guard accelerate import.

* LTX Video 0.9.8  long multi prompt

* Further align comfyui

- Added the “LTXEulerAncestralRFScheduler” scheduler, aligned with [sample_euler_ancestral_RF](7d6103325e/comfy/k_diffusion/sampling.py (L234))

- Updated the LTXI2VLongMultiPromptPipeline.from_pretrained() method:
  - Now uses LTXEulerAncestralRFScheduler by default, for better compatibility with the ComfyUI LTXV workflow.

- Changed the default value of cond_strength from 1.0 to 0.5, aligning with ComfyUI’s default.

- Optimized cross-window overlap blending: moved the latent-space guidance injection to before the UNet and after each step, aligned with[KSamplerX0Inpaint]([ComfyUI/comfy/samplers.py at master · comfyanonymous/ComfyUI](https://github.com/comfyanonymous/ComfyUI/blob/master/comfy/samplers.py#L391))

- Adjusted the default value of skip_steps_sigma_threshold to 1.

* align with diffusers contribute rule

* Add new pipelines and update imports

* Enhance LTXI2VLongMultiPromptPipeline with noise rescaling

Refactor LTXI2VLongMultiPromptPipeline to improve documentation and add noise rescaling functionality.

* Clean up comments in scheduling_ltx_euler_ancestral_rf.py

Removed design notes and limitations from the implementation.

* Enhance video generation example with scheduler

Updated LTXI2VLongMultiPromptPipeline example to include LTXEulerAncestralRFScheduler for ComfyUI parity.

* clean up

* style

* copies

* import ltx scheduler

* copies

* fix

* fix more

* up up

* up up up

* up upup

* Apply suggestions from code review

* Update docs/source/en/api/pipelines/ltx_video.md

* Update docs/source/en/api/pipelines/ltx_video.md

---------

Co-authored-by: yiyixuxu <yixu310@gmail.com>

.github/workflows/codeql.yml

@@ -0,0 +1,22 @@
+---
+name: CodeQL Security Analysis For Github Actions
+
+on:
+  push:
+    branches: ["main"]
+  workflow_dispatch:
+  # pull_request:
+
+jobs:
+  codeql:
+    name: CodeQL Analysis
+    uses: huggingface/security-workflows/.github/workflows/codeql-reusable.yml@v1
+    permissions:
+      security-events: write
+      packages: read
+      actions: read
+      contents: read
+    with:
+      languages: '["actions","python"]'
+      queries: 'security-extended,security-and-quality'
+      runner: 'ubuntu-latest' #optional if need custom runner

.github/workflows/mirror_community_pipeline.yml

@@ -24,7 +24,6 @@ jobs:
   mirror_community_pipeline:
     env:
       SLACK_WEBHOOK_URL: ${{ secrets.SLACK_WEBHOOK_URL_COMMUNITY_MIRROR }}
-
     runs-on: ubuntu-22.04
     steps:
       # Checkout to correct ref
@@ -39,25 +38,28 @@ jobs:
       #     If ref is 'refs/heads/main' => set 'main'
       #     Else it must be a tag => set {tag}
       - name: Set checkout_ref and path_in_repo
+          EVENT_NAME: ${{ github.event_name }}
+          EVENT_INPUT_REF: ${{ github.event.inputs.ref }}
+          GITHUB_REF: ${{ github.ref }}
         run: |
-          if [ "${{ github.event_name }}" == "workflow_dispatch" ]; then
-            if [ -z "${{ github.event.inputs.ref }}" ]; then
+          if [ "$EVENT_NAME" == "workflow_dispatch" ]; then
+            if [ -z "$EVENT_INPUT_REF" ]; then
               echo "Error: Missing ref input"
               exit 1
-            elif [ "${{ github.event.inputs.ref }}" == "main" ]; then
+            elif [ "$EVENT_INPUT_REF" == "main" ]; then
               echo "CHECKOUT_REF=refs/heads/main" >> $GITHUB_ENV
               echo "PATH_IN_REPO=main" >> $GITHUB_ENV
             else
-              echo "CHECKOUT_REF=refs/tags/${{ github.event.inputs.ref }}" >> $GITHUB_ENV
-              echo "PATH_IN_REPO=${{ github.event.inputs.ref }}" >> $GITHUB_ENV
+              echo "CHECKOUT_REF=refs/tags/$EVENT_INPUT_REF" >> $GITHUB_ENV
+              echo "PATH_IN_REPO=$EVENT_INPUT_REF" >> $GITHUB_ENV
             fi
-          elif [ "${{ github.ref }}" == "refs/heads/main" ]; then
-            echo "CHECKOUT_REF=${{ github.ref }}" >> $GITHUB_ENV
+          elif [ "$GITHUB_REF" == "refs/heads/main" ]; then
+            echo "CHECKOUT_REF=$GITHUB_REF" >> $GITHUB_ENV
             echo "PATH_IN_REPO=main" >> $GITHUB_ENV
           else
             # e.g. refs/tags/v0.28.1 -> v0.28.1
-            echo "CHECKOUT_REF=${{ github.ref }}" >> $GITHUB_ENV
-            echo "PATH_IN_REPO=$(echo ${{ github.ref }} | sed 's/^refs\/tags\///')" >> $GITHUB_ENV
+            echo "CHECKOUT_REF=$GITHUB_REF" >> $GITHUB_ENV
+            echo "PATH_IN_REPO=$(echo $GITHUB_REF | sed 's/^refs\/tags\///')" >> $GITHUB_ENV
           fi
       - name: Print env vars
         run: |
@@ -99,4 +101,4 @@ jobs:
       - name: Report failure status
         if: ${{ failure() }}
         run: |
-          pip install requests && python utils/notify_community_pipelines_mirror.py --status=failure
+          pip install requests && python utils/notify_community_pipelines_mirror.py --status=failure

docs/source/en/api/pipelines/ltx_video.md

@@ -136,7 +136,7 @@ export_to_video(video, "output.mp4", fps=24)
   - The recommended dtype for the transformer, VAE, and text encoder is `torch.bfloat16`. The VAE and text encoder can also be `torch.float32` or `torch.float16`.
   - For guidance-distilled variants of LTX-Video, set `guidance_scale` to `1.0`. The `guidance_scale` for any other model should be set higher, like `5.0`, for good generation quality.
   - For timestep-aware VAE variants (LTX-Video 0.9.1 and above), set `decode_timestep` to `0.05` and `image_cond_noise_scale` to `0.025`.
-  - For variants that support interpolation between multiple conditioning images and videos (LTX-Video 0.9.5 and above), use similar images and videos for the best results. Divergence from the conditioning inputs may lead to abrupt transitionts in the generated video.
+  - For variants that support interpolation between multiple conditioning images and videos (LTX-Video 0.9.5 and above), use similar images and videos for the best results. Divergence from the conditioning inputs may lead to abrupt transitions in the generated video.
 
 - LTX-Video 0.9.7 includes a spatial latent upscaler and a 13B parameter transformer. During inference, a low resolution video is quickly generated first and then upscaled and refined.
 
@@ -329,7 +329,7 @@ export_to_video(video, "output.mp4", fps=24)
 
   <details>
   <summary>Show example code</summary>
-  
+
   ```python
   import torch
   from diffusers import LTXConditionPipeline, LTXLatentUpsamplePipeline
@@ -474,6 +474,12 @@ export_to_video(video, "output.mp4", fps=24)
 
   </details>
 
+## LTXI2VLongMultiPromptPipeline
+
+[[autodoc]] LTXI2VLongMultiPromptPipeline
+  - all
+  - __call__
+
 ## LTXPipeline
 
 [[autodoc]] LTXPipeline

docs/source/en/training/distributed_inference.md

@@ -263,8 +263,8 @@ def main():
     world_size = dist.get_world_size()
 
     pipeline = DiffusionPipeline.from_pretrained(
-        "black-forest-labs/FLUX.1-dev", torch_dtype=torch.bfloat16, device_map=device
-    )
+        "black-forest-labs/FLUX.1-dev", torch_dtype=torch.bfloat16
+    ).to(device)
     pipeline.transformer.set_attention_backend("_native_cudnn")
 
     cp_config = ContextParallelConfig(ring_degree=world_size)

examples/community/pipeline_hunyuandit_differential_img2img.py

@@ -21,8 +21,8 @@ from transformers import (
     BertModel,
     BertTokenizer,
     CLIPImageProcessor,
-    MT5Tokenizer,
     T5EncoderModel,
+    T5Tokenizer,
 )
 
 from diffusers.callbacks import MultiPipelineCallbacks, PipelineCallback
@@ -260,7 +260,7 @@ class HunyuanDiTDifferentialImg2ImgPipeline(DiffusionPipeline):
             The HunyuanDiT model designed by Tencent Hunyuan.
         text_encoder_2 (`T5EncoderModel`):
             The mT5 embedder. Specifically, it is 't5-v1_1-xxl'.
-        tokenizer_2 (`MT5Tokenizer`):
+        tokenizer_2 (`T5Tokenizer`):
             The tokenizer for the mT5 embedder.
         scheduler ([`DDPMScheduler`]):
             A scheduler to be used in combination with HunyuanDiT to denoise the encoded image latents.
@@ -295,7 +295,7 @@ class HunyuanDiTDifferentialImg2ImgPipeline(DiffusionPipeline):
         feature_extractor: CLIPImageProcessor,
         requires_safety_checker: bool = True,
         text_encoder_2=T5EncoderModel,
-        tokenizer_2=MT5Tokenizer,
+        tokenizer_2=T5Tokenizer,
     ):
         super().__init__()
 

examples/community/pipeline_z_image_differential_img2img.py

@@ -0,0 +1,844 @@
+# 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 diffusers.image_processor import PipelineImageInput, VaeImageProcessor
+from diffusers.loaders import FromSingleFileMixin, ZImageLoraLoaderMixin
+from diffusers.models.autoencoders import AutoencoderKL
+from diffusers.models.transformers import ZImageTransformer2DModel
+from diffusers.pipelines.pipeline_utils import DiffusionPipeline
+from diffusers.pipelines.z_image.pipeline_output import ZImagePipelineOutput
+from diffusers.schedulers import FlowMatchEulerDiscreteScheduler
+from diffusers.utils import logging, replace_example_docstring
+from diffusers.utils.torch_utils import randn_tensor
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+    Examples:
+        ```py
+        >>> import torch
+        >>> from pipeline_z_image_differential_img2img import ZImageDifferentialImg2ImgPipeline
+        >>> from diffusers.utils import load_image
+
+        >>> pipe = ZImageDifferentialImg2ImgPipeline.from_pretrained("Z-a-o/Z-Image-Turbo", torch_dtype=torch.bfloat16)
+        >>> pipe.to("cuda")
+
+        >>> init_image = load_image(
+        >>>     "https://github.com/exx8/differential-diffusion/blob/main/assets/input.jpg?raw=true",
+        >>> )
+
+        >>> mask = load_image(
+        >>>     "https://github.com/exx8/differential-diffusion/blob/main/assets/map.jpg?raw=true",
+        >>> )
+
+        >>> prompt = "painting of a mountain landscape with a meadow and a forest, meadow background, anime countryside landscape, anime nature wallpap, anime landscape wallpaper, studio ghibli landscape, anime landscape, mountain behind meadow, anime background art, studio ghibli environment, background of flowery hill, anime beautiful peace scene, forrest background, anime scenery, landscape background, background art, anime scenery concept art"
+
+        >>> image = pipe(
+        ...     prompt,
+        ...     image=init_image,
+        ...     mask_image=mask,
+        ...     strength=0.75,
+        ...     num_inference_steps=9,
+        ...     guidance_scale=0.0,
+        ...     generator=torch.Generator("cuda").manual_seed(41),
+        ... ).images[0]
+        >>> image.save("image.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 ZImageDifferentialImg2ImgPipeline(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
+        )
+        latent_channels = self.vae.config.latent_channels if getattr(self, "vae", None) else 16
+        self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor * 2)
+
+        self.mask_processor = VaeImageProcessor(
+            vae_scale_factor=self.vae_scale_factor,
+            vae_latent_channels=latent_channels,
+            do_normalize=False,
+            do_binarize=False,
+            do_convert_grayscale=True,
+        )
+
+    # Copied from diffusers.pipelines.z_image.pipeline_z_image.ZImagePipeline.encode_prompt
+    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
+
+    # Copied from diffusers.pipelines.z_image.pipeline_z_image.ZImagePipeline._encode_prompt
+    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
+
+    @staticmethod
+    def _prepare_latent_image_ids(batch_size, height, width, device, dtype):
+        latent_image_ids = torch.zeros(height // 2, width // 2, 3)
+        latent_image_ids[..., 1] = latent_image_ids[..., 1] + torch.arange(height // 2)[:, None]
+        latent_image_ids[..., 2] = latent_image_ids[..., 2] + torch.arange(width // 2)[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)
+
+    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)
+        latent_image_ids = self._prepare_latent_image_ids(batch_size, height, width, device, dtype)
+
+        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, noise, image_latents, latent_image_ids
+
+    def prepare_mask_latents(
+        self,
+        mask,
+        masked_image,
+        batch_size,
+        num_images_per_prompt,
+        height,
+        width,
+        dtype,
+        device,
+        generator,
+    ):
+        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
+
+        masked_image = masked_image.to(device=device, dtype=dtype)
+
+        if masked_image.shape[1] == 16:
+            masked_image_latents = masked_image
+        else:
+            masked_image_latents = retrieve_latents(self.vae.encode(masked_image), generator=generator)
+
+        masked_image_latents = (masked_image_latents - self.vae.config.shift_factor) * self.vae.config.scaling_factor
+
+        # 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)
+
+        # 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)
+
+        return mask, masked_image_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,
+        mask_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)`.
+            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`. Black pixels in the mask
+                are repainted while white 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)`.
+            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, noise, original_image_latents, latent_image_ids = self.prepare_latents(
+            init_image,
+            latent_timestep,
+            actual_batch_size,
+            num_channels_latents,
+            height,
+            width,
+            prompt_embeds[0].dtype,
+            device,
+            generator,
+            latents,
+        )
+        resize_mode = "default"
+        crops_coords = None
+
+        # start diff diff preparation
+        original_mask = self.mask_processor.preprocess(
+            mask_image, height=height, width=width, resize_mode=resize_mode, crops_coords=crops_coords
+        )
+
+        masked_image = init_image * original_mask
+        original_mask, _ = self.prepare_mask_latents(
+            original_mask,
+            masked_image,
+            batch_size,
+            num_images_per_prompt,
+            height,
+            width,
+            prompt_embeds[0].dtype,
+            device,
+            generator,
+        )
+        mask_thresholds = torch.arange(num_inference_steps, dtype=original_mask.dtype) / num_inference_steps
+        mask_thresholds = mask_thresholds.reshape(-1, 1, 1, 1).to(device)
+        masks = original_mask > mask_thresholds
+        # end diff diff preparation
+
+        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
+
+                # start diff diff
+                image_latent = original_image_latents
+                latents_dtype = latents.dtype
+                if i < len(timesteps) - 1:
+                    noise_timestep = timesteps[i + 1]
+                    image_latent = self.scheduler.scale_noise(
+                        original_image_latents, torch.tensor([noise_timestep]), noise
+                    )
+
+                    mask = masks[i].to(latents_dtype)
+                    latents = image_latent * mask + latents * (1 - mask)
+                # end diff diff
+
+                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)
+                    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)

examples/dreambooth/README_flux2.md

@@ -98,6 +98,9 @@ Flux.2 uses  Mistral Small 3.1 as text encoder which is quite large and can take
 This way, the text encoder model is not loaded into memory during training.
 > [!NOTE] 
 > to enable remote text encoding you must either be logged in to your HuggingFace account (`hf auth login`) OR pass a token with `--hub_token`.
+### FSDP Text Encoder 
+Flux.2 uses  Mistral Small 3.1 as text encoder which is quite large and can take up a lot of memory. To mitigate this, we can use the `--fsdp_text_encoder` flag to enable distributed computation of the prompt embeddings. 
+This way, it distributes the memory cost across multiple nodes.
 ### CPU Offloading 
 To offload parts of the model to CPU memory, you can use `--offload` flag. This will offload the vae and text encoder to CPU memory and only move them to GPU when needed.
 ### Latent Caching 
@@ -166,6 +169,26 @@ To better track our training experiments, we're using the following flags in the
 > [!NOTE]
 > If you want to train using long prompts with the T5 text encoder, you can use `--max_sequence_length` to set the token limit. The default is 77, but it can be increased to as high as 512. Note that this will use more resources and may slow down the training in some cases.
 
+### FSDP on the transformer
+By setting the accelerate configuration with FSDP, the transformer block will be wrapped automatically. E.g. set the configuration to:
+
+```shell
+distributed_type: FSDP
+fsdp_config:
+  fsdp_version: 2
+  fsdp_offload_params: false
+  fsdp_sharding_strategy: HYBRID_SHARD
+  fsdp_auto_wrap_policy: TRANSFOMER_BASED_WRAP
+  fsdp_transformer_layer_cls_to_wrap: Flux2TransformerBlock, Flux2SingleTransformerBlock
+  fsdp_forward_prefetch: true
+  fsdp_sync_module_states: false
+  fsdp_state_dict_type: FULL_STATE_DICT
+  fsdp_use_orig_params: false
+  fsdp_activation_checkpointing: true
+  fsdp_reshard_after_forward: true
+  fsdp_cpu_ram_efficient_loading: false
+```
+
 ## LoRA + DreamBooth
 
 [LoRA](https://huggingface.co/docs/peft/conceptual_guides/adapter#low-rank-adaptation-lora) is a popular parameter-efficient fine-tuning technique that allows you to achieve full-finetuning like performance but with a fraction of learnable parameters.

examples/dreambooth/train_dreambooth_lora_flux2.py

@@ -44,6 +44,7 @@ import shutil
 import warnings
 from contextlib import nullcontext
 from pathlib import Path
+from typing import Any
 
 import numpy as np
 import torch
@@ -75,13 +76,16 @@ from diffusers import (
 from diffusers.optimization import get_scheduler
 from diffusers.training_utils import (
     _collate_lora_metadata,
+    _to_cpu_contiguous,
     cast_training_params,
     compute_density_for_timestep_sampling,
     compute_loss_weighting_for_sd3,
     find_nearest_bucket,
     free_memory,
+    get_fsdp_kwargs_from_accelerator,
     offload_models,
     parse_buckets_string,
+    wrap_with_fsdp,
 )
 from diffusers.utils import (
     check_min_version,
@@ -93,6 +97,9 @@ from diffusers.utils.import_utils import is_torch_npu_available
 from diffusers.utils.torch_utils import is_compiled_module
 
 
+if getattr(torch, "distributed", None) is not None:
+    import torch.distributed as dist
+
 if is_wandb_available():
     import wandb
 
@@ -722,6 +729,7 @@ def parse_args(input_args=None):
     )
     parser.add_argument("--local_rank", type=int, default=-1, help="For distributed training: local_rank")
     parser.add_argument("--enable_npu_flash_attention", action="store_true", help="Enabla Flash Attention for NPU")
+    parser.add_argument("--fsdp_text_encoder", action="store_true", help="Use FSDP for text encoder")
 
     if input_args is not None:
         args = parser.parse_args(input_args)
@@ -1219,7 +1227,11 @@ def main(args):
         if args.bnb_quantization_config_path is not None
         else {"device": accelerator.device, "dtype": weight_dtype}
     )
-    transformer.to(**transformer_to_kwargs)
+
+    is_fsdp = accelerator.state.fsdp_plugin is not None
+    if not is_fsdp:
+        transformer.to(**transformer_to_kwargs)
+
     if args.do_fp8_training:
         convert_to_float8_training(
             transformer, module_filter_fn=module_filter_fn, config=Float8LinearConfig(pad_inner_dim=True)
@@ -1263,17 +1275,42 @@ def main(args):
 
     # create custom saving & loading hooks so that `accelerator.save_state(...)` serializes in a nice format
     def save_model_hook(models, weights, output_dir):
-        if accelerator.is_main_process:
-            transformer_lora_layers_to_save = None
-            modules_to_save = {}
-            for model in models:
-                if isinstance(model, type(unwrap_model(transformer))):
-                    transformer_lora_layers_to_save = get_peft_model_state_dict(model)
-                    modules_to_save["transformer"] = model
-                else:
-                    raise ValueError(f"unexpected save model: {model.__class__}")
+        transformer_cls = type(unwrap_model(transformer))
 
