch.FloatTensor]] = None, + output_type: Optional[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 image-to-image 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. + 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)`. + strength (`float`, *optional*, defaults to 0.6): + 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`. + 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 and validate strength + if strength < 0 or strength > 1: + raise ValueError(f"The value of strength should be in [0.0, 1.0] but is {strength}") + + # 2. Preprocess image + 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}." + ) + + 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 + latents = self.prepare_latents( + init_image, + latent_timestep, + actual_batch_size, + num_channels_latents, + height, + width, + prompt_embeds[0].dtype, + device, + generator, + latents, + ) + + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self._num_timesteps = len(timesteps) + + # 8. 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 + + 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) + + # 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 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) diff --git a/src/diffusers/utils/dummy_torch_and_transformers_objects.py b/src/diffusers/utils/dummy_torch_and_transformers_objects.py index 65306b839015..cb6decbc648c 100644 --- a/src/diffusers/utils/dummy_torch_and_transformers_objects.py +++ b/src/diffusers/utils/dummy_torch_and_transformers_objects.py @@ -3692,6 +3692,21 @@ def from_pretrained(cls, *args, **kwargs): requires_backends(cls, ["torch", "transformers"]) +class ZImageImg2ImgPipeline(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 ZImagePipeline(metaclass=DummyObject): _backends = ["torch", "transformers"] diff --git a/tests/pipelines/z_image/test_z_image_img2img.py b/tests/pipelines/z_image/test_z_image_img2img.py new file mode 100644 index 000000000000..91b3025b17e8 --- /dev/null +++ b/tests/pipelines/z_image/test_z_image_img2img.py @@ -0,0 +1,358 @@ +# 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 gc +import os +import unittest + +import numpy as np +import torch +from transformers import Qwen2Tokenizer, Qwen3Config, Qwen3Model + +from diffusers import ( + AutoencoderKL, + FlowMatchEulerDiscreteScheduler, + ZImageImg2ImgPipeline, + ZImageTransformer2DModel, +) +from diffusers.utils.testing_utils import floats_tensor + +from ...testing_utils import torch_device +from ..pipeline_params import ( + IMAGE_TO_IMAGE_IMAGE_PARAMS, + TEXT_GUIDED_IMAGE_VARIATION_BATCH_PARAMS, + TEXT_GUIDED_IMAGE_VARIATION_PARAMS, +) +from ..test_pipelines_common import PipelineTesterMixin, to_np + + +# Z-Image requires torch.use_deterministic_algorithms(False) due to complex64 RoPE operations +# Cannot use enable_full_determinism() which sets it to True +# Note: Z-Image does not support FP16 inference due to complex64 RoPE embeddings +os.environ["CUDA_LAUNCH_BLOCKING"] = "1" +os.environ["CUBLAS_WORKSPACE_CONFIG"] = ":16:8" +torch.use_deterministic_algorithms(False) +torch.backends.cudnn.deterministic = True +torch.backends.cudnn.benchmark = False +if hasattr(torch.backends, "cuda"): + torch.backends.cuda.matmul.allow_tf32 = False + + +class ZImageImg2ImgPipelineFastTests(PipelineTesterMixin, unittest.TestCase): + pipeline_class = ZImageImg2ImgPipeline + params = TEXT_GUIDED_IMAGE_VARIATION_PARAMS - {"cross_attention_kwargs"} + batch_params = TEXT_GUIDED_IMAGE_VARIATION_BATCH_PARAMS + image_params = IMAGE_TO_IMAGE_IMAGE_PARAMS + image_latents_params = IMAGE_TO_IMAGE_IMAGE_PARAMS + required_optional_params = frozenset( + [ + "num_inference_steps", + "strength", + "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 setUp(self): + gc.collect() + if torch.cuda.is_available(): + torch.cuda.empty_cache() + torch.cuda.synchronize() + torch.manual_seed(0) + if torch.cuda.is_available(): + torch.cuda.manual_seed_all(0) + + def tearDown(self): + super().tearDown() + gc.collect() + if torch.cuda.is_available(): + torch.cuda.empty_cache() + torch.cuda.synchronize() + torch.manual_seed(0) + if torch.cuda.is_available(): + torch.cuda.manual_seed_all(0) + + def get_dummy_components(self): + torch.manual_seed(0) + transformer = ZImageTransformer2DModel( + all_patch_size=(2,), + all_f_patch_size=(1,), + in_channels=16, + dim=32, + n_layers=2, + n_refiner_layers=1, + n_heads=2, + n_kv_heads=2, + norm_eps=1e-5, + qk_norm=True, + cap_feat_dim=16, + rope_theta=256.0, + t_scale=1000.0, + axes_dims=[8, 4, 4], + axes_lens=[256, 32, 32], + ) + + torch.manual_seed(0) + vae = AutoencoderKL( + in_channels=3, + out_channels=3, + down_block_types=["DownEncoderBlock2D", "DownEncoderBlock2D"], + up_block_types=["UpDecoderBlock2D", "UpDecoderBlock2D"], + block_out_channels=[32, 64], + layers_per_block=1, + latent_channels=16, + norm_num_groups=32, + sample_size=32, + scaling_factor=0.3611, + shift_factor=0.1159, + ) + + torch.manual_seed(0) + scheduler = FlowMatchEulerDiscreteScheduler() + + torch.manual_seed(0) + config = Qwen3Config( + hidden_size=16, + intermediate_size=16, + num_hidden_layers=2, + num_attention_heads=2, + num_key_value_heads=2, + vocab_size=151936, + max_position_embeddings=512, + ) + text_encoder = Qwen3Model(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): + import random + + if str(device).startswith("mps"): + generator = torch.manual_seed(seed) + else: + generator = torch.Generator(device=device).manual_seed(seed) + + image = floats_tensor((1, 3, 32, 32), rng=random.Random(seed)).to(device) + + inputs = { + "prompt": "dance monkey", + "negative_prompt": "bad quality", + "image": image, + "strength": 0.6, + "generator": generator, + "num_inference_steps": 2, + "guidance_scale": 3.0, + "cfg_normalization": False, + "cfg_truncation": 1.0, + "height": 32, + "width": 32, + "max_sequence_length": 16, + "output_type": "np", + } + + 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, (32, 32, 3)) + + def test_inference_batch_single_identical(self): + gc.collect() + if torch.cuda.is_available(): + torch.cuda.empty_cache() + torch.cuda.synchronize() + torch.manual_seed(0) + if torch.cuda.is_available(): + torch.cuda.manual_seed_all(0) + self._test_inference_batch_single_identical(batch_size=3, expected_max_diff=1e-1) + + def test_num_images_per_prompt(self): + import inspect + + sig = inspect.signature(self.pipeline_class.__call__) + + if "num_images_per_prompt" not in sig.parameters: + return + + components = self.get_dummy_components() + pipe = self.pipeline_class(**components) + pipe = pipe.to(torch_device) + pipe.set_progress_bar_config(disable=None) + + batch_sizes = [1, 2] + num_images_per_prompts = [1, 2] + + for batch_size in batch_sizes: + for num_images_per_prompt in num_images_per_prompts: + inputs = self.get_dummy_inputs(torch_device) + + for key in inputs.keys(): + if key in self.batch_params: + inputs[key] = batch_size * [inputs[key]] + + images = pipe(**inputs, num_images_per_prompt=num_images_per_prompt)[0] + + assert images.shape[0] == batch_size * num_images_per_prompt + + del pipe + gc.collect() + if torch.cuda.is_available(): + torch.cuda.empty_cache() + torch.cuda.synchronize() + + 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 = 0.3): + import random + + 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 + # Generate a larger image for the input + inputs["image"] = floats_tensor((1, 3, 128, 128), rng=random.Random(0)).to("cpu") + output_without_tiling = pipe(**inputs)[0] + + # With tiling (standard AutoencoderKL doesn't accept parameters) + pipe.vae.enable_tiling() + inputs = self.get_dummy_inputs(generator_device) + inputs["height"] = inputs["width"] = 128 + inputs["image"] = floats_tensor((1, 3, 128, 128), rng=random.Random(0)).to("cpu") + 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", + ) + + def test_pipeline_with_accelerator_device_map(self, expected_max_difference=5e-4): + # Z-Image RoPE embeddings (complex64) have slightly higher numerical tolerance + super().test_pipeline_with_accelerator_device_map(expected_max_difference=expected_max_difference) + + def test_group_offloading_inference(self): + # Block-level offloading conflicts with RoPE cache. Pipeline-level offloading (tested separately) works fine. + self.skipTest("Using test_pipeline_level_group_offloading_inference instead") + + def test_save_load_float16(self, expected_max_diff=1e-2): + # Z-Image does not support FP16 due to complex64 RoPE embeddings + self.skipTest("Z-Image does not support FP16 inference") + + def test_float16_inference(self, expected_max_diff=5e-2): + # Z-Image does not support FP16 due to complex64 RoPE embeddings + self.skipTest("Z-Image does not support FP16 inference") + + def test_strength_parameter(self): + """Test that strength parameter affects the output correctly.""" + device = "cpu" + components = self.get_dummy_components() + pipe = self.pipeline_class(**components) + pipe.to(device) + pipe.set_progress_bar_config(disable=None) + + # Test with different strength values + inputs_low_strength = self.get_dummy_inputs(device) + inputs_low_strength["strength"] = 0.2 + + inputs_high_strength = self.get_dummy_inputs(device) + inputs_high_strength["strength"] = 0.8 + + # Both should complete without errors + output_low = pipe(**inputs_low_strength).images[0] + output_high = pipe(**inputs_high_strength).images[0] + + # Outputs should be different (different amount of transformation) + self.assertFalse(np.allclose(output_low, output_high, atol=1e-3)) + + def test_invalid_strength(self): + """Test that invalid strength values raise appropriate errors.""" + device = "cpu" + components = self.get_dummy_components() + pipe = self.pipeline_class(**components) + pipe.to(device) + + inputs = self.get_dummy_inputs(device) + + # Test strength < 0 + inputs["strength"] = -0.1 + with self.assertRaises(ValueError): + pipe(**inputs) + + # Test strength > 1 + inputs["strength"] = 1.5 + with self.assertRaises(ValueError): + pipe(**inputs)