-                # make sure to pop weight so that corresponding model is not saved again
+        # 1) Validate and pick the transformer model
+        modules_to_save: dict[str, Any] = {}
+        transformer_model = None
+
+        for model in models:
+            if isinstance(unwrap_model(model), transformer_cls):
+                transformer_model = model
+                modules_to_save["transformer"] = model
+            else:
+                raise ValueError(f"unexpected save model: {model.__class__}")
+
+        if transformer_model is None:
+            raise ValueError("No transformer model found in 'models'")
+
+        # 2) Optionally gather FSDP state dict once
+        state_dict = accelerator.get_state_dict(model) if is_fsdp else None
+
+        # 3) Only main process materializes the LoRA state dict
+        transformer_lora_layers_to_save = None
+        if accelerator.is_main_process:
+            peft_kwargs = {}
+            if is_fsdp:
+                peft_kwargs["state_dict"] = state_dict
+
+            transformer_lora_layers_to_save = get_peft_model_state_dict(
+                unwrap_model(transformer_model) if is_fsdp else transformer_model,
+                **peft_kwargs,
+            )
+
+            if is_fsdp:
+                transformer_lora_layers_to_save = _to_cpu_contiguous(transformer_lora_layers_to_save)
+
+            # make sure to pop weight so that corresponding model is not saved again
+            if weights:
                 weights.pop()
 
             Flux2Pipeline.save_lora_weights(
@@ -1285,13 +1322,20 @@ def main(args):
     def load_model_hook(models, input_dir):
         transformer_ = None
 
-        while len(models) > 0:
-            model = models.pop()
+        if not is_fsdp:
+            while len(models) > 0:
+                model = models.pop()
 
-            if isinstance(model, type(unwrap_model(transformer))):
-                transformer_ = model
-            else:
-                raise ValueError(f"unexpected save model: {model.__class__}")
+                if isinstance(unwrap_model(model), type(unwrap_model(transformer))):
+                    transformer_ = unwrap_model(model)
+                else:
+                    raise ValueError(f"unexpected save model: {model.__class__}")
+        else:
+            transformer_ = Flux2Transformer2DModel.from_pretrained(
+                args.pretrained_model_name_or_path,
+                subfolder="transformer",
+            )
+            transformer_.add_adapter(transformer_lora_config)
 
         lora_state_dict = Flux2Pipeline.lora_state_dict(input_dir)
 
@@ -1507,6 +1551,21 @@ def main(args):
                     args.validation_prompt, text_encoding_pipeline
                 )
 
+    # Init FSDP for text encoder
+    if args.fsdp_text_encoder:
+        fsdp_kwargs = get_fsdp_kwargs_from_accelerator(accelerator)
+        text_encoder_fsdp = wrap_with_fsdp(
+            model=text_encoding_pipeline.text_encoder,
+            device=accelerator.device,
+            offload=args.offload,
+            limit_all_gathers=True,
+            use_orig_params=True,
+            fsdp_kwargs=fsdp_kwargs,
+        )
+
+        text_encoding_pipeline.text_encoder = text_encoder_fsdp
+        dist.barrier()
+
     # If custom instance prompts are NOT provided (i.e. the instance prompt is used for all images),
     # pack the statically computed variables appropriately here. This is so that we don't
     # have to pass them to the dataloader.
@@ -1536,6 +1595,8 @@ def main(args):
                 if train_dataset.custom_instance_prompts:
                     if args.remote_text_encoder:
                         prompt_embeds, text_ids = compute_remote_text_embeddings(batch["prompts"])
+                    elif args.fsdp_text_encoder:
+                        prompt_embeds, text_ids = compute_text_embeddings(batch["prompts"], text_encoding_pipeline)
                     else:
                         with offload_models(text_encoding_pipeline, device=accelerator.device, offload=args.offload):
                             prompt_embeds, text_ids = compute_text_embeddings(batch["prompts"], text_encoding_pipeline)
@@ -1777,7 +1838,7 @@ def main(args):
                 progress_bar.update(1)
                 global_step += 1
 
-                if accelerator.is_main_process:
+                if accelerator.is_main_process or is_fsdp:
                     if global_step % args.checkpointing_steps == 0:
                         # _before_ saving state, check if this save would set us over the `checkpoints_total_limit`
                         if args.checkpoints_total_limit is not None:
@@ -1836,15 +1897,41 @@ def main(args):
 
     # Save the lora layers
     accelerator.wait_for_everyone()
+
+    if is_fsdp:
+        transformer = unwrap_model(transformer)
+        state_dict = accelerator.get_state_dict(transformer)
     if accelerator.is_main_process:
         modules_to_save = {}
-        transformer = unwrap_model(transformer)
-        if args.bnb_quantization_config_path is None:
-            if args.upcast_before_saving:
-                transformer.to(torch.float32)
-            else:
-                transformer = transformer.to(weight_dtype)
-        transformer_lora_layers = get_peft_model_state_dict(transformer)
+        if is_fsdp:
+            if args.bnb_quantization_config_path is None:
+                if args.upcast_before_saving:
+                    state_dict = {
+                        k: v.to(torch.float32) if isinstance(v, torch.Tensor) else v for k, v in state_dict.items()
+                    }
+                else:
+                    state_dict = {
+                        k: v.to(weight_dtype) if isinstance(v, torch.Tensor) else v for k, v in state_dict.items()
+                    }
+
+            transformer_lora_layers = get_peft_model_state_dict(
+                transformer,
+                state_dict=state_dict,
+            )
+            transformer_lora_layers = {
+                k: v.detach().cpu().contiguous() if isinstance(v, torch.Tensor) else v
+                for k, v in transformer_lora_layers.items()
+            }
+
+        else:
+            transformer = unwrap_model(transformer)
+            if args.bnb_quantization_config_path is None:
+                if args.upcast_before_saving:
+                    transformer.to(torch.float32)
+                else:
+                    transformer = transformer.to(weight_dtype)
+            transformer_lora_layers = get_peft_model_state_dict(transformer)
+
         modules_to_save["transformer"] = transformer
 
         Flux2Pipeline.save_lora_weights(

examples/dreambooth/train_dreambooth_lora_flux2_img2img.py

@@ -43,6 +43,7 @@ import random
 import shutil
 from contextlib import nullcontext
 from pathlib import Path
+from typing import Any
 
 import numpy as np
 import torch
@@ -74,13 +75,16 @@ from diffusers.optimization import get_scheduler
 from diffusers.pipelines.flux2.image_processor import Flux2ImageProcessor
 from diffusers.training_utils import (
     _collate_lora_metadata,
+    _to_cpu_contiguous,
     cast_training_params,
     compute_density_for_timestep_sampling,
     compute_loss_weighting_for_sd3,
     find_nearest_bucket,
     free_memory,
+    get_fsdp_kwargs_from_accelerator,
     offload_models,
     parse_buckets_string,
+    wrap_with_fsdp,
 )
 from diffusers.utils import (
     check_min_version,
@@ -93,6 +97,9 @@ from diffusers.utils.import_utils import is_torch_npu_available
 from diffusers.utils.torch_utils import is_compiled_module
 
 
+if getattr(torch, "distributed", None) is not None:
+    import torch.distributed as dist
+
 if is_wandb_available():
     import wandb
 
@@ -691,6 +698,7 @@ def parse_args(input_args=None):
 
     parser.add_argument("--local_rank", type=int, default=-1, help="For distributed training: local_rank")
     parser.add_argument("--enable_npu_flash_attention", action="store_true", help="Enabla Flash Attention for NPU")
+    parser.add_argument("--fsdp_text_encoder", action="store_true", help="Use FSDP for text encoder")
 
     if input_args is not None:
         args = parser.parse_args(input_args)
@@ -1156,7 +1164,11 @@ def main(args):
         if args.bnb_quantization_config_path is not None
         else {"device": accelerator.device, "dtype": weight_dtype}
     )
-    transformer.to(**transformer_to_kwargs)
+
+    is_fsdp = accelerator.state.fsdp_plugin is not None
+    if not is_fsdp:
+        transformer.to(**transformer_to_kwargs)
+
     if args.do_fp8_training:
         convert_to_float8_training(
             transformer, module_filter_fn=module_filter_fn, config=Float8LinearConfig(pad_inner_dim=True)
@@ -1200,17 +1212,42 @@ def main(args):
 
     # create custom saving & loading hooks so that `accelerator.save_state(...)` serializes in a nice format
     def save_model_hook(models, weights, output_dir):
-        if accelerator.is_main_process:
-            transformer_lora_layers_to_save = None
-            modules_to_save = {}
-            for model in models:
-                if isinstance(model, type(unwrap_model(transformer))):
-                    transformer_lora_layers_to_save = get_peft_model_state_dict(model)
-                    modules_to_save["transformer"] = model
-                else:
-                    raise ValueError(f"unexpected save model: {model.__class__}")
+        transformer_cls = type(unwrap_model(transformer))
 
-                # make sure to pop weight so that corresponding model is not saved again
+        # 1) Validate and pick the transformer model
+        modules_to_save: dict[str, Any] = {}
+        transformer_model = None
+
+        for model in models:
+            if isinstance(unwrap_model(model), transformer_cls):
+                transformer_model = model
+                modules_to_save["transformer"] = model
+            else:
+                raise ValueError(f"unexpected save model: {model.__class__}")
+
+        if transformer_model is None:
+            raise ValueError("No transformer model found in 'models'")
+
+        # 2) Optionally gather FSDP state dict once
+        state_dict = accelerator.get_state_dict(model) if is_fsdp else None
+
+        # 3) Only main process materializes the LoRA state dict
+        transformer_lora_layers_to_save = None
+        if accelerator.is_main_process:
+            peft_kwargs = {}
+            if is_fsdp:
+                peft_kwargs["state_dict"] = state_dict
+
+            transformer_lora_layers_to_save = get_peft_model_state_dict(
+                unwrap_model(transformer_model) if is_fsdp else transformer_model,
+                **peft_kwargs,
+            )
+
+            if is_fsdp:
+                transformer_lora_layers_to_save = _to_cpu_contiguous(transformer_lora_layers_to_save)
+
+            # make sure to pop weight so that corresponding model is not saved again
+            if weights:
                 weights.pop()
 
             Flux2Pipeline.save_lora_weights(
@@ -1222,13 +1259,20 @@ def main(args):
     def load_model_hook(models, input_dir):
         transformer_ = None
 
-        while len(models) > 0:
-            model = models.pop()
+        if not is_fsdp:
+            while len(models) > 0:
+                model = models.pop()
 
-            if isinstance(model, type(unwrap_model(transformer))):
-                transformer_ = model
-            else:
-                raise ValueError(f"unexpected save model: {model.__class__}")
+                if isinstance(unwrap_model(model), type(unwrap_model(transformer))):
+                    transformer_ = unwrap_model(model)
+                else:
+                    raise ValueError(f"unexpected save model: {model.__class__}")
+        else:
+            transformer_ = Flux2Transformer2DModel.from_pretrained(
+                args.pretrained_model_name_or_path,
+                subfolder="transformer",
+            )
+            transformer_.add_adapter(transformer_lora_config)
 
         lora_state_dict = Flux2Pipeline.lora_state_dict(input_dir)
 
@@ -1430,6 +1474,21 @@ def main(args):
                     args.validation_prompt, text_encoding_pipeline
                 )
 
+    # Init FSDP for text encoder
+    if args.fsdp_text_encoder:
+        fsdp_kwargs = get_fsdp_kwargs_from_accelerator(accelerator)
+        text_encoder_fsdp = wrap_with_fsdp(
+            model=text_encoding_pipeline.text_encoder,
+            device=accelerator.device,
+            offload=args.offload,
+            limit_all_gathers=True,
+            use_orig_params=True,
+            fsdp_kwargs=fsdp_kwargs,
+        )
+
+        text_encoding_pipeline.text_encoder = text_encoder_fsdp
+        dist.barrier()
+
     # If custom instance prompts are NOT provided (i.e. the instance prompt is used for all images),
     # pack the statically computed variables appropriately here. This is so that we don't
     # have to pass them to the dataloader.
@@ -1461,6 +1520,8 @@ def main(args):
                 if train_dataset.custom_instance_prompts:
                     if args.remote_text_encoder:
                         prompt_embeds, text_ids = compute_remote_text_embeddings(batch["prompts"])
+                    elif args.fsdp_text_encoder:
+                        prompt_embeds, text_ids = compute_text_embeddings(batch["prompts"], text_encoding_pipeline)
                     else:
                         with offload_models(text_encoding_pipeline, device=accelerator.device, offload=args.offload):
                             prompt_embeds, text_ids = compute_text_embeddings(batch["prompts"], text_encoding_pipeline)
@@ -1700,7 +1761,7 @@ def main(args):
                 progress_bar.update(1)
                 global_step += 1
 
-                if accelerator.is_main_process:
+                if accelerator.is_main_process or is_fsdp:
                     if global_step % args.checkpointing_steps == 0:
                         # _before_ saving state, check if this save would set us over the `checkpoints_total_limit`
                         if args.checkpoints_total_limit is not None:
@@ -1759,15 +1820,41 @@ def main(args):
 
     # Save the lora layers
     accelerator.wait_for_everyone()
+
+    if is_fsdp:
+        transformer = unwrap_model(transformer)
+        state_dict = accelerator.get_state_dict(transformer)
     if accelerator.is_main_process:
         modules_to_save = {}
-        transformer = unwrap_model(transformer)
-        if args.bnb_quantization_config_path is None:
-            if args.upcast_before_saving:
-                transformer.to(torch.float32)
-            else:
-                transformer = transformer.to(weight_dtype)
-        transformer_lora_layers = get_peft_model_state_dict(transformer)
+        if is_fsdp:
+            if args.bnb_quantization_config_path is None:
+                if args.upcast_before_saving:
+                    state_dict = {
+                        k: v.to(torch.float32) if isinstance(v, torch.Tensor) else v for k, v in state_dict.items()
+                    }
+                else:
+                    state_dict = {
+                        k: v.to(weight_dtype) if isinstance(v, torch.Tensor) else v for k, v in state_dict.items()
+                    }
+
+            transformer_lora_layers = get_peft_model_state_dict(
+                transformer,
+                state_dict=state_dict,
+            )
+            transformer_lora_layers = {
+                k: v.detach().cpu().contiguous() if isinstance(v, torch.Tensor) else v
+                for k, v in transformer_lora_layers.items()
+            }
+
+        else:
+            transformer = unwrap_model(transformer)
+            if args.bnb_quantization_config_path is None:
+                if args.upcast_before_saving:
+                    transformer.to(torch.float32)
+                else:
+                    transformer = transformer.to(weight_dtype)
+            transformer_lora_layers = get_peft_model_state_dict(transformer)
+
         modules_to_save["transformer"] = transformer
 
         Flux2Pipeline.save_lora_weights(

scripts/convert_cosmos_to_diffusers.py

@@ -29,13 +29,52 @@ hf download nvidia/Cosmos-Predict2.5-2B
 
 Convert checkpoint
 ```bash
+# pre-trained
 transformer_ckpt_path=~/.cache/huggingface/hub/models--nvidia--Cosmos-Predict2.5-2B/snapshots/865baf084d4c9e850eac59a021277d5a9b9e8b63/base/pre-trained/d20b7120-df3e-4911-919d-db6e08bad31c_ema_bf16.pt
 
 python scripts/convert_cosmos_to_diffusers.py \
     --transformer_type Cosmos-2.5-Predict-Base-2B \
     --transformer_ckpt_path $transformer_ckpt_path \
     --vae_type wan2.1 \
-    --output_path converted/cosmos-p2.5-base-2b \
+    --output_path converted/2b/d20b7120-df3e-4911-919d-db6e08bad31c \
+    --save_pipeline
+
+# post-trained
+transformer_ckpt_path=~/.cache/huggingface/hub/models--nvidia--Cosmos-Predict2.5-2B/snapshots/865baf084d4c9e850eac59a021277d5a9b9e8b63/base/post-trained/81edfebe-bd6a-4039-8c1d-737df1a790bf_ema_bf16.pt
+
+python scripts/convert_cosmos_to_diffusers.py \
+    --transformer_type Cosmos-2.5-Predict-Base-2B \
+    --transformer_ckpt_path $transformer_ckpt_path \
+    --vae_type wan2.1 \
+    --output_path converted/2b/81edfebe-bd6a-4039-8c1d-737df1a790bf \
+    --save_pipeline
+```
+
+## 14B
+
+```bash
+hf download nvidia/Cosmos-Predict2.5-14B
+```
+
+```bash
+# pre-trained
+transformer_ckpt_path=~/.cache/huggingface/hub/models--nvidia--Cosmos-Predict2.5-14B/snapshots/71ebf3e8af30ecfe440bf0481115975fcc052b46/base/pre-trained/54937b8c-29de-4f04-862c-e67b04ec41e8_ema_bf16.pt
+
+python scripts/convert_cosmos_to_diffusers.py \
+    --transformer_type Cosmos-2.5-Predict-Base-14B \
+    --transformer_ckpt_path $transformer_ckpt_path \
+    --vae_type wan2.1 \
+    --output_path converted/14b/54937b8c-29de-4f04-862c-e67b04ec41e8/ \
+    --save_pipeline
+
+# post-trained
+transformer_ckpt_path=~/.cache/huggingface/hub/models--nvidia--Cosmos-Predict2.5-14B/snapshots/71ebf3e8af30ecfe440bf0481115975fcc052b46/base/post-trained/e21d2a49-4747-44c8-ba44-9f6f9243715f_ema_bf16.pt
+
+python scripts/convert_cosmos_to_diffusers.py \
+    --transformer_type Cosmos-2.5-Predict-Base-14B \
+    --transformer_ckpt_path $transformer_ckpt_path \
+    --vae_type wan2.1 \
+    --output_path converted/14b/e21d2a49-4747-44c8-ba44-9f6f9243715f/ \
     --save_pipeline
 ```
 
@@ -298,6 +337,25 @@ TRANSFORMER_CONFIGS = {
         "crossattn_proj_in_channels": 100352,
         "encoder_hidden_states_channels": 1024,
     },
+    "Cosmos-2.5-Predict-Base-14B": {
+        "in_channels": 16 + 1,
+        "out_channels": 16,
+        "num_attention_heads": 40,
+        "attention_head_dim": 128,
+        "num_layers": 36,
+        "mlp_ratio": 4.0,
+        "text_embed_dim": 1024,
+        "adaln_lora_dim": 256,
+        "max_size": (128, 240, 240),
+        "patch_size": (1, 2, 2),
+        "rope_scale": (1.0, 3.0, 3.0),
+        "concat_padding_mask": True,
+        # NOTE: source config has pos_emb_learnable: 'True' - but params are missing
+        "extra_pos_embed_type": None,
+        "use_crossattn_projection": True,
+        "crossattn_proj_in_channels": 100352,
+        "encoder_hidden_states_channels": 1024,
+    },
 }
 
 VAE_KEYS_RENAME_DICT = {

setup.py

@@ -274,7 +274,7 @@ version_range_max = max(sys.version_info[1], 10) + 1
 
 setup(
     name="diffusers",
-    version="0.36.0.dev0",  # expected format is one of x.y.z.dev0, or x.y.z.rc1 or x.y.z (no to dashes, yes to dots)
+    version="0.37.0.dev0",  # expected format is one of x.y.z.dev0, or x.y.z.rc1 or x.y.z (no to dashes, yes to dots)
     description="State-of-the-art diffusion in PyTorch and JAX.",
     long_description=open("README.md", "r", encoding="utf-8").read(),
     long_description_content_type="text/markdown",

src/diffusers/__init__.py

@@ -353,6 +353,7 @@ else:
             "KDPM2AncestralDiscreteScheduler",
             "KDPM2DiscreteScheduler",
             "LCMScheduler",
+            "LTXEulerAncestralRFScheduler",
             "PNDMScheduler",
             "RePaintScheduler",
             "SASolverScheduler",
@@ -538,6 +539,7 @@ else:
             "LongCatImageEditPipeline",
             "LongCatImagePipeline",
             "LTXConditionPipeline",
+            "LTXI2VLongMultiPromptPipeline",
             "LTXImageToVideoPipeline",
             "LTXLatentUpsamplePipeline",
             "LTXPipeline",
@@ -675,6 +677,7 @@ else:
             "ZImageControlNetInpaintPipeline",
             "ZImageControlNetPipeline",
             "ZImageImg2ImgPipeline",
+            "ZImageOmniPipeline",
             "ZImagePipeline",
         ]
     )
@@ -1087,6 +1090,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             KDPM2AncestralDiscreteScheduler,
             KDPM2DiscreteScheduler,
             LCMScheduler,
+            LTXEulerAncestralRFScheduler,
             PNDMScheduler,
             RePaintScheduler,
             SASolverScheduler,
@@ -1251,6 +1255,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             LongCatImageEditPipeline,
             LongCatImagePipeline,
             LTXConditionPipeline,
+            LTXI2VLongMultiPromptPipeline,
             LTXImageToVideoPipeline,
             LTXLatentUpsamplePipeline,
             LTXPipeline,
@@ -1386,6 +1391,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             ZImageControlNetInpaintPipeline,
             ZImageControlNetPipeline,
             ZImageImg2ImgPipeline,
+            ZImageOmniPipeline,
             ZImagePipeline,
         )
 

src/diffusers/loaders/single_file_model.py

@@ -162,7 +162,7 @@ SINGLE_FILE_LOADABLE_CLASSES = {
         "default_subfolder": "transformer",
     },
     "QwenImageTransformer2DModel": {
-        "checkpoint_mapping_fn": lambda x: x,
+        "checkpoint_mapping_fn": lambda checkpoint, **kwargs: checkpoint,
         "default_subfolder": "transformer",
     },
     "Flux2Transformer2DModel": {

src/diffusers/loaders/single_file_utils.py

@@ -120,7 +120,10 @@ CHECKPOINT_KEY_NAMES = {
     "hunyuan-video": "txt_in.individual_token_refiner.blocks.0.adaLN_modulation.1.bias",
     "instruct-pix2pix": "model.diffusion_model.input_blocks.0.0.weight",
     "lumina2": ["model.diffusion_model.cap_embedder.0.weight", "cap_embedder.0.weight"],
-    "z-image-turbo": "cap_embedder.0.weight",
+    "z-image-turbo": [
+        "model.diffusion_model.layers.0.adaLN_modulation.0.weight",
+        "layers.0.adaLN_modulation.0.weight",
+    ],
     "z-image-turbo-controlnet": "control_all_x_embedder.2-1.weight",
     "z-image-turbo-controlnet-2.x": "control_layers.14.adaLN_modulation.0.weight",
     "sana": [
@@ -223,7 +226,8 @@ DIFFUSERS_DEFAULT_PIPELINE_PATHS = {
     "cosmos-2.0-v2w-14B": {"pretrained_model_name_or_path": "nvidia/Cosmos-Predict2-14B-Video2World"},
     "z-image-turbo": {"pretrained_model_name_or_path": "Tongyi-MAI/Z-Image-Turbo"},
     "z-image-turbo-controlnet": {"pretrained_model_name_or_path": "hlky/Z-Image-Turbo-Fun-Controlnet-Union"},
-    "z-image-turbo-controlnet-2.x": {"pretrained_model_name_or_path": "hlky/Z-Image-Turbo-Fun-Controlnet-Union-2.1"},
+    "z-image-turbo-controlnet-2.0": {"pretrained_model_name_or_path": "hlky/Z-Image-Turbo-Fun-Controlnet-Union-2.0"},
+    "z-image-turbo-controlnet-2.1": {"pretrained_model_name_or_path": "hlky/Z-Image-Turbo-Fun-Controlnet-Union-2.1"},
 }
 
 # Use to configure model sample size when original config is provided
@@ -727,10 +731,7 @@ def infer_diffusers_model_type(checkpoint):
     ):
         model_type = "instruct-pix2pix"
 
-    elif (
-        CHECKPOINT_KEY_NAMES["z-image-turbo"] in checkpoint
-        and checkpoint[CHECKPOINT_KEY_NAMES["z-image-turbo"]].shape[0] == 2560
-    ):
+    elif any(key in checkpoint for key in CHECKPOINT_KEY_NAMES["z-image-turbo"]):
         model_type = "z-image-turbo"
 
     elif any(key in checkpoint for key in CHECKPOINT_KEY_NAMES["lumina2"]):
@@ -784,7 +785,13 @@ def infer_diffusers_model_type(checkpoint):
             raise ValueError(f"Unexpected x_embedder shape: {x_embedder_shape} when loading Cosmos 2.0 model.")
 
     elif CHECKPOINT_KEY_NAMES["z-image-turbo-controlnet-2.x"] in checkpoint:
-        model_type = "z-image-turbo-controlnet-2.x"
+        before_proj_weight = checkpoint.get("control_noise_refiner.0.before_proj.weight", None)
+        if before_proj_weight is None:
+            model_type = "z-image-turbo-controlnet-2.0"
+        elif before_proj_weight is not None and torch.all(before_proj_weight == 0.0):
+            model_type = "z-image-turbo-controlnet-2.0"
+        else:
+            model_type = "z-image-turbo-controlnet-2.1"
 
     elif CHECKPOINT_KEY_NAMES["z-image-turbo-controlnet"] in checkpoint:
         model_type = "z-image-turbo-controlnet"
@@ -3852,6 +3859,7 @@ def convert_z_image_transformer_checkpoint_to_diffusers(checkpoint, **kwargs):
         ".attention.k_norm.weight": ".attention.norm_k.weight",
         ".attention.q_norm.weight": ".attention.norm_q.weight",
         ".attention.out.weight": ".attention.to_out.0.weight",
+        "model.diffusion_model.": "",
     }
 
     def convert_z_image_fused_attention(key: str, state_dict: dict[str, object]) -> None:
@@ -3886,6 +3894,9 @@ def convert_z_image_transformer_checkpoint_to_diffusers(checkpoint, **kwargs):
 
         update_state_dict(converted_state_dict, key, new_key)
 
+    if "norm_final.weight" in converted_state_dict.keys():
+        _ = converted_state_dict.pop("norm_final.weight")
+
     # Handle any special logic which can't be expressed by a simple 1:1 remapping with the handlers in
     # special_keys_remap
     for key in list(converted_state_dict.keys()):

src/diffusers/models/attention_dispatch.py

@@ -1420,6 +1420,7 @@ def _flash_attention(
     query: torch.Tensor,
     key: torch.Tensor,
     value: torch.Tensor,
+    attn_mask: Optional[torch.Tensor] = None,
     dropout_p: float = 0.0,
     is_causal: bool = False,
     scale: Optional[float] = None,
@@ -1427,6 +1428,9 @@ def _flash_attention(
     _parallel_config: Optional["ParallelConfig"] = None,
 ) -> torch.Tensor:
     lse = None
+    if attn_mask is not None:
+        raise ValueError("`attn_mask` is not supported for flash-attn 2.")
+
     if _parallel_config is None:
         out = flash_attn_func(
             q=query,
@@ -1469,6 +1473,7 @@ def _flash_attention_hub(
     query: torch.Tensor,
     key: torch.Tensor,
     value: torch.Tensor,
+    attn_mask: Optional[torch.Tensor] = None,
     dropout_p: float = 0.0,
     is_causal: bool = False,
     scale: Optional[float] = None,
@@ -1476,6 +1481,9 @@ def _flash_attention_hub(
     _parallel_config: Optional["ParallelConfig"] = None,
 ) -> torch.Tensor:
     lse = None
+    if attn_mask is not None:
+        raise ValueError("`attn_mask` is not supported for flash-attn 2.")
+
     func = _HUB_KERNELS_REGISTRY[AttentionBackendName.FLASH_HUB].kernel_fn
     out = func(
         q=query,
@@ -1612,11 +1620,15 @@ def _flash_attention_3(
     query: torch.Tensor,
     key: torch.Tensor,
     value: torch.Tensor,
+    attn_mask: Optional[torch.Tensor] = None,
     scale: Optional[float] = None,
     is_causal: bool = False,
     return_lse: bool = False,
     _parallel_config: Optional["ParallelConfig"] = None,
 ) -> torch.Tensor:
+    if attn_mask is not None:
+        raise ValueError("`attn_mask` is not supported for flash-attn 3.")
+
     out, lse = _wrapped_flash_attn_3(
         q=query,
         k=key,
@@ -1636,6 +1648,7 @@ def _flash_attention_3_hub(
     query: torch.Tensor,
     key: torch.Tensor,
     value: torch.Tensor,
+    attn_mask: Optional[torch.Tensor] = None,
     scale: Optional[float] = None,
     is_causal: bool = False,
     window_size: Tuple[int, int] = (-1, -1),
@@ -1646,6 +1659,8 @@ def _flash_attention_3_hub(
 ) -> torch.Tensor:
     if _parallel_config:
         raise NotImplementedError(f"{AttentionBackendName._FLASH_3_HUB.value} is not implemented for parallelism yet.")
+    if attn_mask is not None:
+        raise ValueError("`attn_mask` is not supported for flash-attn 3.")
 
     func = _HUB_KERNELS_REGISTRY[AttentionBackendName._FLASH_3_HUB].kernel_fn
     out = func(
@@ -1785,12 +1800,16 @@ def _aiter_flash_attention(
     query: torch.Tensor,
     key: torch.Tensor,
     value: torch.Tensor,
+    attn_mask: Optional[torch.Tensor] = None,
     dropout_p: float = 0.0,
     is_causal: bool = False,
     scale: Optional[float] = None,
     return_lse: bool = False,
     _parallel_config: Optional["ParallelConfig"] = None,
 ) -> torch.Tensor:
+    if attn_mask is not None:
+        raise ValueError("`attn_mask` is not supported for aiter attention")
+
     if not return_lse and torch.is_grad_enabled():
         # aiter requires return_lse=True by assertion when gradients are enabled.
         out, lse, *_ = aiter_flash_attn_func(
@@ -2028,6 +2047,7 @@ def _native_flash_attention(
     query: torch.Tensor,
     key: torch.Tensor,
     value: torch.Tensor,
+    attn_mask: Optional[torch.Tensor] = None,
     dropout_p: float = 0.0,
     is_causal: bool = False,
     scale: Optional[float] = None,
@@ -2035,6 +2055,9 @@ def _native_flash_attention(
     return_lse: bool = False,
     _parallel_config: Optional["ParallelConfig"] = None,
 ) -> torch.Tensor:
+    if attn_mask is not None:
+        raise ValueError("`attn_mask` is not supported for aiter attention")
+
     lse = None
     if _parallel_config is None and not return_lse:
         query, key, value = (x.permute(0, 2, 1, 3) for x in (query, key, value))
@@ -2113,11 +2136,14 @@ def _native_npu_attention(
     query: torch.Tensor,
     key: torch.Tensor,
     value: torch.Tensor,
+    attn_mask: Optional[torch.Tensor] = None,
     dropout_p: float = 0.0,
     scale: Optional[float] = None,
     return_lse: bool = False,
     _parallel_config: Optional["ParallelConfig"] = None,
 ) -> torch.Tensor:
+    if attn_mask is not None:
+        raise ValueError("`attn_mask` is not supported for NPU attention")
     if return_lse:
         raise ValueError("NPU attention backend does not support setting `return_lse=True`.")
     query, key, value = (x.transpose(1, 2).contiguous() for x in (query, key, value))
@@ -2148,10 +2174,13 @@ def _native_xla_attention(
     query: torch.Tensor,
     key: torch.Tensor,
     value: torch.Tensor,
+    attn_mask: Optional[torch.Tensor] = None,
     is_causal: bool = False,
     return_lse: bool = False,
     _parallel_config: Optional["ParallelConfig"] = None,
 ) -> torch.Tensor:
+    if attn_mask is not None:
+        raise ValueError("`attn_mask` is not supported for XLA attention")
     if return_lse:
         raise ValueError("XLA attention backend does not support setting `return_lse=True`.")
     query, key, value = (x.permute(0, 2, 1, 3) for x in (query, key, value))
@@ -2175,11 +2204,14 @@ def _sage_attention(
     query: torch.Tensor,
     key: torch.Tensor,
     value: torch.Tensor,
+    attn_mask: Optional[torch.Tensor] = None,
     is_causal: bool = False,
     scale: Optional[float] = None,
     return_lse: bool = False,
     _parallel_config: Optional["ParallelConfig"] = None,
 ) -> torch.Tensor:
+    if attn_mask is not None:
+        raise ValueError("`attn_mask` is not supported for sage attention")
     lse = None
     if _parallel_config is None:
         out = sageattn(
@@ -2223,11 +2255,14 @@ def _sage_attention_hub(
     query: torch.Tensor,
     key: torch.Tensor,
     value: torch.Tensor,
+    attn_mask: Optional[torch.Tensor] = None,
     is_causal: bool = False,
     scale: Optional[float] = None,
     return_lse: bool = False,
     _parallel_config: Optional["ParallelConfig"] = None,
 ) -> torch.Tensor:
+    if attn_mask is not None:
+        raise ValueError("`attn_mask` is not supported for sage attention")
     lse = None
     func = _HUB_KERNELS_REGISTRY[AttentionBackendName.SAGE_HUB].kernel_fn
     if _parallel_config is None:
@@ -2309,11 +2344,14 @@ def _sage_qk_int8_pv_fp8_cuda_attention(
     query: torch.Tensor,
     key: torch.Tensor,
     value: torch.Tensor,
+    attn_mask: Optional[torch.Tensor] = None,
     is_causal: bool = False,
     scale: Optional[float] = None,
     return_lse: bool = False,
     _parallel_config: Optional["ParallelConfig"] = None,
 ) -> torch.Tensor:
+    if attn_mask is not None:
+        raise ValueError("`attn_mask` is not supported for sage attention")
     return sageattn_qk_int8_pv_fp8_cuda(
         q=query,
         k=key,
@@ -2333,11 +2371,14 @@ def _sage_qk_int8_pv_fp8_cuda_sm90_attention(
     query: torch.Tensor,
     key: torch.Tensor,
     value: torch.Tensor,
+    attn_mask: Optional[torch.Tensor] = None,
     is_causal: bool = False,
     scale: Optional[float] = None,
     return_lse: bool = False,
     _parallel_config: Optional["ParallelConfig"] = None,
 ) -> torch.Tensor:
+    if attn_mask is not None:
+        raise ValueError("`attn_mask` is not supported for sage attention")
     return sageattn_qk_int8_pv_fp8_cuda_sm90(
         q=query,
         k=key,
@@ -2357,11 +2398,14 @@ def _sage_qk_int8_pv_fp16_cuda_attention(
     query: torch.Tensor,
     key: torch.Tensor,
     value: torch.Tensor,
+    attn_mask: Optional[torch.Tensor] = None,
     is_causal: bool = False,
     scale: Optional[float] = None,
     return_lse: bool = False,
     _parallel_config: Optional["ParallelConfig"] = None,
 ) -> torch.Tensor:
+    if attn_mask is not None:
+        raise ValueError("`attn_mask` is not supported for sage attention")
     return sageattn_qk_int8_pv_fp16_cuda(
         q=query,
         k=key,
@@ -2381,11 +2425,14 @@ def _sage_qk_int8_pv_fp16_triton_attention(
     query: torch.Tensor,
     key: torch.Tensor,
     value: torch.Tensor,
+    attn_mask: Optional[torch.Tensor] = None,
     is_causal: bool = False,
     scale: Optional[float] = None,
     return_lse: bool = False,
     _parallel_config: Optional["ParallelConfig"] = None,
 ) -> torch.Tensor:
+    if attn_mask is not None:
+        raise ValueError("`attn_mask` is not supported for sage attention")
     return sageattn_qk_int8_pv_fp16_triton(
         q=query,
         k=key,

src/diffusers/models/autoencoders/autoencoder_kl_hunyuanimage.py

@@ -27,7 +27,7 @@ from ...utils.accelerate_utils import apply_forward_hook
 from ..activations import get_activation
 from ..modeling_outputs import AutoencoderKLOutput
 from ..modeling_utils import ModelMixin
-from .vae import DecoderOutput, DiagonalGaussianDistribution
+from .vae import AutoencoderMixin, DecoderOutput, DiagonalGaussianDistribution
 
 
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
@@ -410,7 +410,7 @@ class HunyuanImageDecoder2D(nn.Module):
         return h
 
 
-class AutoencoderKLHunyuanImage(ModelMixin, ConfigMixin, FromOriginalModelMixin):
+class AutoencoderKLHunyuanImage(ModelMixin, AutoencoderMixin, ConfigMixin, FromOriginalModelMixin):
     r"""
     A VAE model for 2D images with spatial tiling support.
 
@@ -486,27 +486,6 @@ class AutoencoderKLHunyuanImage(ModelMixin, ConfigMixin, FromOriginalModelMixin)
         self.tile_overlap_factor = tile_overlap_factor or self.tile_overlap_factor
         self.tile_latent_min_size = self.tile_sample_min_size // self.config.spatial_compression_ratio
 
-    def disable_tiling(self) -> None:
-        r"""
-        Disable tiled VAE decoding. If `enable_tiling` was previously enabled, this method will go back to computing
-        decoding in one step.
-        """
-        self.use_tiling = False
-
-    def enable_slicing(self) -> None:
-        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.use_slicing = True
-
-    def disable_slicing(self) -> None:
-        r"""
-        Disable sliced VAE decoding. If `enable_slicing` was previously enabled, this method will go back to computing
-        decoding in one step.
-        """
-        self.use_slicing = False
-
     def _encode(self, x: torch.Tensor):
 
         batch_size, num_channels, height, width = x.shape

src/diffusers/models/autoencoders/autoencoder_kl_hunyuanimage_refiner.py

@@ -26,7 +26,7 @@ from ...utils.accelerate_utils import apply_forward_hook
 from ..activations import get_activation
 from ..modeling_outputs import AutoencoderKLOutput
 from ..modeling_utils import ModelMixin
-from .vae import DecoderOutput, DiagonalGaussianDistribution
+from .vae import AutoencoderMixin, DecoderOutput, DiagonalGaussianDistribution
 
 
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
@@ -584,7 +584,7 @@ class HunyuanImageRefinerDecoder3D(nn.Module):
         return hidden_states
 
 
-class AutoencoderKLHunyuanImageRefiner(ModelMixin, ConfigMixin):
+class AutoencoderKLHunyuanImageRefiner(ModelMixin, AutoencoderMixin, ConfigMixin):
     r"""
     A VAE model with KL loss for encoding videos into latents and decoding latent representations into videos. Used for
     HunyuanImage-2.1 Refiner.
@@ -685,27 +685,6 @@ class AutoencoderKLHunyuanImageRefiner(ModelMixin, ConfigMixin):
         self.tile_sample_stride_width = tile_sample_stride_width or self.tile_sample_stride_width
         self.tile_overlap_factor = tile_overlap_factor or self.tile_overlap_factor
 
-    def disable_tiling(self) -> None:
-        r"""
-        Disable tiled VAE decoding. If `enable_tiling` was previously enabled, this method will go back to computing
-        decoding in one step.
-        """
-        self.use_tiling = False
-
-    def enable_slicing(self) -> None:
-        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.use_slicing = True
-
-    def disable_slicing(self) -> None:
-        r"""
-        Disable sliced VAE decoding. If `enable_slicing` was previously enabled, this method will go back to computing
-        decoding in one step.
-        """
-        self.use_slicing = False
-
     def _encode(self, x: torch.Tensor) -> torch.Tensor:
         _, _, _, height, width = x.shape
 

src/diffusers/models/autoencoders/autoencoder_kl_hunyuanvideo15.py

@@ -26,7 +26,7 @@ from ...utils.accelerate_utils import apply_forward_hook
 from ..activations import get_activation
 from ..modeling_outputs import AutoencoderKLOutput
 from ..modeling_utils import ModelMixin
-from .vae import DecoderOutput, DiagonalGaussianDistribution
+from .vae import AutoencoderMixin, DecoderOutput, DiagonalGaussianDistribution
 
 
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
@@ -625,7 +625,7 @@ class HunyuanVideo15Decoder3D(nn.Module):
         return hidden_states
 
 
-class AutoencoderKLHunyuanVideo15(ModelMixin, ConfigMixin):
+class AutoencoderKLHunyuanVideo15(ModelMixin, AutoencoderMixin, ConfigMixin):
     r"""
     A VAE model with KL loss for encoding videos into latents and decoding latent representations into videos. Used for
     HunyuanVideo-1.5.
@@ -723,27 +723,6 @@ class AutoencoderKLHunyuanVideo15(ModelMixin, ConfigMixin):
         self.tile_latent_min_width = tile_latent_min_width or self.tile_latent_min_width
         self.tile_overlap_factor = tile_overlap_factor or self.tile_overlap_factor
 
-    def disable_tiling(self) -> None:
-        r"""
-        Disable tiled VAE decoding. If `enable_tiling` was previously enabled, this method will go back to computing
-        decoding in one step.
-        """
-        self.use_tiling = False
-
-    def enable_slicing(self) -> None:
-        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.use_slicing = True
-
-    def disable_slicing(self) -> None:
-        r"""
-        Disable sliced VAE decoding. If `enable_slicing` was previously enabled, this method will go back to computing
-        decoding in one step.
-        """
-        self.use_slicing = False
-
     def _encode(self, x: torch.Tensor) -> torch.Tensor:
         _, _, _, height, width = x.shape
 

src/diffusers/models/controlnets/controlnet_z_image.py

@@ -13,7 +13,7 @@
 # limitations under the License.
 
 import math
-from typing import List, Literal, Optional
+from typing import List, Literal, Optional, Tuple
 
 import torch
 import torch.nn as nn
@@ -170,6 +170,21 @@ class FeedForward(nn.Module):
         return self.w2(self._forward_silu_gating(self.w1(x), self.w3(x)))
 
 
+# Copied from diffusers.models.transformers.transformer_z_image.select_per_token
+def select_per_token(
+    value_noisy: torch.Tensor,
+    value_clean: torch.Tensor,
+    noise_mask: torch.Tensor,
+    seq_len: int,
+) -> torch.Tensor:
+    noise_mask_expanded = noise_mask.unsqueeze(-1)  # (batch, seq_len, 1)
+    return torch.where(
+        noise_mask_expanded == 1,
+        value_noisy.unsqueeze(1).expand(-1, seq_len, -1),
+        value_clean.unsqueeze(1).expand(-1, seq_len, -1),
+    )
+
+
 @maybe_allow_in_graph
 # Copied from diffusers.models.transformers.transformer_z_image.ZImageTransformerBlock
 class ZImageTransformerBlock(nn.Module):
@@ -220,12 +235,37 @@ class ZImageTransformerBlock(nn.Module):
         attn_mask: torch.Tensor,
         freqs_cis: torch.Tensor,
         adaln_input: Optional[torch.Tensor] = None,
+        noise_mask: Optional[torch.Tensor] = None,
+        adaln_noisy: Optional[torch.Tensor] = None,
+        adaln_clean: Optional[torch.Tensor] = None,
     ):
         if self.modulation:
-            assert adaln_input is not None
-            scale_msa, gate_msa, scale_mlp, gate_mlp = self.adaLN_modulation(adaln_input).unsqueeze(1).chunk(4, dim=2)
-            gate_msa, gate_mlp = gate_msa.tanh(), gate_mlp.tanh()
-            scale_msa, scale_mlp = 1.0 + scale_msa, 1.0 + scale_mlp
+            seq_len = x.shape[1]
+
+            if noise_mask is not None:
+                # Per-token modulation: different modulation for noisy/clean tokens
+                mod_noisy = self.adaLN_modulation(adaln_noisy)
+                mod_clean = self.adaLN_modulation(adaln_clean)
+
+                scale_msa_noisy, gate_msa_noisy, scale_mlp_noisy, gate_mlp_noisy = mod_noisy.chunk(4, dim=1)
+                scale_msa_clean, gate_msa_clean, scale_mlp_clean, gate_mlp_clean = mod_clean.chunk(4, dim=1)
+
+                gate_msa_noisy, gate_mlp_noisy = gate_msa_noisy.tanh(), gate_mlp_noisy.tanh()
+                gate_msa_clean, gate_mlp_clean = gate_msa_clean.tanh(), gate_mlp_clean.tanh()
+
+                scale_msa_noisy, scale_mlp_noisy = 1.0 + scale_msa_noisy, 1.0 + scale_mlp_noisy
+                scale_msa_clean, scale_mlp_clean = 1.0 + scale_msa_clean, 1.0 + scale_mlp_clean
+
+                scale_msa = select_per_token(scale_msa_noisy, scale_msa_clean, noise_mask, seq_len)
+                scale_mlp = select_per_token(scale_mlp_noisy, scale_mlp_clean, noise_mask, seq_len)
+                gate_msa = select_per_token(gate_msa_noisy, gate_msa_clean, noise_mask, seq_len)
+                gate_mlp = select_per_token(gate_mlp_noisy, gate_mlp_clean, noise_mask, seq_len)
+            else:
+                # Global modulation: same modulation for all tokens (avoid double select)
+                mod = self.adaLN_modulation(adaln_input)
+                scale_msa, gate_msa, scale_mlp, gate_mlp = mod.unsqueeze(1).chunk(4, dim=2)
+                gate_msa, gate_mlp = gate_msa.tanh(), gate_mlp.tanh()
+                scale_msa, scale_mlp = 1.0 + scale_msa, 1.0 + scale_mlp
 
             # Attention block
             attn_out = self.attention(
@@ -493,112 +533,93 @@ class ZImageControlNetModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOrigi
     def create_coordinate_grid(size, start=None, device=None):
         if start is None:
             start = (0 for _ in size)
-
         axes = [torch.arange(x0, x0 + span, dtype=torch.int32, device=device) for x0, span in zip(start, size)]
         grids = torch.meshgrid(axes, indexing="ij")
         return torch.stack(grids, dim=-1)
 
+    # Copied from diffusers.models.transformers.transformer_z_image.ZImageTransformer2DModel._patchify_image
+    def _patchify_image(self, image: torch.Tensor, patch_size: int, f_patch_size: int):
+        """Patchify a single image tensor: (C, F, H, W) -> (num_patches, patch_dim)."""
+        pH, pW, pF = patch_size, patch_size, f_patch_size
+        C, F, H, W = image.size()
+        F_tokens, H_tokens, W_tokens = F // pF, H // pH, W // pW
+        image = image.view(C, F_tokens, pF, H_tokens, pH, W_tokens, pW)
+        image = image.permute(1, 3, 5, 2, 4, 6, 0).reshape(F_tokens * H_tokens * W_tokens, pF * pH * pW * C)
+        return image, (F, H, W), (F_tokens, H_tokens, W_tokens)
+
+    # Copied from diffusers.models.transformers.transformer_z_image.ZImageTransformer2DModel._pad_with_ids
+    def _pad_with_ids(
+        self,
+        feat: torch.Tensor,
+        pos_grid_size: Tuple,
+        pos_start: Tuple,
+        device: torch.device,
+        noise_mask_val: Optional[int] = None,
+    ):
+        """Pad feature to SEQ_MULTI_OF, create position IDs and pad mask."""
+        ori_len = len(feat)
+        pad_len = (-ori_len) % SEQ_MULTI_OF
+        total_len = ori_len + pad_len
+
+        # Pos IDs
+        ori_pos_ids = self.create_coordinate_grid(size=pos_grid_size, start=pos_start, device=device).flatten(0, 2)
+        if pad_len > 0:
+            pad_pos_ids = (
+                self.create_coordinate_grid(size=(1, 1, 1), start=(0, 0, 0), device=device)
+                .flatten(0, 2)
+                .repeat(pad_len, 1)
+            )
+            pos_ids = torch.cat([ori_pos_ids, pad_pos_ids], dim=0)
+            padded_feat = torch.cat([feat, feat[-1:].repeat(pad_len, 1)], dim=0)
+            pad_mask = torch.cat(
+                [
+                    torch.zeros(ori_len, dtype=torch.bool, device=device),
+                    torch.ones(pad_len, dtype=torch.bool, device=device),
+                ]
+            )
+        else:
+            pos_ids = ori_pos_ids
+            padded_feat = feat
+            pad_mask = torch.zeros(ori_len, dtype=torch.bool, device=device)
+
+        noise_mask = [noise_mask_val] * total_len if noise_mask_val is not None else None  # token level
+        return padded_feat, pos_ids, pad_mask, total_len, noise_mask
+
     # Copied from diffusers.models.transformers.transformer_z_image.ZImageTransformer2DModel.patchify_and_embed
     def patchify_and_embed(
-        self,
-        all_image: List[torch.Tensor],
-        all_cap_feats: List[torch.Tensor],
-        patch_size: int,
-        f_patch_size: int,
+        self, all_image: List[torch.Tensor], all_cap_feats: List[torch.Tensor], patch_size: int, f_patch_size: int
     ):
-        pH = pW = patch_size
-        pF = f_patch_size
+        """Patchify for basic mode: single image per batch item."""
         device = all_image[0].device
+        all_img_out, all_img_size, all_img_pos_ids, all_img_pad_mask = [], [], [], []
+        all_cap_out, all_cap_pos_ids, all_cap_pad_mask = [], [], []
 
-        all_image_out = []
-        all_image_size = []
-        all_image_pos_ids = []
-        all_image_pad_mask = []
-        all_cap_pos_ids = []
-        all_cap_pad_mask = []
-        all_cap_feats_out = []
-
-        for i, (image, cap_feat) in enumerate(zip(all_image, all_cap_feats)):
-            ### Process Caption
-            cap_ori_len = len(cap_feat)
-            cap_padding_len = (-cap_ori_len) % SEQ_MULTI_OF
-            # padded position ids
-            cap_padded_pos_ids = self.create_coordinate_grid(
-                size=(cap_ori_len + cap_padding_len, 1, 1),
-                start=(1, 0, 0),
-                device=device,
-            ).flatten(0, 2)
-            all_cap_pos_ids.append(cap_padded_pos_ids)
-            # pad mask
-            cap_pad_mask = torch.cat(
-                [
-                    torch.zeros((cap_ori_len,), dtype=torch.bool, device=device),
-                    torch.ones((cap_padding_len,), dtype=torch.bool, device=device),
-                ],
-                dim=0,
-            )
-            all_cap_pad_mask.append(
-                cap_pad_mask if cap_padding_len > 0 else torch.zeros((cap_ori_len,), dtype=torch.bool, device=device)
+        for image, cap_feat in zip(all_image, all_cap_feats):
+            # Caption
+            cap_out, cap_pos_ids, cap_pad_mask, cap_len, _ = self._pad_with_ids(
+                cap_feat, (len(cap_feat) + (-len(cap_feat)) % SEQ_MULTI_OF, 1, 1), (1, 0, 0), device
             )
+            all_cap_out.append(cap_out)
+            all_cap_pos_ids.append(cap_pos_ids)
+            all_cap_pad_mask.append(cap_pad_mask)
 
-            # padded feature
-            cap_padded_feat = torch.cat([cap_feat, cap_feat[-1:].repeat(cap_padding_len, 1)], dim=0)
-            all_cap_feats_out.append(cap_padded_feat)
-
-            ### Process Image
-            C, F, H, W = image.size()
-            all_image_size.append((F, H, W))
-            F_tokens, H_tokens, W_tokens = F // pF, H // pH, W // pW
-
-            image = image.view(C, F_tokens, pF, H_tokens, pH, W_tokens, pW)
-            # "c f pf h ph w pw -> (f h w) (pf ph pw c)"
-            image = image.permute(1, 3, 5, 2, 4, 6, 0).reshape(F_tokens * H_tokens * W_tokens, pF * pH * pW * C)
-
-            image_ori_len = len(image)
-            image_padding_len = (-image_ori_len) % SEQ_MULTI_OF
-
-            image_ori_pos_ids = self.create_coordinate_grid(
-                size=(F_tokens, H_tokens, W_tokens),
-                start=(cap_ori_len + cap_padding_len + 1, 0, 0),
-                device=device,
-            ).flatten(0, 2)
-            image_padded_pos_ids = torch.cat(
-                [
-                    image_ori_pos_ids,
-                    self.create_coordinate_grid(size=(1, 1, 1), start=(0, 0, 0), device=device)
-                    .flatten(0, 2)
-                    .repeat(image_padding_len, 1),
-                ],
-                dim=0,
+            # Image
+            img_patches, size, (F_t, H_t, W_t) = self._patchify_image(image, patch_size, f_patch_size)
+            img_out, img_pos_ids, img_pad_mask, _, _ = self._pad_with_ids(
+                img_patches, (F_t, H_t, W_t), (cap_len + 1, 0, 0), device
             )
-            all_image_pos_ids.append(image_padded_pos_ids if image_padding_len > 0 else image_ori_pos_ids)
-            # pad mask
-            image_pad_mask = torch.cat(
-                [
-                    torch.zeros((image_ori_len,), dtype=torch.bool, device=device),
-                    torch.ones((image_padding_len,), dtype=torch.bool, device=device),
-                ],
-                dim=0,
-            )
-            all_image_pad_mask.append(
-                image_pad_mask
-                if image_padding_len > 0
-                else torch.zeros((image_ori_len,), dtype=torch.bool, device=device)
-            )
-            # padded feature
-            image_padded_feat = torch.cat(
-                [image, image[-1:].repeat(image_padding_len, 1)],
-                dim=0,
-            )
-            all_image_out.append(image_padded_feat if image_padding_len > 0 else image)
+            all_img_out.append(img_out)
+            all_img_size.append(size)
+            all_img_pos_ids.append(img_pos_ids)
+            all_img_pad_mask.append(img_pad_mask)
 
         return (
-            all_image_out,
-            all_cap_feats_out,
-            all_image_size,
-            all_image_pos_ids,
+            all_img_out,
+            all_cap_out,
+            all_img_size,
+            all_img_pos_ids,
             all_cap_pos_ids,
-            all_image_pad_mask,
+            all_img_pad_mask,
             all_cap_pad_mask,
         )
 

src/diffusers/models/transformers/transformer_flux.py

@@ -22,7 +22,7 @@ import torch.nn.functional as F
 
 from ...configuration_utils import ConfigMixin, register_to_config
 from ...loaders import FluxTransformer2DLoadersMixin, FromOriginalModelMixin, PeftAdapterMixin
-from ...utils import USE_PEFT_BACKEND, is_torch_npu_available, logging, scale_lora_layers, unscale_lora_layers
+from ...utils import USE_PEFT_BACKEND, logging, scale_lora_layers, unscale_lora_layers
 from ...utils.torch_utils import maybe_allow_in_graph
 from .._modeling_parallel import ContextParallelInput, ContextParallelOutput
 from ..attention import AttentionMixin, AttentionModuleMixin, FeedForward
@@ -717,11 +717,7 @@ class FluxTransformer2DModel(
             img_ids = img_ids[0]
 
         ids = torch.cat((txt_ids, img_ids), dim=0)
-        if is_torch_npu_available():
-            freqs_cos, freqs_sin = self.pos_embed(ids.cpu())
-            image_rotary_emb = (freqs_cos.npu(), freqs_sin.npu())
-        else:
-            image_rotary_emb = self.pos_embed(ids)
+        image_rotary_emb = self.pos_embed(ids)
 
         if joint_attention_kwargs is not None and "ip_adapter_image_embeds" in joint_attention_kwargs:
             ip_adapter_image_embeds = joint_attention_kwargs.pop("ip_adapter_image_embeds")

src/diffusers/models/transformers/transformer_flux2.py

@@ -21,7 +21,7 @@ import torch.nn.functional as F
 
 from ...configuration_utils import ConfigMixin, register_to_config
 from ...loaders import FluxTransformer2DLoadersMixin, FromOriginalModelMixin, PeftAdapterMixin
-from ...utils import USE_PEFT_BACKEND, is_torch_npu_available, logging, scale_lora_layers, unscale_lora_layers
+from ...utils import USE_PEFT_BACKEND, logging, scale_lora_layers, unscale_lora_layers
 from .._modeling_parallel import ContextParallelInput, ContextParallelOutput
 from ..attention import AttentionMixin, AttentionModuleMixin
 from ..attention_dispatch import dispatch_attention_fn
@@ -835,14 +835,8 @@ class Flux2Transformer2DModel(
         if txt_ids.ndim == 3:
             txt_ids = txt_ids[0]
 
-        if is_torch_npu_available():
-            freqs_cos_image, freqs_sin_image = self.pos_embed(img_ids.cpu())
-            image_rotary_emb = (freqs_cos_image.npu(), freqs_sin_image.npu())
-            freqs_cos_text, freqs_sin_text = self.pos_embed(txt_ids.cpu())
-            text_rotary_emb = (freqs_cos_text.npu(), freqs_sin_text.npu())
-        else:
-            image_rotary_emb = self.pos_embed(img_ids)
-            text_rotary_emb = self.pos_embed(txt_ids)
+        image_rotary_emb = self.pos_embed(img_ids)
+        text_rotary_emb = self.pos_embed(txt_ids)
         concat_rotary_emb = (
             torch.cat([text_rotary_emb[0], image_rotary_emb[0]], dim=0),
             torch.cat([text_rotary_emb[1], image_rotary_emb[1]], dim=0),

src/diffusers/models/transformers/transformer_longcat_image.py

@@ -21,7 +21,7 @@ import torch.nn.functional as F
 
 from ...configuration_utils import ConfigMixin, register_to_config
 from ...loaders import FromOriginalModelMixin, PeftAdapterMixin
-from ...utils import is_torch_npu_available, logging
+from ...utils import logging
 from ...utils.torch_utils import maybe_allow_in_graph
 from ..attention import AttentionModuleMixin, FeedForward
 from ..attention_dispatch import dispatch_attention_fn
@@ -499,11 +499,7 @@ class LongCatImageTransformer2DModel(
         encoder_hidden_states = self.context_embedder(encoder_hidden_states)
 
         ids = torch.cat((txt_ids, img_ids), dim=0)
-        if is_torch_npu_available():
-            freqs_cos, freqs_sin = self.pos_embed(ids.cpu())
-            image_rotary_emb = (freqs_cos.npu(), freqs_sin.npu())
-        else:
-            image_rotary_emb = self.pos_embed(ids)
+        image_rotary_emb = self.pos_embed(ids)
 
         for index_block, block in enumerate(self.transformer_blocks):
             if torch.is_grad_enabled() and self.gradient_checkpointing and self.use_checkpoint[index_block]:

src/diffusers/models/transformers/transformer_ovis_image.py

@@ -21,7 +21,7 @@ import torch.nn.functional as F
 
 from ...configuration_utils import ConfigMixin, register_to_config
 from ...loaders import FromOriginalModelMixin, PeftAdapterMixin
-from ...utils import is_torch_npu_available, logging
+from ...utils import logging
 from ...utils.torch_utils import maybe_allow_in_graph
 from ..attention import AttentionModuleMixin, FeedForward
 from ..attention_dispatch import dispatch_attention_fn
@@ -530,11 +530,7 @@ class OvisImageTransformer2DModel(
             img_ids = img_ids[0]
 
         ids = torch.cat((txt_ids, img_ids), dim=0)
-        if is_torch_npu_available():
-            freqs_cos, freqs_sin = self.pos_embed(ids.cpu())
-            image_rotary_emb = (freqs_cos.npu(), freqs_sin.npu())
-        else:
-            image_rotary_emb = self.pos_embed(ids)
+        image_rotary_emb = self.pos_embed(ids)
 
         for index_block, block in enumerate(self.transformer_blocks):
             if torch.is_grad_enabled() and self.gradient_checkpointing:

src/diffusers/models/transformers/transformer_wan.py

@@ -134,7 +134,8 @@ class WanAttnProcessor:
                 dropout_p=0.0,
                 is_causal=False,
                 backend=self._attention_backend,
-                parallel_config=self._parallel_config,
+                # Reference: https://github.com/huggingface/diffusers/pull/12909
+                parallel_config=None,
             )
             hidden_states_img = hidden_states_img.flatten(2, 3)
             hidden_states_img = hidden_states_img.type_as(query)
@@ -147,7 +148,8 @@ class WanAttnProcessor:
             dropout_p=0.0,
             is_causal=False,
             backend=self._attention_backend,
-            parallel_config=self._parallel_config,
+            # Reference: https://github.com/huggingface/diffusers/pull/12909
+            parallel_config=(self._parallel_config if encoder_hidden_states is None else None),
         )
         hidden_states = hidden_states.flatten(2, 3)
         hidden_states = hidden_states.type_as(query)
@@ -552,9 +554,11 @@ class WanTransformer3DModel(
         "blocks.0": {
             "hidden_states": ContextParallelInput(split_dim=1, expected_dims=3, split_output=False),
         },
-        "blocks.*": {
-            "encoder_hidden_states": ContextParallelInput(split_dim=1, expected_dims=3, split_output=False),
-        },
+        # Reference: https://github.com/huggingface/diffusers/pull/12909
+        # We need to disable the splitting of encoder_hidden_states because the image_encoder
+        # (Wan 2.1 I2V) consistently generates 257 tokens for image_embed. This causes the shape
+        # of encoder_hidden_states—whose token count is always 769 (512 + 257) after concatenation
+        # —to be indivisible by the number of devices in the CP.
         "proj_out": ContextParallelOutput(gather_dim=1, expected_dims=3),
         "": {
             "timestep": ContextParallelInput(split_dim=1, expected_dims=2, split_output=False),

src/diffusers/models/transformers/transformer_wan_animate.py

@@ -609,7 +609,8 @@ class WanAttnProcessor:
                 dropout_p=0.0,
                 is_causal=False,
                 backend=self._attention_backend,
-                parallel_config=self._parallel_config,
+                # Reference: https://github.com/huggingface/diffusers/pull/12909
+                parallel_config=None,
             )
             hidden_states_img = hidden_states_img.flatten(2, 3)
             hidden_states_img = hidden_states_img.type_as(query)
@@ -622,7 +623,8 @@ class WanAttnProcessor:
             dropout_p=0.0,
             is_causal=False,
             backend=self._attention_backend,
-            parallel_config=self._parallel_config,
+            # Reference: https://github.com/huggingface/diffusers/pull/12909
+            parallel_config=(self._parallel_config if encoder_hidden_states is None else None),
         )
         hidden_states = hidden_states.flatten(2, 3)
         hidden_states = hidden_states.type_as(query)

src/diffusers/models/transformers/transformer_z_image.py

@@ -13,7 +13,7 @@
 # limitations under the License.
 
 import math
-from typing import Dict, List, Optional, Tuple
+from typing import Dict, List, Optional, Tuple, Union
 
 import torch
 import torch.nn as nn
@@ -32,6 +32,7 @@ from ..modeling_outputs import Transformer2DModelOutput
 
 ADALN_EMBED_DIM = 256
 SEQ_MULTI_OF = 32
+X_PAD_DIM = 64
 
 
 class TimestepEmbedder(nn.Module):
@@ -152,6 +153,20 @@ class ZSingleStreamAttnProcessor:
         return output
 
 
+def select_per_token(
+    value_noisy: torch.Tensor,
+    value_clean: torch.Tensor,
+    noise_mask: torch.Tensor,
+    seq_len: int,
+) -> torch.Tensor:
+    noise_mask_expanded = noise_mask.unsqueeze(-1)  # (batch, seq_len, 1)
+    return torch.where(
+        noise_mask_expanded == 1,
+        value_noisy.unsqueeze(1).expand(-1, seq_len, -1),
+        value_clean.unsqueeze(1).expand(-1, seq_len, -1),
+    )
+
+
 class FeedForward(nn.Module):
     def __init__(self, dim: int, hidden_dim: int):
         super().__init__()
@@ -215,12 +230,37 @@ class ZImageTransformerBlock(nn.Module):
         attn_mask: torch.Tensor,
         freqs_cis: torch.Tensor,
         adaln_input: Optional[torch.Tensor] = None,
+        noise_mask: Optional[torch.Tensor] = None,
+        adaln_noisy: Optional[torch.Tensor] = None,
+        adaln_clean: Optional[torch.Tensor] = None,
     ):
         if self.modulation:
-            assert adaln_input is not None
-            scale_msa, gate_msa, scale_mlp, gate_mlp = self.adaLN_modulation(adaln_input).unsqueeze(1).chunk(4, dim=2)
-            gate_msa, gate_mlp = gate_msa.tanh(), gate_mlp.tanh()
-            scale_msa, scale_mlp = 1.0 + scale_msa, 1.0 + scale_mlp
+            seq_len = x.shape[1]
+
+            if noise_mask is not None:
+                # Per-token modulation: different modulation for noisy/clean tokens
+                mod_noisy = self.adaLN_modulation(adaln_noisy)
+                mod_clean = self.adaLN_modulation(adaln_clean)
+
+                scale_msa_noisy, gate_msa_noisy, scale_mlp_noisy, gate_mlp_noisy = mod_noisy.chunk(4, dim=1)
+                scale_msa_clean, gate_msa_clean, scale_mlp_clean, gate_mlp_clean = mod_clean.chunk(4, dim=1)
+
+                gate_msa_noisy, gate_mlp_noisy = gate_msa_noisy.tanh(), gate_mlp_noisy.tanh()
+                gate_msa_clean, gate_mlp_clean = gate_msa_clean.tanh(), gate_mlp_clean.tanh()
+
+                scale_msa_noisy, scale_mlp_noisy = 1.0 + scale_msa_noisy, 1.0 + scale_mlp_noisy
+                scale_msa_clean, scale_mlp_clean = 1.0 + scale_msa_clean, 1.0 + scale_mlp_clean
+
+                scale_msa = select_per_token(scale_msa_noisy, scale_msa_clean, noise_mask, seq_len)
+                scale_mlp = select_per_token(scale_mlp_noisy, scale_mlp_clean, noise_mask, seq_len)
+                gate_msa = select_per_token(gate_msa_noisy, gate_msa_clean, noise_mask, seq_len)
+                gate_mlp = select_per_token(gate_mlp_noisy, gate_mlp_clean, noise_mask, seq_len)
+            else:
+                # Global modulation: same modulation for all tokens (avoid double select)
+                mod = self.adaLN_modulation(adaln_input)
+                scale_msa, gate_msa, scale_mlp, gate_mlp = mod.unsqueeze(1).chunk(4, dim=2)
+                gate_msa, gate_mlp = gate_msa.tanh(), gate_mlp.tanh()
+                scale_msa, scale_mlp = 1.0 + scale_msa, 1.0 + scale_mlp
 
             # Attention block
             attn_out = self.attention(
@@ -252,9 +292,21 @@ class FinalLayer(nn.Module):
             nn.Linear(min(hidden_size, ADALN_EMBED_DIM), hidden_size, bias=True),
         )
 
-    def forward(self, x, c):
-        scale = 1.0 + self.adaLN_modulation(c)
-        x = self.norm_final(x) * scale.unsqueeze(1)
+    def forward(self, x, c=None, noise_mask=None, c_noisy=None, c_clean=None):
+        seq_len = x.shape[1]
+
+        if noise_mask is not None:
+            # Per-token modulation
+            scale_noisy = 1.0 + self.adaLN_modulation(c_noisy)
+            scale_clean = 1.0 + self.adaLN_modulation(c_clean)
+            scale = select_per_token(scale_noisy, scale_clean, noise_mask, seq_len)
+        else:
+            # Original global modulation
+            assert c is not None, "Either c or (c_noisy, c_clean) must be provided"
+            scale = 1.0 + self.adaLN_modulation(c)
+            scale = scale.unsqueeze(1)
+
+        x = self.norm_final(x) * scale
         x = self.linear(x)
         return x
 
@@ -325,6 +377,7 @@ class ZImageTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOr
         norm_eps=1e-5,
         qk_norm=True,
         cap_feat_dim=2560,
+        siglip_feat_dim=None,  # Optional: set to enable SigLIP support for Omni
         rope_theta=256.0,
         t_scale=1000.0,
         axes_dims=[32, 48, 48],
@@ -386,6 +439,31 @@ class ZImageTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOr
         self.t_embedder = TimestepEmbedder(min(dim, ADALN_EMBED_DIM), mid_size=1024)
         self.cap_embedder = nn.Sequential(RMSNorm(cap_feat_dim, eps=norm_eps), nn.Linear(cap_feat_dim, dim, bias=True))
 
+        # Optional SigLIP components (for Omni variant)
+        if siglip_feat_dim is not None:
+            self.siglip_embedder = nn.Sequential(
+                RMSNorm(siglip_feat_dim, eps=norm_eps), nn.Linear(siglip_feat_dim, dim, bias=True)
+            )
+            self.siglip_refiner = nn.ModuleList(
+                [
+                    ZImageTransformerBlock(
+                        2000 + layer_id,
+                        dim,
+                        n_heads,
+                        n_kv_heads,
+                        norm_eps,
+                        qk_norm,
+                        modulation=False,
+                    )
+                    for layer_id in range(n_refiner_layers)
+                ]
+            )
+            self.siglip_pad_token = nn.Parameter(torch.empty((1, dim)))
+        else:
+            self.siglip_embedder = None
+            self.siglip_refiner = None
+            self.siglip_pad_token = None
+
         self.x_pad_token = nn.Parameter(torch.empty((1, dim)))
         self.cap_pad_token = nn.Parameter(torch.empty((1, dim)))
 
@@ -402,259 +480,561 @@ class ZImageTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOr
 
         self.rope_embedder = RopeEmbedder(theta=rope_theta, axes_dims=axes_dims, axes_lens=axes_lens)
 
-    def unpatchify(self, x: List[torch.Tensor], size: List[Tuple], patch_size, f_patch_size) -> List[torch.Tensor]:
+    def unpatchify(
+        self,
+        x: List[torch.Tensor],
+        size: List[Tuple],
+        patch_size,
+        f_patch_size,
+        x_pos_offsets: Optional[List[Tuple[int, int]]] = None,
+    ) -> List[torch.Tensor]:
         pH = pW = patch_size
         pF = f_patch_size
         bsz = len(x)
         assert len(size) == bsz
-        for i in range(bsz):
-            F, H, W = size[i]
-            ori_len = (F // pF) * (H // pH) * (W // pW)
-            # "f h w pf ph pw c -> c (f pf) (h ph) (w pw)"
-            x[i] = (
-                x[i][:ori_len]
-                .view(F // pF, H // pH, W // pW, pF, pH, pW, self.out_channels)
-                .permute(6, 0, 3, 1, 4, 2, 5)
-                .reshape(self.out_channels, F, H, W)
-            )
-        return x
+
+        if x_pos_offsets is not None:
+            # Omni: extract target image from unified sequence (cond_images + target)
+            result = []
+            for i in range(bsz):
+                unified_x = x[i][x_pos_offsets[i][0] : x_pos_offsets[i][1]]
+                cu_len = 0
+                x_item = None
+                for j in range(len(size[i])):
+                    if size[i][j] is None:
+                        ori_len = 0
+                        pad_len = SEQ_MULTI_OF
+                        cu_len += pad_len + ori_len
+                    else:
+                        F, H, W = size[i][j]
+                        ori_len = (F // pF) * (H // pH) * (W // pW)
+                        pad_len = (-ori_len) % SEQ_MULTI_OF
+                        x_item = (
+                            unified_x[cu_len : cu_len + ori_len]
+                            .view(F // pF, H // pH, W // pW, pF, pH, pW, self.out_channels)
+                            .permute(6, 0, 3, 1, 4, 2, 5)
+                            .reshape(self.out_channels, F, H, W)
+                        )
+                        cu_len += ori_len + pad_len
+                result.append(x_item)  # Return only the last (target) image
+            return result
+        else:
+            # Original mode: simple unpatchify
+            for i in range(bsz):
+                F, H, W = size[i]
+                ori_len = (F // pF) * (H // pH) * (W // pW)
+                # "f h w pf ph pw c -> c (f pf) (h ph) (w pw)"
+                x[i] = (
+                    x[i][:ori_len]
+                    .view(F // pF, H // pH, W // pW, pF, pH, pW, self.out_channels)
+                    .permute(6, 0, 3, 1, 4, 2, 5)
+                    .reshape(self.out_channels, F, H, W)
+                )
+            return x
 
     @staticmethod
     def create_coordinate_grid(size, start=None, device=None):
         if start is None:
             start = (0 for _ in size)
-
         axes = [torch.arange(x0, x0 + span, dtype=torch.int32, device=device) for x0, span in zip(start, size)]
         grids = torch.meshgrid(axes, indexing="ij")
         return torch.stack(grids, dim=-1)
 
-    def patchify_and_embed(
+    def _patchify_image(self, image: torch.Tensor, patch_size: int, f_patch_size: int):
+        """Patchify a single image tensor: (C, F, H, W) -> (num_patches, patch_dim)."""
+        pH, pW, pF = patch_size, patch_size, f_patch_size
+        C, F, H, W = image.size()
+        F_tokens, H_tokens, W_tokens = F // pF, H // pH, W // pW
+        image = image.view(C, F_tokens, pF, H_tokens, pH, W_tokens, pW)
+        image = image.permute(1, 3, 5, 2, 4, 6, 0).reshape(F_tokens * H_tokens * W_tokens, pF * pH * pW * C)
+        return image, (F, H, W), (F_tokens, H_tokens, W_tokens)
+
+    def _pad_with_ids(
         self,
-        all_image: List[torch.Tensor],
-        all_cap_feats: List[torch.Tensor],
-        patch_size: int,
-        f_patch_size: int,
+        feat: torch.Tensor,
+        pos_grid_size: Tuple,
+        pos_start: Tuple,
+        device: torch.device,
+        noise_mask_val: Optional[int] = None,
     ):
-        pH = pW = patch_size
-        pF = f_patch_size
+        """Pad feature to SEQ_MULTI_OF, create position IDs and pad mask."""
+        ori_len = len(feat)
+        pad_len = (-ori_len) % SEQ_MULTI_OF
+        total_len = ori_len + pad_len
+
+        # Pos IDs
+        ori_pos_ids = self.create_coordinate_grid(size=pos_grid_size, start=pos_start, device=device).flatten(0, 2)
+        if pad_len > 0:
+            pad_pos_ids = (
+                self.create_coordinate_grid(size=(1, 1, 1), start=(0, 0, 0), device=device)
+                .flatten(0, 2)
+                .repeat(pad_len, 1)
+            )
+            pos_ids = torch.cat([ori_pos_ids, pad_pos_ids], dim=0)
+            padded_feat = torch.cat([feat, feat[-1:].repeat(pad_len, 1)], dim=0)
+            pad_mask = torch.cat(
+                [
+                    torch.zeros(ori_len, dtype=torch.bool, device=device),
+                    torch.ones(pad_len, dtype=torch.bool, device=device),
+                ]
+            )
+        else:
+            pos_ids = ori_pos_ids
+            padded_feat = feat
+            pad_mask = torch.zeros(ori_len, dtype=torch.bool, device=device)
+
+        noise_mask = [noise_mask_val] * total_len if noise_mask_val is not None else None  # token level
+        return padded_feat, pos_ids, pad_mask, total_len, noise_mask
+
+    def patchify_and_embed(
+        self, all_image: List[torch.Tensor], all_cap_feats: List[torch.Tensor], patch_size: int, f_patch_size: int
+    ):
+        """Patchify for basic mode: single image per batch item."""
         device = all_image[0].device
+        all_img_out, all_img_size, all_img_pos_ids, all_img_pad_mask = [], [], [], []
+        all_cap_out, all_cap_pos_ids, all_cap_pad_mask = [], [], []
 
-        all_image_out = []
-        all_image_size = []
-        all_image_pos_ids = []
-        all_image_pad_mask = []
-        all_cap_pos_ids = []
-        all_cap_pad_mask = []
-        all_cap_feats_out = []
-
-        for i, (image, cap_feat) in enumerate(zip(all_image, all_cap_feats)):
-            ### Process Caption
-            cap_ori_len = len(cap_feat)
-            cap_padding_len = (-cap_ori_len) % SEQ_MULTI_OF
-            # padded position ids
-            cap_padded_pos_ids = self.create_coordinate_grid(
-                size=(cap_ori_len + cap_padding_len, 1, 1),
-                start=(1, 0, 0),
-                device=device,
-            ).flatten(0, 2)
-            all_cap_pos_ids.append(cap_padded_pos_ids)
-            # pad mask
-            cap_pad_mask = torch.cat(
-                [
-                    torch.zeros((cap_ori_len,), dtype=torch.bool, device=device),
-                    torch.ones((cap_padding_len,), dtype=torch.bool, device=device),
-                ],
-                dim=0,
-            )
-            all_cap_pad_mask.append(
-                cap_pad_mask if cap_padding_len > 0 else torch.zeros((cap_ori_len,), dtype=torch.bool, device=device)
+        for image, cap_feat in zip(all_image, all_cap_feats):
+            # Caption
+            cap_out, cap_pos_ids, cap_pad_mask, cap_len, _ = self._pad_with_ids(
+                cap_feat, (len(cap_feat) + (-len(cap_feat)) % SEQ_MULTI_OF, 1, 1), (1, 0, 0), device
             )
+            all_cap_out.append(cap_out)
+            all_cap_pos_ids.append(cap_pos_ids)
+            all_cap_pad_mask.append(cap_pad_mask)
 
-            # padded feature
-            cap_padded_feat = torch.cat([cap_feat, cap_feat[-1:].repeat(cap_padding_len, 1)], dim=0)
-            all_cap_feats_out.append(cap_padded_feat)
-
-            ### Process Image
-            C, F, H, W = image.size()
-            all_image_size.append((F, H, W))
-            F_tokens, H_tokens, W_tokens = F // pF, H // pH, W // pW
-
-            image = image.view(C, F_tokens, pF, H_tokens, pH, W_tokens, pW)
-            # "c f pf h ph w pw -> (f h w) (pf ph pw c)"
-            image = image.permute(1, 3, 5, 2, 4, 6, 0).reshape(F_tokens * H_tokens * W_tokens, pF * pH * pW * C)
-
-            image_ori_len = len(image)
-            image_padding_len = (-image_ori_len) % SEQ_MULTI_OF
-
-            image_ori_pos_ids = self.create_coordinate_grid(
-                size=(F_tokens, H_tokens, W_tokens),
-                start=(cap_ori_len + cap_padding_len + 1, 0, 0),
-                device=device,
-            ).flatten(0, 2)
-            image_padded_pos_ids = torch.cat(
-                [
-                    image_ori_pos_ids,
-                    self.create_coordinate_grid(size=(1, 1, 1), start=(0, 0, 0), device=device)
-                    .flatten(0, 2)
-                    .repeat(image_padding_len, 1),
-                ],
-                dim=0,
+            # Image
+            img_patches, size, (F_t, H_t, W_t) = self._patchify_image(image, patch_size, f_patch_size)
+            img_out, img_pos_ids, img_pad_mask, _, _ = self._pad_with_ids(
+                img_patches, (F_t, H_t, W_t), (cap_len + 1, 0, 0), device
             )
-            all_image_pos_ids.append(image_padded_pos_ids if image_padding_len > 0 else image_ori_pos_ids)
-            # pad mask
-            image_pad_mask = torch.cat(
-                [
-                    torch.zeros((image_ori_len,), dtype=torch.bool, device=device),
-                    torch.ones((image_padding_len,), dtype=torch.bool, device=device),
-                ],
-                dim=0,
-            )
-            all_image_pad_mask.append(
-                image_pad_mask
-                if image_padding_len > 0
-                else torch.zeros((image_ori_len,), dtype=torch.bool, device=device)
-            )
-            # padded feature
-            image_padded_feat = torch.cat(
-                [image, image[-1:].repeat(image_padding_len, 1)],
-                dim=0,
-            )
-            all_image_out.append(image_padded_feat if image_padding_len > 0 else image)
+            all_img_out.append(img_out)
+            all_img_size.append(size)
+            all_img_pos_ids.append(img_pos_ids)
+            all_img_pad_mask.append(img_pad_mask)
 
         return (
-            all_image_out,
-            all_cap_feats_out,
-            all_image_size,
-            all_image_pos_ids,
+            all_img_out,
+            all_cap_out,
+            all_img_size,
+            all_img_pos_ids,
             all_cap_pos_ids,
-            all_image_pad_mask,
+            all_img_pad_mask,
             all_cap_pad_mask,
         )
 
-    def forward(
+    def patchify_and_embed_omni(
         self,
-        x: List[torch.Tensor],
-        t,
-        cap_feats: List[torch.Tensor],
-        controlnet_block_samples: Optional[Dict[int, torch.Tensor]] = None,
-        patch_size=2,
-        f_patch_size=1,
-        return_dict: bool = True,
+        all_x: List[List[torch.Tensor]],
+        all_cap_feats: List[List[torch.Tensor]],
+        all_siglip_feats: List[List[torch.Tensor]],
+        patch_size: int,
+        f_patch_size: int,
+        images_noise_mask: List[List[int]],
     ):
-        assert patch_size in self.all_patch_size
-        assert f_patch_size in self.all_f_patch_size
+        """Patchify for omni mode: multiple images per batch item with noise masks."""
+        bsz = len(all_x)
+        device = all_x[0][-1].device
+        dtype = all_x[0][-1].dtype
 
-        bsz = len(x)
-        device = x[0].device
-        t = t * self.t_scale
-        t = self.t_embedder(t)
+        all_x_out, all_x_size, all_x_pos_ids, all_x_pad_mask, all_x_len, all_x_noise_mask = [], [], [], [], [], []
+        all_cap_out, all_cap_pos_ids, all_cap_pad_mask, all_cap_len, all_cap_noise_mask = [], [], [], [], []
+        all_sig_out, all_sig_pos_ids, all_sig_pad_mask, all_sig_len, all_sig_noise_mask = [], [], [], [], []
 
-        (
-            x,
-            cap_feats,
-            x_size,
-            x_pos_ids,
-            cap_pos_ids,
-            x_inner_pad_mask,
-            cap_inner_pad_mask,
-        ) = self.patchify_and_embed(x, cap_feats, patch_size, f_patch_size)
+        for i in range(bsz):
+            num_images = len(all_x[i])
+            cap_feats_list, cap_pos_list, cap_mask_list, cap_lens, cap_noise = [], [], [], [], []
+            cap_end_pos = []
+            cap_cu_len = 1
 
-        # x embed & refine
-        x_item_seqlens = [len(_) for _ in x]
-        assert all(_ % SEQ_MULTI_OF == 0 for _ in x_item_seqlens)
-        x_max_item_seqlen = max(x_item_seqlens)
+            # Process captions
+            for j, cap_item in enumerate(all_cap_feats[i]):
+                noise_val = images_noise_mask[i][j] if j < len(images_noise_mask[i]) else 1
+                cap_out, cap_pos, cap_mask, cap_len, cap_nm = self._pad_with_ids(
+                    cap_item,
+                    (len(cap_item) + (-len(cap_item)) % SEQ_MULTI_OF, 1, 1),
+                    (cap_cu_len, 0, 0),
+                    device,
+                    noise_val,
+                )
+                cap_feats_list.append(cap_out)
+                cap_pos_list.append(cap_pos)
+                cap_mask_list.append(cap_mask)
+                cap_lens.append(cap_len)
+                cap_noise.extend(cap_nm)
+                cap_cu_len += len(cap_item)
+                cap_end_pos.append(cap_cu_len)
+                cap_cu_len += 2  # for image vae and siglip tokens
 
-        x = torch.cat(x, dim=0)
-        x = self.all_x_embedder[f"{patch_size}-{f_patch_size}"](x)
+            all_cap_out.append(torch.cat(cap_feats_list, dim=0))
+            all_cap_pos_ids.append(torch.cat(cap_pos_list, dim=0))
+            all_cap_pad_mask.append(torch.cat(cap_mask_list, dim=0))
+            all_cap_len.append(cap_lens)
+            all_cap_noise_mask.append(cap_noise)
 
-        # Match t_embedder output dtype to x for layerwise casting compatibility
-        adaln_input = t.type_as(x)
-        x[torch.cat(x_inner_pad_mask)] = self.x_pad_token
-        x = list(x.split(x_item_seqlens, dim=0))
-        x_freqs_cis = list(self.rope_embedder(torch.cat(x_pos_ids, dim=0)).split([len(_) for _ in x_pos_ids], dim=0))
+            # Process images
+            x_feats_list, x_pos_list, x_mask_list, x_lens, x_size, x_noise = [], [], [], [], [], []
+            for j, x_item in enumerate(all_x[i]):
+                noise_val = images_noise_mask[i][j]
+                if x_item is not None:
+                    x_patches, size, (F_t, H_t, W_t) = self._patchify_image(x_item, patch_size, f_patch_size)
+                    x_out, x_pos, x_mask, x_len, x_nm = self._pad_with_ids(
+                        x_patches, (F_t, H_t, W_t), (cap_end_pos[j], 0, 0), device, noise_val
+                    )
+                    x_size.append(size)
+                else:
+                    x_len = SEQ_MULTI_OF
+                    x_out = torch.zeros((x_len, X_PAD_DIM), dtype=dtype, device=device)
+                    x_pos = self.create_coordinate_grid((1, 1, 1), (0, 0, 0), device).flatten(0, 2).repeat(x_len, 1)
+                    x_mask = torch.ones(x_len, dtype=torch.bool, device=device)
+                    x_nm = [noise_val] * x_len
+                    x_size.append(None)
+                x_feats_list.append(x_out)
+                x_pos_list.append(x_pos)
+                x_mask_list.append(x_mask)
+                x_lens.append(x_len)
+                x_noise.extend(x_nm)
 
-        x = pad_sequence(x, batch_first=True, padding_value=0.0)
-        x_freqs_cis = pad_sequence(x_freqs_cis, batch_first=True, padding_value=0.0)
-        # Clarify the length matches to satisfy Dynamo due to "Symbolic Shape Inference" to avoid compilation errors
-        x_freqs_cis = x_freqs_cis[:, : x.shape[1]]
+            all_x_out.append(torch.cat(x_feats_list, dim=0))
+            all_x_pos_ids.append(torch.cat(x_pos_list, dim=0))
+            all_x_pad_mask.append(torch.cat(x_mask_list, dim=0))
+            all_x_size.append(x_size)
+            all_x_len.append(x_lens)
+            all_x_noise_mask.append(x_noise)
 
-        x_attn_mask = torch.zeros((bsz, x_max_item_seqlen), dtype=torch.bool, device=device)
-        for i, seq_len in enumerate(x_item_seqlens):
-            x_attn_mask[i, :seq_len] = 1
+            # Process siglip
+            if all_siglip_feats[i] is None:
+                all_sig_len.append([0] * num_images)
+                all_sig_out.append(None)
+            else:
+                sig_feats_list, sig_pos_list, sig_mask_list, sig_lens, sig_noise = [], [], [], [], []
+                for j, sig_item in enumerate(all_siglip_feats[i]):
+                    noise_val = images_noise_mask[i][j]
+                    if sig_item is not None:
+                        sig_H, sig_W, sig_C = sig_item.size()
+                        sig_flat = sig_item.permute(2, 0, 1).reshape(sig_H * sig_W, sig_C)
+                        sig_out, sig_pos, sig_mask, sig_len, sig_nm = self._pad_with_ids(
+                            sig_flat, (1, sig_H, sig_W), (cap_end_pos[j] + 1, 0, 0), device, noise_val
+                        )
+                        # Scale position IDs to match x resolution
+                        if x_size[j] is not None:
+                            sig_pos = sig_pos.float()
+                            sig_pos[..., 1] = sig_pos[..., 1] / max(sig_H - 1, 1) * (x_size[j][1] - 1)
+                            sig_pos[..., 2] = sig_pos[..., 2] / max(sig_W - 1, 1) * (x_size[j][2] - 1)
+                            sig_pos = sig_pos.to(torch.int32)
+                    else:
+                        sig_len = SEQ_MULTI_OF
+                        sig_out = torch.zeros((sig_len, self.config.siglip_feat_dim), dtype=dtype, device=device)
+                        sig_pos = (
+                            self.create_coordinate_grid((1, 1, 1), (0, 0, 0), device).flatten(0, 2).repeat(sig_len, 1)
+                        )
+                        sig_mask = torch.ones(sig_len, dtype=torch.bool, device=device)
+                        sig_nm = [noise_val] * sig_len
+                    sig_feats_list.append(sig_out)
+                    sig_pos_list.append(sig_pos)
+                    sig_mask_list.append(sig_mask)
+                    sig_lens.append(sig_len)
+                    sig_noise.extend(sig_nm)
 
-        if torch.is_grad_enabled() and self.gradient_checkpointing:
-            for layer in self.noise_refiner:
-                x = self._gradient_checkpointing_func(layer, x, x_attn_mask, x_freqs_cis, adaln_input)
-        else:
-            for layer in self.noise_refiner:
-                x = layer(x, x_attn_mask, x_freqs_cis, adaln_input)
+                all_sig_out.append(torch.cat(sig_feats_list, dim=0))
+                all_sig_pos_ids.append(torch.cat(sig_pos_list, dim=0))
+                all_sig_pad_mask.append(torch.cat(sig_mask_list, dim=0))
+                all_sig_len.append(sig_lens)
+                all_sig_noise_mask.append(sig_noise)
 
-        # cap embed & refine
-        cap_item_seqlens = [len(_) for _ in cap_feats]
-        cap_max_item_seqlen = max(cap_item_seqlens)
+        # Compute x position offsets
+        all_x_pos_offsets = [(sum(all_cap_len[i]), sum(all_cap_len[i]) + sum(all_x_len[i])) for i in range(bsz)]
 
-        cap_feats = torch.cat(cap_feats, dim=0)
-        cap_feats = self.cap_embedder(cap_feats)
-        cap_feats[torch.cat(cap_inner_pad_mask)] = self.cap_pad_token
-        cap_feats = list(cap_feats.split(cap_item_seqlens, dim=0))
-        cap_freqs_cis = list(
-            self.rope_embedder(torch.cat(cap_pos_ids, dim=0)).split([len(_) for _ in cap_pos_ids], dim=0)
+        return (
+            all_x_out,
+            all_cap_out,
+            all_sig_out,
+            all_x_size,
+            all_x_pos_ids,
+            all_cap_pos_ids,
+            all_sig_pos_ids,
+            all_x_pad_mask,
+            all_cap_pad_mask,
+            all_sig_pad_mask,
+            all_x_pos_offsets,
+            all_x_noise_mask,
+            all_cap_noise_mask,
+            all_sig_noise_mask,
         )
 
-        cap_feats = pad_sequence(cap_feats, batch_first=True, padding_value=0.0)
-        cap_freqs_cis = pad_sequence(cap_freqs_cis, batch_first=True, padding_value=0.0)
-        # Clarify the length matches to satisfy Dynamo due to "Symbolic Shape Inference" to avoid compilation errors
-        cap_freqs_cis = cap_freqs_cis[:, : cap_feats.shape[1]]
+    def _prepare_sequence(
+        self,
+        feats: List[torch.Tensor],
+        pos_ids: List[torch.Tensor],
+        inner_pad_mask: List[torch.Tensor],
+        pad_token: torch.nn.Parameter,
+        noise_mask: Optional[List[List[int]]] = None,
+        device: torch.device = None,
+    ):
+        """Prepare sequence: apply pad token, RoPE embed, pad to batch, create attention mask."""
+        item_seqlens = [len(f) for f in feats]
+        max_seqlen = max(item_seqlens)
+        bsz = len(feats)
 
-        cap_attn_mask = torch.zeros((bsz, cap_max_item_seqlen), dtype=torch.bool, device=device)
-        for i, seq_len in enumerate(cap_item_seqlens):
-            cap_attn_mask[i, :seq_len] = 1
+        # Pad token
+        feats_cat = torch.cat(feats, dim=0)
+        feats_cat[torch.cat(inner_pad_mask)] = pad_token
+        feats = list(feats_cat.split(item_seqlens, dim=0))
 
-        if torch.is_grad_enabled() and self.gradient_checkpointing:
-            for layer in self.context_refiner:
-                cap_feats = self._gradient_checkpointing_func(layer, cap_feats, cap_attn_mask, cap_freqs_cis)
-        else:
-            for layer in self.context_refiner:
-                cap_feats = layer(cap_feats, cap_attn_mask, cap_freqs_cis)
+        # RoPE
+        freqs_cis = list(self.rope_embedder(torch.cat(pos_ids, dim=0)).split([len(p) for p in pos_ids], dim=0))
 
-        # unified
+        # Pad to batch
+        feats = pad_sequence(feats, batch_first=True, padding_value=0.0)
+        freqs_cis = pad_sequence(freqs_cis, batch_first=True, padding_value=0.0)[:, : feats.shape[1]]
+
+        # Attention mask
+        attn_mask = torch.zeros((bsz, max_seqlen), dtype=torch.bool, device=device)
+        for i, seq_len in enumerate(item_seqlens):
+            attn_mask[i, :seq_len] = 1
+
+        # Noise mask
+        noise_mask_tensor = None
+        if noise_mask is not None:
+            noise_mask_tensor = pad_sequence(
+                [torch.tensor(m, dtype=torch.long, device=device) for m in noise_mask],
+                batch_first=True,
+                padding_value=0,
+            )[:, : feats.shape[1]]
+
+        return feats, freqs_cis, attn_mask, item_seqlens, noise_mask_tensor
+
+    def _build_unified_sequence(
+        self,
+        x: torch.Tensor,
+        x_freqs: torch.Tensor,
+        x_seqlens: List[int],
+        x_noise_mask: Optional[List[List[int]]],
+        cap: torch.Tensor,
+        cap_freqs: torch.Tensor,
+        cap_seqlens: List[int],
+        cap_noise_mask: Optional[List[List[int]]],
+        siglip: Optional[torch.Tensor],
+        siglip_freqs: Optional[torch.Tensor],
+        siglip_seqlens: Optional[List[int]],
+        siglip_noise_mask: Optional[List[List[int]]],
+        omni_mode: bool,
+        device: torch.device,
+    ):
+        """Build unified sequence: x, cap, and optionally siglip.
+        Basic mode order: [x, cap]; Omni mode order: [cap, x, siglip]
+        """
+        bsz = len(x_seqlens)
         unified = []
-        unified_freqs_cis = []
+        unified_freqs = []
+        unified_noise_mask = []
+
         for i in range(bsz):
-            x_len = x_item_seqlens[i]
-            cap_len = cap_item_seqlens[i]
-            unified.append(torch.cat([x[i][:x_len], cap_feats[i][:cap_len]]))
-            unified_freqs_cis.append(torch.cat([x_freqs_cis[i][:x_len], cap_freqs_cis[i][:cap_len]]))
-        unified_item_seqlens = [a + b for a, b in zip(cap_item_seqlens, x_item_seqlens)]
-        assert unified_item_seqlens == [len(_) for _ in unified]
-        unified_max_item_seqlen = max(unified_item_seqlens)
+            x_len, cap_len = x_seqlens[i], cap_seqlens[i]
 
-        unified = pad_sequence(unified, batch_first=True, padding_value=0.0)
-        unified_freqs_cis = pad_sequence(unified_freqs_cis, batch_first=True, padding_value=0.0)
-        unified_attn_mask = torch.zeros((bsz, unified_max_item_seqlen), dtype=torch.bool, device=device)
-        for i, seq_len in enumerate(unified_item_seqlens):
-            unified_attn_mask[i, :seq_len] = 1
+            if omni_mode:
+                # Omni: [cap, x, siglip]
+                if siglip is not None and siglip_seqlens is not None:
+                    sig_len = siglip_seqlens[i]
+                    unified.append(torch.cat([cap[i][:cap_len], x[i][:x_len], siglip[i][:sig_len]]))
+                    unified_freqs.append(
+                        torch.cat([cap_freqs[i][:cap_len], x_freqs[i][:x_len], siglip_freqs[i][:sig_len]])
+                    )
+                    unified_noise_mask.append(
+                        torch.tensor(
+                            cap_noise_mask[i] + x_noise_mask[i] + siglip_noise_mask[i], dtype=torch.long, device=device
+                        )
+                    )
+                else:
+                    unified.append(torch.cat([cap[i][:cap_len], x[i][:x_len]]))
+                    unified_freqs.append(torch.cat([cap_freqs[i][:cap_len], x_freqs[i][:x_len]]))
+                    unified_noise_mask.append(
+                        torch.tensor(cap_noise_mask[i] + x_noise_mask[i], dtype=torch.long, device=device)
+                    )
+            else:
+                # Basic: [x, cap]
+                unified.append(torch.cat([x[i][:x_len], cap[i][:cap_len]]))
+                unified_freqs.append(torch.cat([x_freqs[i][:x_len], cap_freqs[i][:cap_len]]))
 
-        if torch.is_grad_enabled() and self.gradient_checkpointing:
-            for layer_idx, layer in enumerate(self.layers):
-                unified = self._gradient_checkpointing_func(
-                    layer, unified, unified_attn_mask, unified_freqs_cis, adaln_input
-                )
-                if controlnet_block_samples is not None:
-                    if layer_idx in controlnet_block_samples:
-                        unified = unified + controlnet_block_samples[layer_idx]
+        # Compute unified seqlens
+        if omni_mode:
+            if siglip is not None and siglip_seqlens is not None:
+                unified_seqlens = [a + b + c for a, b, c in zip(cap_seqlens, x_seqlens, siglip_seqlens)]
+            else:
+                unified_seqlens = [a + b for a, b in zip(cap_seqlens, x_seqlens)]
         else:
-            for layer_idx, layer in enumerate(self.layers):
-                unified = layer(unified, unified_attn_mask, unified_freqs_cis, adaln_input)
-                if controlnet_block_samples is not None:
-                    if layer_idx in controlnet_block_samples:
-                        unified = unified + controlnet_block_samples[layer_idx]
+            unified_seqlens = [a + b for a, b in zip(x_seqlens, cap_seqlens)]
 
-        unified = self.all_final_layer[f"{patch_size}-{f_patch_size}"](unified, adaln_input)
-        unified = list(unified.unbind(dim=0))
-        x = self.unpatchify(unified, x_size, patch_size, f_patch_size)
+        max_seqlen = max(unified_seqlens)
 
-        if not return_dict:
-            return (x,)
+        # Pad to batch
+        unified = pad_sequence(unified, batch_first=True, padding_value=0.0)
+        unified_freqs = pad_sequence(unified_freqs, batch_first=True, padding_value=0.0)
 
-        return Transformer2DModelOutput(sample=x)
+        # Attention mask
+        attn_mask = torch.zeros((bsz, max_seqlen), dtype=torch.bool, device=device)
+        for i, seq_len in enumerate(unified_seqlens):
+            attn_mask[i, :seq_len] = 1
+
+        # Noise mask
+        noise_mask_tensor = None
+        if omni_mode:
+            noise_mask_tensor = pad_sequence(unified_noise_mask, batch_first=True, padding_value=0)[
+                :, : unified.shape[1]
+            ]
+
+        return unified, unified_freqs, attn_mask, noise_mask_tensor
+
+    def forward(
+        self,
+        x: Union[List[torch.Tensor], List[List[torch.Tensor]]],
+        t,
+        cap_feats: Union[List[torch.Tensor], List[List[torch.Tensor]]],
+        return_dict: bool = True,
+        controlnet_block_samples: Optional[Dict[int, torch.Tensor]] = None,
+        siglip_feats: Optional[List[List[torch.Tensor]]] = None,
+        image_noise_mask: Optional[List[List[int]]] = None,
+        patch_size: int = 2,
+        f_patch_size: int = 1,
+    ):
+        """
+        Flow: patchify -> t_embed -> x_embed -> x_refine -> cap_embed -> cap_refine
+              -> [siglip_embed -> siglip_refine] -> build_unified -> main_layers -> final_layer -> unpatchify
+        """
+        assert patch_size in self.all_patch_size and f_patch_size in self.all_f_patch_size
+        omni_mode = isinstance(x[0], list)
+        device = x[0][-1].device if omni_mode else x[0].device
+
+        if omni_mode:
+            # Dual embeddings: noisy (t) and clean (t=1)
+            t_noisy = self.t_embedder(t * self.t_scale).type_as(x[0][-1])
+            t_clean = self.t_embedder(torch.ones_like(t) * self.t_scale).type_as(x[0][-1])
+            adaln_input = None
+        else:
+            # Single embedding for all tokens
+            adaln_input = self.t_embedder(t * self.t_scale).type_as(x[0])
+            t_noisy = t_clean = None
+
+        # Patchify
+        if omni_mode:
+            (
+                x,
+                cap_feats,
+                siglip_feats,
+                x_size,
+                x_pos_ids,
+                cap_pos_ids,
+                siglip_pos_ids,
+                x_pad_mask,
+                cap_pad_mask,
+                siglip_pad_mask,
+                x_pos_offsets,
+                x_noise_mask,
+                cap_noise_mask,
+                siglip_noise_mask,
+            ) = self.patchify_and_embed_omni(x, cap_feats, siglip_feats, patch_size, f_patch_size, image_noise_mask)
+        else:
+            (
+                x,
+                cap_feats,
+                x_size,
+                x_pos_ids,
+                cap_pos_ids,
+                x_pad_mask,
+                cap_pad_mask,
+            ) = self.patchify_and_embed(x, cap_feats, patch_size, f_patch_size)
+            x_pos_offsets = x_noise_mask = cap_noise_mask = siglip_noise_mask = None
+
+        # X embed & refine
+        x_seqlens = [len(xi) for xi in x]
+        x = self.all_x_embedder[f"{patch_size}-{f_patch_size}"](torch.cat(x, dim=0))  # embed
+        x, x_freqs, x_mask, _, x_noise_tensor = self._prepare_sequence(
+            list(x.split(x_seqlens, dim=0)), x_pos_ids, x_pad_mask, self.x_pad_token, x_noise_mask, device
+        )
+
+        for layer in self.noise_refiner:
+            x = (
+                self._gradient_checkpointing_func(
+                    layer, x, x_mask, x_freqs, adaln_input, x_noise_tensor, t_noisy, t_clean
+                )
+                if torch.is_grad_enabled() and self.gradient_checkpointing
+                else layer(x, x_mask, x_freqs, adaln_input, x_noise_tensor, t_noisy, t_clean)
+            )
+
+        # Cap embed & refine
+        cap_seqlens = [len(ci) for ci in cap_feats]
+        cap_feats = self.cap_embedder(torch.cat(cap_feats, dim=0))  # embed
+        cap_feats, cap_freqs, cap_mask, _, _ = self._prepare_sequence(
+            list(cap_feats.split(cap_seqlens, dim=0)), cap_pos_ids, cap_pad_mask, self.cap_pad_token, None, device
+        )
+
+        for layer in self.context_refiner:
+            cap_feats = (
+                self._gradient_checkpointing_func(layer, cap_feats, cap_mask, cap_freqs)
+                if torch.is_grad_enabled() and self.gradient_checkpointing
+                else layer(cap_feats, cap_mask, cap_freqs)
+            )
+
+        # Siglip embed & refine
+        siglip_seqlens = siglip_freqs = None
+        if omni_mode and siglip_feats[0] is not None and self.siglip_embedder is not None:
+            siglip_seqlens = [len(si) for si in siglip_feats]
+            siglip_feats = self.siglip_embedder(torch.cat(siglip_feats, dim=0))  # embed
+            siglip_feats, siglip_freqs, siglip_mask, _, _ = self._prepare_sequence(
+                list(siglip_feats.split(siglip_seqlens, dim=0)),
+                siglip_pos_ids,
+                siglip_pad_mask,
+                self.siglip_pad_token,
+                None,
+                device,
+            )
+
+            for layer in self.siglip_refiner:
+                siglip_feats = (
+                    self._gradient_checkpointing_func(layer, siglip_feats, siglip_mask, siglip_freqs)
+                    if torch.is_grad_enabled() and self.gradient_checkpointing
+                    else layer(siglip_feats, siglip_mask, siglip_freqs)
+                )
+
+        # Unified sequence
+        unified, unified_freqs, unified_mask, unified_noise_tensor = self._build_unified_sequence(
+            x,
+            x_freqs,
+            x_seqlens,
+            x_noise_mask,
+            cap_feats,
+            cap_freqs,
+            cap_seqlens,
+            cap_noise_mask,
+            siglip_feats,
+            siglip_freqs,
+            siglip_seqlens,
+            siglip_noise_mask,
+            omni_mode,
+            device,
+        )
+
+        # Main transformer layers
+        for layer_idx, layer in enumerate(self.layers):
+            unified = (
+                self._gradient_checkpointing_func(
+                    layer, unified, unified_mask, unified_freqs, adaln_input, unified_noise_tensor, t_noisy, t_clean
+                )
+                if torch.is_grad_enabled() and self.gradient_checkpointing
+                else layer(unified, unified_mask, unified_freqs, adaln_input, unified_noise_tensor, t_noisy, t_clean)
+            )
+            if controlnet_block_samples is not None and layer_idx in controlnet_block_samples:
+                unified = unified + controlnet_block_samples[layer_idx]
+
+        unified = (
+            self.all_final_layer[f"{patch_size}-{f_patch_size}"](
+                unified, noise_mask=unified_noise_tensor, c_noisy=t_noisy, c_clean=t_clean
+            )
+            if omni_mode
+            else self.all_final_layer[f"{patch_size}-{f_patch_size}"](unified, c=adaln_input)
+        )
+
+        # Unpatchify
+        x = self.unpatchify(list(unified.unbind(dim=0)), x_size, patch_size, f_patch_size, x_pos_offsets)
+
+        return (x,) if not return_dict else Transformer2DModelOutput(sample=x)

src/diffusers/pipelines/__init__.py

@@ -288,6 +288,7 @@ else:
         "LTXImageToVideoPipeline",
         "LTXConditionPipeline",
         "LTXLatentUpsamplePipeline",
+        "LTXI2VLongMultiPromptPipeline",
     ]
     _import_structure["lumina"] = ["LuminaPipeline", "LuminaText2ImgPipeline"]
     _import_structure["lumina2"] = ["Lumina2Pipeline", "Lumina2Text2ImgPipeline"]
@@ -411,6 +412,7 @@ else:
         "ZImagePipeline",
         "ZImageControlNetPipeline",
         "ZImageControlNetInpaintPipeline",
+        "ZImageOmniPipeline",
     ]
     _import_structure["skyreels_v2"] = [
         "SkyReelsV2DiffusionForcingPipeline",
@@ -728,7 +730,13 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             LEditsPPPipelineStableDiffusionXL,
         )
         from .longcat_image import LongCatImageEditPipeline, LongCatImagePipeline
-        from .ltx import LTXConditionPipeline, LTXImageToVideoPipeline, LTXLatentUpsamplePipeline, LTXPipeline
+        from .ltx import (
+            LTXConditionPipeline,
+            LTXI2VLongMultiPromptPipeline,
+            LTXImageToVideoPipeline,
+            LTXLatentUpsamplePipeline,
+            LTXPipeline,
+        )
         from .lucy import LucyEditPipeline
         from .lumina import LuminaPipeline, LuminaText2ImgPipeline
         from .lumina2 import Lumina2Pipeline, Lumina2Text2ImgPipeline
@@ -856,6 +864,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             ZImageControlNetInpaintPipeline,
             ZImageControlNetPipeline,
             ZImageImg2ImgPipeline,
+            ZImageOmniPipeline,
             ZImagePipeline,
         )
 

src/diffusers/pipelines/auto_pipeline.py

@@ -73,6 +73,7 @@ from .kandinsky3 import Kandinsky3Img2ImgPipeline, Kandinsky3Pipeline
 from .latent_consistency_models import LatentConsistencyModelImg2ImgPipeline, LatentConsistencyModelPipeline
 from .lumina import LuminaPipeline
 from .lumina2 import Lumina2Pipeline
+from .ovis_image import OvisImagePipeline
 from .pag import (
     HunyuanDiTPAGPipeline,
     PixArtSigmaPAGPipeline,
@@ -119,7 +120,13 @@ from .stable_diffusion_xl import (
 )
 from .wan import WanImageToVideoPipeline, WanPipeline, WanVideoToVideoPipeline
 from .wuerstchen import WuerstchenCombinedPipeline, WuerstchenDecoderPipeline
-from .z_image import ZImageImg2ImgPipeline, ZImagePipeline
+from .z_image import (
+    ZImageControlNetInpaintPipeline,
+    ZImageControlNetPipeline,
+    ZImageImg2ImgPipeline,
+    ZImageOmniPipeline,
+    ZImagePipeline,
+)
 
 
 AUTO_TEXT2IMAGE_PIPELINES_MAPPING = OrderedDict(
@@ -164,6 +171,10 @@ AUTO_TEXT2IMAGE_PIPELINES_MAPPING = OrderedDict(
         ("qwenimage", QwenImagePipeline),
         ("qwenimage-controlnet", QwenImageControlNetPipeline),
         ("z-image", ZImagePipeline),
+        ("z-image-controlnet", ZImageControlNetPipeline),
+        ("z-image-controlnet-inpaint", ZImageControlNetInpaintPipeline),
+        ("z-image-omni", ZImageOmniPipeline),
+        ("ovis", OvisImagePipeline),
     ]
 )
 

src/diffusers/pipelines/controlnet_hunyuandit/pipeline_hunyuandit_controlnet.py

@@ -17,7 +17,7 @@ from typing import Callable, Dict, List, Optional, Tuple, Union
 
 import numpy as np
 import torch
-from transformers import BertModel, BertTokenizer, CLIPImageProcessor, MT5Tokenizer, T5EncoderModel
+from transformers import BertModel, BertTokenizer, CLIPImageProcessor, T5EncoderModel, T5Tokenizer
 
 from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput
 
@@ -185,7 +185,7 @@ class HunyuanDiTControlNetPipeline(DiffusionPipeline):
             The HunyuanDiT model designed by Tencent Hunyuan.
         text_encoder_2 (`T5EncoderModel`):
             The mT5 embedder. Specifically, it is 't5-v1_1-xxl'.
-        tokenizer_2 (`MT5Tokenizer`):
+        tokenizer_2 (`T5Tokenizer`):
             The tokenizer for the mT5 embedder.
         scheduler ([`DDPMScheduler`]):
             A scheduler to be used in combination with HunyuanDiT to denoise the encoded image latents.
@@ -229,7 +229,7 @@ class HunyuanDiTControlNetPipeline(DiffusionPipeline):
             HunyuanDiT2DMultiControlNetModel,
         ],
         text_encoder_2: Optional[T5EncoderModel] = None,
-        tokenizer_2: Optional[MT5Tokenizer] = None,
+        tokenizer_2: Optional[T5Tokenizer] = None,
         requires_safety_checker: bool = True,
     ):
         super().__init__()

src/diffusers/pipelines/cosmos/pipeline_cosmos2_5_predict.py

@@ -76,7 +76,7 @@ EXAMPLE_DOC_STRING = """
 
         >>> model_id = "nvidia/Cosmos-Predict2.5-2B"
         >>> pipe = Cosmos2_5_PredictBasePipeline.from_pretrained(
-        ...     model_id, revision="diffusers/base/pre-trianed", torch_dtype=torch.bfloat16
+        ...     model_id, revision="diffusers/base/post-trained", torch_dtype=torch.bfloat16
         ... )
         >>> pipe = pipe.to("cuda")
 
@@ -133,7 +133,7 @@ EXAMPLE_DOC_STRING = """
         ...     num_frames=93,
         ...     generator=torch.Generator().manual_seed(1),
         ... ).frames[0]
-        >>> # export_to_video(video, "image2world.mp4", fps=16)
+        >>> export_to_video(video, "image2world.mp4", fps=16)
 
         >>> # Video2World: condition on an input clip and predict a 93-frame world video.
         >>> prompt = (

src/diffusers/pipelines/hunyuandit/pipeline_hunyuandit.py

@@ -17,7 +17,7 @@ from typing import Callable, Dict, List, Optional, Tuple, Union
 
 import numpy as np
 import torch
-from transformers import BertModel, BertTokenizer, CLIPImageProcessor, MT5Tokenizer, T5EncoderModel
+from transformers import BertModel, BertTokenizer, CLIPImageProcessor, T5EncoderModel, T5Tokenizer
 
 from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput
 
@@ -169,7 +169,7 @@ class HunyuanDiTPipeline(DiffusionPipeline):
             The HunyuanDiT model designed by Tencent Hunyuan.
         text_encoder_2 (`T5EncoderModel`):
             The mT5 embedder. Specifically, it is 't5-v1_1-xxl'.
-        tokenizer_2 (`MT5Tokenizer`):
+        tokenizer_2 (`T5Tokenizer`):
             The tokenizer for the mT5 embedder.
         scheduler ([`DDPMScheduler`]):
             A scheduler to be used in combination with HunyuanDiT to denoise the encoded image latents.
@@ -204,7 +204,7 @@ class HunyuanDiTPipeline(DiffusionPipeline):
         feature_extractor: CLIPImageProcessor,
         requires_safety_checker: bool = True,
         text_encoder_2: Optional[T5EncoderModel] = None,
-        tokenizer_2: Optional[MT5Tokenizer] = None,
+        tokenizer_2: Optional[T5Tokenizer] = None,
     ):
         super().__init__()
 

src/diffusers/pipelines/ltx/__init__.py

@@ -25,6 +25,7 @@ else:
     _import_structure["modeling_latent_upsampler"] = ["LTXLatentUpsamplerModel"]
     _import_structure["pipeline_ltx"] = ["LTXPipeline"]
     _import_structure["pipeline_ltx_condition"] = ["LTXConditionPipeline"]
+    _import_structure["pipeline_ltx_i2v_long_multi_prompt"] = ["LTXI2VLongMultiPromptPipeline"]
     _import_structure["pipeline_ltx_image2video"] = ["LTXImageToVideoPipeline"]
     _import_structure["pipeline_ltx_latent_upsample"] = ["LTXLatentUpsamplePipeline"]
 
@@ -39,6 +40,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
         from .modeling_latent_upsampler import LTXLatentUpsamplerModel
         from .pipeline_ltx import LTXPipeline
         from .pipeline_ltx_condition import LTXConditionPipeline
+        from .pipeline_ltx_i2v_long_multi_prompt import LTXI2VLongMultiPromptPipeline
         from .pipeline_ltx_image2video import LTXImageToVideoPipeline
         from .pipeline_ltx_latent_upsample import LTXLatentUpsamplePipeline
 

src/diffusers/pipelines/pag/pipeline_pag_hunyuandit.py

@@ -17,7 +17,7 @@ from typing import Callable, Dict, List, Optional, Tuple, Union
 
 import numpy as np
 import torch
-from transformers import BertModel, BertTokenizer, CLIPImageProcessor, MT5Tokenizer, T5EncoderModel
+from transformers import BertModel, BertTokenizer, CLIPImageProcessor, T5EncoderModel, T5Tokenizer
 
 from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput
 
@@ -173,7 +173,7 @@ class HunyuanDiTPAGPipeline(DiffusionPipeline, PAGMixin):
             The HunyuanDiT model designed by Tencent Hunyuan.
         text_encoder_2 (`T5EncoderModel`):
             The mT5 embedder. Specifically, it is 't5-v1_1-xxl'.
-        tokenizer_2 (`MT5Tokenizer`):
+        tokenizer_2 (`T5Tokenizer`):
             The tokenizer for the mT5 embedder.
         scheduler ([`DDPMScheduler`]):
             A scheduler to be used in combination with HunyuanDiT to denoise the encoded image latents.
@@ -208,7 +208,7 @@ class HunyuanDiTPAGPipeline(DiffusionPipeline, PAGMixin):
         feature_extractor: Optional[CLIPImageProcessor] = None,
         requires_safety_checker: bool = True,
         text_encoder_2: Optional[T5EncoderModel] = None,
-        tokenizer_2: Optional[MT5Tokenizer] = None,
+        tokenizer_2: Optional[T5Tokenizer] = None,
         pag_applied_layers: Union[str, List[str]] = "blocks.1",  # "blocks.16.attn1", "blocks.16", "16", 16
     ):
         super().__init__()

src/diffusers/pipelines/z_image/__init__.py

@@ -26,6 +26,7 @@ else:
     _import_structure["pipeline_z_image_controlnet"] = ["ZImageControlNetPipeline"]
     _import_structure["pipeline_z_image_controlnet_inpaint"] = ["ZImageControlNetInpaintPipeline"]
     _import_structure["pipeline_z_image_img2img"] = ["ZImageImg2ImgPipeline"]
+    _import_structure["pipeline_z_image_omni"] = ["ZImageOmniPipeline"]
 
 
 if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
@@ -41,7 +42,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
         from .pipeline_z_image_controlnet import ZImageControlNetPipeline
         from .pipeline_z_image_controlnet_inpaint import ZImageControlNetInpaintPipeline
         from .pipeline_z_image_img2img import ZImageImg2ImgPipeline
-
+        from .pipeline_z_image_omni import ZImageOmniPipeline
 else:
     import sys
 

src/diffusers/pipelines/z_image/pipeline_z_image_controlnet.py

@@ -58,14 +58,13 @@ EXAMPLE_DOC_STRING = """
         >>> #     torch_dtype=torch.bfloat16,
         >>> # )
 
-        >>> # 2.0 - `config` is required
+        >>> # 2.0
         >>> # controlnet = ZImageControlNetModel.from_single_file(
         >>> #     hf_hub_download(
         >>> #         "alibaba-pai/Z-Image-Turbo-Fun-Controlnet-Union-2.0",
         >>> #         filename="Z-Image-Turbo-Fun-Controlnet-Union-2.0.safetensors",
         >>> #     ),
         >>> #     torch_dtype=torch.bfloat16,
-        >>> #     config="hlky/Z-Image-Turbo-Fun-Controlnet-Union-2.0",
         >>> # )
 
         >>> pipe = ZImageControlNetPipeline.from_pretrained(

src/diffusers/pipelines/z_image/pipeline_z_image_controlnet_inpaint.py

@@ -50,14 +50,13 @@ EXAMPLE_DOC_STRING = """
         ...     torch_dtype=torch.bfloat16,
         ... )
 
-        >>> # 2.0 - `config` is required
+        >>> # 2.0
         >>> # controlnet = ZImageControlNetModel.from_single_file(
         >>> #     hf_hub_download(
         >>> #         "alibaba-pai/Z-Image-Turbo-Fun-Controlnet-Union-2.0",
         >>> #         filename="Z-Image-Turbo-Fun-Controlnet-Union-2.0.safetensors",
         >>> #     ),
         >>> #     torch_dtype=torch.bfloat16,
-        >>> #     config="hlky/Z-Image-Turbo-Fun-Controlnet-Union-2.0",
         >>> # )
 
         >>> pipe = ZImageControlNetInpaintPipeline.from_pretrained(

src/diffusers/pipelines/z_image/pipeline_z_image_omni.py

@@ -0,0 +1,742 @@
+# 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 PIL
+import torch
+from transformers import AutoTokenizer, PreTrainedModel, Siglip2ImageProcessorFast, Siglip2VisionModel
+
+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 ..flux2.image_processor import Flux2ImageProcessor
+from .pipeline_output import ZImagePipelineOutput
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+    Examples:
+        ```py
+        >>> import torch
+        >>> from diffusers import ZImageOmniPipeline
+
+        >>> pipe = ZImageOmniPipeline.from_pretrained("Z-a-o/Z-Image-Turbo", torch_dtype=torch.bfloat16)
+        >>> pipe.to("cuda")
+
+        >>> # Optionally, set the attention backend to flash-attn 2 or 3, default is SDPA in PyTorch.
+        >>> # (1) Use flash attention 2
+        >>> # pipe.transformer.set_attention_backend("flash")
+        >>> # (2) Use flash attention 3
+        >>> # pipe.transformer.set_attention_backend("_flash_3")
+
+        >>> prompt = "一幅为名为“造相「Z-IMAGE-TURBO」”的项目设计的创意海报。画面巧妙地将文字概念视觉化：一辆复古蒸汽小火车化身为巨大的拉链头，正拉开厚厚的冬日积雪，展露出一个生机盎然的春天。"
+        >>> image = pipe(
+        ...     prompt,
+        ...     height=1024,
+        ...     width=1024,
+        ...     num_inference_steps=9,
+        ...     guidance_scale=0.0,
+        ...     generator=torch.Generator("cuda").manual_seed(42),
+        ... ).images[0]
+        >>> image.save("zimage.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.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 ZImageOmniPipeline(DiffusionPipeline, ZImageLoraLoaderMixin, FromSingleFileMixin):
+    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,
+        siglip: Siglip2VisionModel,
+        siglip_processor: Siglip2ImageProcessorFast,
+    ):
+        super().__init__()
+
+        self.register_modules(
+            vae=vae,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+            scheduler=scheduler,
+            transformer=transformer,
+            siglip=siglip,
+            siglip_processor=siglip_processor,
+        )
+        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.image_processor = Flux2ImageProcessor(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,
+        num_condition_images: int = 0,
+    ):
+        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,
+            num_condition_images=num_condition_images,
+        )
+
+        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,
+                num_condition_images=num_condition_images,
+            )
+        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,
+        num_condition_images: int = 0,
+    ) -> 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):
+            if num_condition_images == 0:
+                prompt[i] = ["<|im_start|>user\n" + prompt_item + "<|im_end|>\n<|im_start|>assistant\n"]
+            elif num_condition_images > 0:
+                prompt_list = ["<|im_start|>user\n<|vision_start|>"]
+                prompt_list += ["<|vision_end|><|vision_start|>"] * (num_condition_images - 1)
+                prompt_list += ["<|vision_end|>" + prompt_item + "<|im_end|>\n<|im_start|>assistant\n<|vision_start|>"]
+                prompt_list += ["<|vision_end|><|im_end|>"]
+                prompt[i] = prompt_list
+
+        flattened_prompt = []
+        prompt_list_lengths = []
+
+        for i in range(len(prompt)):
+            prompt_list_lengths.append(len(prompt[i]))
+            flattened_prompt.extend(prompt[i])
+
+        text_inputs = self.tokenizer(
+            flattened_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 = []
+        start_idx = 0
+        for i in range(len(prompt_list_lengths)):
+            batch_embeddings = []
+            end_idx = start_idx + prompt_list_lengths[i]
+            for j in range(start_idx, end_idx):
+                batch_embeddings.append(prompt_embeds[j][prompt_masks[j]])
+            embeddings_list.append(batch_embeddings)
+            start_idx = end_idx
+
+        return embeddings_list
+
+    def prepare_latents(
+        self,
+        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 None:
+            latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+        else:
+            if latents.shape != shape:
+                raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}")
+            latents = latents.to(device)
+        return latents
+
+    def prepare_image_latents(
+        self,
+        images: List[torch.Tensor],
+        batch_size,
+        device,
+        dtype,
+    ):
+        image_latents = []
+        for image in images:
+            image = image.to(device=device, dtype=dtype)
+            image_latent = (
+                self.vae.encode(image.bfloat16()).latent_dist.mode()[0] - self.vae.config.shift_factor
+            ) * self.vae.config.scaling_factor
+            image_latent = image_latent.unsqueeze(1).to(dtype)
+            image_latents.append(image_latent)  # (16, 128, 128)
+
+        # image_latents = [image_latents] * batch_size
+        image_latents = [image_latents.copy() for _ in range(batch_size)]
+
+        return image_latents
+
+    def prepare_siglip_embeds(
+        self,
+        images: List[torch.Tensor],
+        batch_size,
+        device,
+        dtype,
+    ):
+        siglip_embeds = []
+        for image in images:
+            siglip_inputs = self.siglip_processor(images=[image], return_tensors="pt").to(device)
+            shape = siglip_inputs.spatial_shapes[0]
+            hidden_state = self.siglip(**siglip_inputs).last_hidden_state
+            B, N, C = hidden_state.shape
+            hidden_state = hidden_state[:, : shape[0] * shape[1]]
+            hidden_state = hidden_state.view(shape[0], shape[1], C)
+            siglip_embeds.append(hidden_state.to(dtype))
+
+        # siglip_embeds = [siglip_embeds] * batch_size
+        siglip_embeds = [siglip_embeds.copy() for _ in range(batch_size)]
+
+        return siglip_embeds
+
+    @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,
+        image: Optional[Union[List[PIL.Image.Image], PIL.Image.Image]] = None,
+        prompt: Union[str, List[str]] = None,
+        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 generation.
+
+        Args:
+            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.
+            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.
+            height (`int`, *optional*, defaults to 1024):
+                The height in pixels of the generated image.
+            width (`int`, *optional*, defaults to 1024):
+                The width in pixels of the generated 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 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.
+        """
+
+        if image is not None and not isinstance(image, list):
+            image = [image]
+        num_condition_images = len(image) if image is not None else 0
+
+        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
+
+        # 2. Define call parameters
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = 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,
+                num_condition_images=num_condition_images,
+            )
+
+        # 3. Process condition images. Copied from diffusers.pipelines.flux2.pipeline_flux2
+        condition_images = []
+        resized_images = []
+        if image is not None:
+            for img in image:
+                self.image_processor.check_image_input(img)
+            for img in image:
+                image_width, image_height = img.size
+                if image_width * image_height > 1024 * 1024:
+                    if height is not None and width is not None:
+                        img = self.image_processor._resize_to_target_area(img, height * width)
+                    else:
+                        img = self.image_processor._resize_to_target_area(img, 1024 * 1024)
+                    image_width, image_height = img.size
+                resized_images.append(img)
+
+                multiple_of = self.vae_scale_factor * 2
+                image_width = (image_width // multiple_of) * multiple_of
+                image_height = (image_height // multiple_of) * multiple_of
+                img = self.image_processor.preprocess(img, height=image_height, width=image_width, resize_mode="crop")
+                condition_images.append(img)
+
+            if len(condition_images) > 0:
+                height = height or image_height
+                width = width or image_width
+
+        else:
+            height = height or 1024
+            width = width or 1024
+
+        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}."
+            )
+
+        # 4. Prepare latent variables
+        num_channels_latents = self.transformer.in_channels
+
+        latents = self.prepare_latents(
+            batch_size * num_images_per_prompt,
+            num_channels_latents,
+            height,
+            width,
+            torch.float32,
+            device,
+            generator,
+            latents,
+        )
+
+        condition_latents = self.prepare_image_latents(
+            images=condition_images,
+            batch_size=batch_size * num_images_per_prompt,
+            device=device,
+            dtype=torch.float32,
+        )
+        condition_latents = [[lat.to(self.transformer.dtype) for lat in lats] for lats in condition_latents]
+        if self.do_classifier_free_guidance:
+            negative_condition_latents = [[lat.clone() for lat in batch] for batch in condition_latents]
+
+        condition_siglip_embeds = self.prepare_siglip_embeds(
+            images=resized_images,
+            batch_size=batch_size * num_images_per_prompt,
+            device=device,
+            dtype=torch.float32,
+        )
+        condition_siglip_embeds = [[se.to(self.transformer.dtype) for se in sels] for sels in condition_siglip_embeds]
+        if self.do_classifier_free_guidance:
+            negative_condition_siglip_embeds = [[se.clone() for se in batch] for batch in condition_siglip_embeds]
+
+        # 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)]
+
+        condition_siglip_embeds = [None if sels == [] else sels + [None] for sels in condition_siglip_embeds]
+        negative_condition_siglip_embeds = [
+            None if sels == [] else sels + [None] for sels in negative_condition_siglip_embeds
+        ]
+
+        actual_batch_size = batch_size * num_images_per_prompt
+        image_seq_len = (latents.shape[2] // 2) * (latents.shape[3] // 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,
+        )
+        num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
+        self._num_timesteps = len(timesteps)
+
+        # 6. 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
+                    condition_latents_model_input = condition_latents + negative_condition_latents
+                    condition_siglip_embeds_model_input = condition_siglip_embeds + negative_condition_siglip_embeds
+                    timestep_model_input = timestep.repeat(2)
+                else:
+                    latent_model_input = latents.to(self.transformer.dtype)
+                    prompt_embeds_model_input = prompt_embeds
+                    condition_latents_model_input = condition_latents
+                    condition_siglip_embeds_model_input = condition_siglip_embeds
+                    timestep_model_input = timestep
+
+                latent_model_input = latent_model_input.unsqueeze(2)
+                latent_model_input_list = list(latent_model_input.unbind(dim=0))
+
+                # Combine condition latents with target latent
+                current_batch_size = len(latent_model_input_list)
+                x_combined = [
+                    condition_latents_model_input[i] + [latent_model_input_list[i]] for i in range(current_batch_size)
+                ]
+                # Create noise mask: 0 for condition images (clean), 1 for target image (noisy)
+                image_noise_mask = [
+                    [0] * len(condition_latents_model_input[i]) + [1] for i in range(current_batch_size)
+                ]
+
+                model_out_list = self.transformer(
+                    x=x_combined,
+                    t=timestep_model_input,
+                    cap_feats=prompt_embeds_model_input,
+                    siglip_feats=condition_siglip_embeds_model_input,
+                    image_noise_mask=image_noise_mask,
+                    return_dict=False,
+                )[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)

src/diffusers/quantizers/torchao/torchao_quantizer.py

@@ -36,6 +36,9 @@ from ...utils import (
 from ..base import DiffusersQuantizer
 
 
+logger = logging.get_logger(__name__)
+
+
 if TYPE_CHECKING:
     from ...models.modeling_utils import ModelMixin
 
@@ -83,11 +86,19 @@ def _update_torch_safe_globals():
     ]
     try:
         from torchao.dtypes import NF4Tensor
-        from torchao.dtypes.floatx.float8_layout import Float8AQTTensorImpl
-        from torchao.dtypes.uintx.uint4_layout import UInt4Tensor
         from torchao.dtypes.uintx.uintx_layout import UintxAQTTensorImpl, UintxTensor
 
-        safe_globals.extend([UintxTensor, UInt4Tensor, UintxAQTTensorImpl, Float8AQTTensorImpl, NF4Tensor])
+        safe_globals.extend([UintxTensor, UintxAQTTensorImpl, NF4Tensor])
+
+        # note: is_torchao_version(">=", "0.16.0") does not work correctly
+        # with torchao nightly, so using a ">" check which does work correctly
+        if is_torchao_version(">", "0.15.0"):
+            pass
+        else:
+            from torchao.dtypes.floatx.float8_layout import Float8AQTTensorImpl
+            from torchao.dtypes.uintx.uint4_layout import UInt4Tensor
+
+            safe_globals.extend([UInt4Tensor, Float8AQTTensorImpl])
 
     except (ImportError, ModuleNotFoundError) as e:
         logger.warning(
@@ -123,9 +134,6 @@ def fuzzy_match_size(config_name: str) -> Optional[str]:
     return None
 
 
-logger = logging.get_logger(__name__)
-
-
 def _quantization_type(weight):
     from torchao.dtypes import AffineQuantizedTensor
     from torchao.quantization.linear_activation_quantized_tensor import LinearActivationQuantizedTensor

src/diffusers/schedulers/__init__.py

@@ -66,6 +66,7 @@ else:
     _import_structure["scheduling_k_dpm_2_ancestral_discrete"] = ["KDPM2AncestralDiscreteScheduler"]
     _import_structure["scheduling_k_dpm_2_discrete"] = ["KDPM2DiscreteScheduler"]
     _import_structure["scheduling_lcm"] = ["LCMScheduler"]
+    _import_structure["scheduling_ltx_euler_ancestral_rf"] = ["LTXEulerAncestralRFScheduler"]
     _import_structure["scheduling_pndm"] = ["PNDMScheduler"]
     _import_structure["scheduling_repaint"] = ["RePaintScheduler"]
     _import_structure["scheduling_sasolver"] = ["SASolverScheduler"]
@@ -168,6 +169,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
         from .scheduling_k_dpm_2_ancestral_discrete import KDPM2AncestralDiscreteScheduler
         from .scheduling_k_dpm_2_discrete import KDPM2DiscreteScheduler
         from .scheduling_lcm import LCMScheduler
+        from .scheduling_ltx_euler_ancestral_rf import LTXEulerAncestralRFScheduler
         from .scheduling_pndm import PNDMScheduler
         from .scheduling_repaint import RePaintScheduler
         from .scheduling_sasolver import SASolverScheduler

src/diffusers/schedulers/scheduling_ltx_euler_ancestral_rf.py

@@ -0,0 +1,386 @@
+# Copyright 2025 Lightricks 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.
+
+"""
+LTXEulerAncestralRFScheduler
+
+This scheduler implements a K-diffusion style Euler-Ancestral sampler specialized for flow / CONST parameterization,
+closely mirroring ComfyUI's `sample_euler_ancestral_RF` implementation used for LTX-Video.
+
+Reference implementation (ComfyUI):
+    comfy.k_diffusion.sampling.sample_euler_ancestral_RF
+"""
+
+from dataclasses import dataclass
+from typing import List, Optional, Tuple, Union
+
+import torch
+
+from ..configuration_utils import ConfigMixin, register_to_config
+from ..utils import BaseOutput, logging
+from ..utils.torch_utils import randn_tensor
+from .scheduling_utils import SchedulerMixin
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+@dataclass
+class LTXEulerAncestralRFSchedulerOutput(BaseOutput):
+    """
+    Output class for the scheduler's `step` function output.
+
+    Args:
+        prev_sample (`torch.FloatTensor`):
+            Updated sample for the next step in the denoising process.
+    """
+
+    prev_sample: torch.FloatTensor
+
+
+class LTXEulerAncestralRFScheduler(SchedulerMixin, ConfigMixin):
+    """
+    Euler-Ancestral scheduler for LTX-Video (RF / CONST parametrization).
+
+    This scheduler is intended for models where the network is trained with a CONST-like parameterization (as in LTXV /
+    FLUX). It approximates ComfyUI's `sample_euler_ancestral_RF` sampler and is useful when reproducing ComfyUI
+    workflows inside diffusers.
+
+    The scheduler can either:
+    - reuse the [`FlowMatchEulerDiscreteScheduler`] sigma / timestep logic when only `num_inference_steps` is provided
+      (default diffusers-style usage), or
+    - follow an explicit ComfyUI-style sigma schedule when `sigmas` (or `timesteps`) are passed to [`set_timesteps`].
+
+    Args:
+        num_train_timesteps (`int`, defaults to 1000):
+            Included for config compatibility; not used to build the schedule.
+        eta (`float`, defaults to 1.0):
+            Stochasticity parameter. `eta=0.0` yields deterministic DDIM-like sampling; `eta=1.0` matches ComfyUI's
+            default RF behavior.
+        s_noise (`float`, defaults to 1.0):
+            Global scaling factor for the stochastic noise term.
+    """
+
+    # Allow config migration from the flow-match scheduler and back.
+    _compatibles = ["FlowMatchEulerDiscreteScheduler"]
+    order = 1
+
+    @register_to_config
+    def __init__(
+        self,
+        num_train_timesteps: int = 1000,
+        eta: float = 1.0,
+        s_noise: float = 1.0,
+    ):
+        # Note: num_train_timesteps is kept only for config compatibility.
+        self.num_inference_steps: Optional[int] = None
+        self.sigmas: Optional[torch.Tensor] = None
+        self.timesteps: Optional[torch.Tensor] = None
+        self._step_index: Optional[int] = None
+        self._begin_index: Optional[int] = None
+
+    @property
+    def step_index(self) -> Optional[int]:
+        return self._step_index
+
+    @property
+    def begin_index(self) -> Optional[int]:
+        """
+        The index for the first timestep. It can be set from a pipeline with `set_begin_index` to support
+        image-to-image like workflows that start denoising part-way through the schedule.
+        """
+        return self._begin_index
+
+    def set_begin_index(self, begin_index: int = 0):
+        """
+        Included for API compatibility; not strictly needed here but kept to allow pipelines that call
+        `set_begin_index`.
+        """
+        self._begin_index = begin_index
+
+    def index_for_timestep(
+        self, timestep: Union[float, torch.Tensor], schedule_timesteps: Optional[torch.Tensor] = None
+    ) -> int:
+        """
+        Map a (continuous) `timestep` value to an index into `self.timesteps`.
+
+        This follows the convention used in other discrete schedulers: if the same timestep value appears multiple
+        times in the schedule (which can happen when starting in the middle of the schedule), the *second* occurrence
+        is used for the first `step` call so that no sigma is accidentally skipped.
+        """
+        if schedule_timesteps is None:
+            if self.timesteps is None:
+                raise ValueError("Timesteps have not been set. Call `set_timesteps` first.")
+            schedule_timesteps = self.timesteps
+
+        if isinstance(timestep, torch.Tensor):
+            timestep = timestep.to(schedule_timesteps.device)
+
+        indices = (schedule_timesteps == timestep).nonzero()
+
+        # The sigma index that is taken for the **very** first `step`
+        # is always the second index (or the last index if there is only 1)
+        # This way we can ensure we don't accidentally skip a sigma in
+        # case we start in the middle of the denoising schedule (e.g. for image-to-image)
+        pos = 1 if len(indices) > 1 else 0
+
+        if len(indices) == 0:
+            raise ValueError(
+                "Passed `timestep` is not in `self.timesteps`. Make sure to use values from `scheduler.timesteps`."
+            )
+
+        return indices[pos].item()
+
+    def _init_step_index(self, timestep: Union[float, torch.Tensor]):
+        """
+        Initialize the internal step index based on a given timestep.
+        """
+        if self.timesteps is None:
+            raise ValueError("Timesteps have not been set. Call `set_timesteps` first.")
+
+        if self.begin_index is None:
+            if isinstance(timestep, torch.Tensor):
+                timestep = timestep.to(self.timesteps.device)
+            self._step_index = self.index_for_timestep(timestep)
+        else:
+            self._step_index = self._begin_index
+
+    def set_timesteps(
+        self,
+        num_inference_steps: Optional[int] = None,
+        device: Union[str, torch.device, None] = None,
+        sigmas: Optional[Union[List[float], torch.Tensor]] = None,
+        timesteps: Optional[Union[List[float], torch.Tensor]] = None,
+        mu: Optional[float] = None,
+        **kwargs,
+    ):
+        """
+        Set the sigma / timestep schedule for sampling.
+
+        When `sigmas` or `timesteps` are provided explicitly, they are used as the RF sigma schedule (ComfyUI-style)
+        and are expected to include the terminal 0.0. When both are `None`, the scheduler reuses the
+        [`FlowMatchEulerDiscreteScheduler`] logic to generate sigmas from `num_inference_steps` and the stored config
+        (including any resolution-dependent shifting, Karras/beta schedules, etc.).
+
+        Args:
+            num_inference_steps (`int`, *optional*):
+                Number of denoising steps. If provided together with explicit `sigmas`/`timesteps`, they are expected
+                to be consistent and are otherwise ignored with a warning.
+            device (`str` or `torch.device`, *optional*):
+                Device to move the internal tensors to.
+            sigmas (`List[float]` or `torch.Tensor`, *optional*):
+                Explicit sigma schedule, e.g. `[1.0, 0.99, ..., 0.0]`.
+            timesteps (`List[float]` or `torch.Tensor`, *optional*):
+                Optional alias for `sigmas`. If `sigmas` is None and `timesteps` is provided, timesteps are treated as
+                sigmas.
+            mu (`float`, *optional*):
+                Optional shift parameter used when delegating to [`FlowMatchEulerDiscreteScheduler.set_timesteps`] and
+                `config.use_dynamic_shifting` is `True`.
+        """
+        # 1. Auto-generate schedule (FlowMatch-style) when no explicit sigmas/timesteps are given
+        if sigmas is None and timesteps is None:
+            if num_inference_steps is None:
+                raise ValueError(
+                    "LTXEulerAncestralRFScheduler.set_timesteps requires either explicit `sigmas`/`timesteps` "
+                    "or a `num_inference_steps` value."
+                )
+
+            # We reuse FlowMatchEulerDiscreteScheduler to construct a sigma schedule that is
+            # consistent with the original LTX training setup (including optional time shifting,
+            # Karras / exponential / beta schedules, etc.).
+            from .scheduling_flow_match_euler_discrete import FlowMatchEulerDiscreteScheduler
+
+            base_scheduler = FlowMatchEulerDiscreteScheduler.from_config(self.config)
+            base_scheduler.set_timesteps(
+                num_inference_steps=num_inference_steps,
+                device=device,
+                sigmas=None,
+                mu=mu,
+                timesteps=None,
+            )
+
+            self.num_inference_steps = base_scheduler.num_inference_steps
+            # Keep sigmas / timesteps on the requested device so step() can operate on-device without
+            # extra transfers.
+            self.sigmas = base_scheduler.sigmas.to(device=device)
+            self.timesteps = base_scheduler.timesteps.to(device=device)
+            self._step_index = None
+            self._begin_index = None
+            return
+
+        # 2. Explicit sigma schedule (ComfyUI-style path)
+        if sigmas is None:
+            # `timesteps` is treated as sigmas in RF / flow-matching setups.
+            sigmas = timesteps
+
+        if isinstance(sigmas, list):
+            sigmas_tensor = torch.tensor(sigmas, dtype=torch.float32)
+        elif isinstance(sigmas, torch.Tensor):
+            sigmas_tensor = sigmas.to(dtype=torch.float32)
+        else:
+            raise TypeError(f"`sigmas` must be a list or torch.Tensor, got {type(sigmas)}.")
+
+        if sigmas_tensor.ndim != 1:
+            raise ValueError(f"`sigmas` must be a 1D tensor, got shape {tuple(sigmas_tensor.shape)}.")
+
+        if sigmas_tensor[-1].abs().item() > 1e-6:
+            logger.warning(
+                "The last sigma in the schedule is not zero (%.6f). "
+                "For best compatibility with ComfyUI's RF sampler, the terminal sigma "
+                "should be 0.0.",
+                sigmas_tensor[-1].item(),
+            )
+
+        # Move to device once, then derive timesteps.
+        if device is not None:
+            sigmas_tensor = sigmas_tensor.to(device)
+
+        # Internal sigma schedule stays in [0, 1] (as provided).
+        self.sigmas = sigmas_tensor
+        # Timesteps are scaled to match the training setup of LTX (FlowMatch-style),
+        # where the network expects timesteps on [0, num_train_timesteps].
+        # This keeps the transformer conditioning in the expected range while the RF
+        # scheduler still operates on the raw sigma values.
+        num_train = float(getattr(self.config, "num_train_timesteps", 1000))
+        self.timesteps = sigmas_tensor * num_train
+
+        if num_inference_steps is not None and num_inference_steps != len(sigmas) - 1:
+            logger.warning(
+                "Provided `num_inference_steps=%d` does not match `len(sigmas)-1=%d`. "
+                "Overriding `num_inference_steps` with `len(sigmas)-1`.",
+                num_inference_steps,
+                len(sigmas) - 1,
+            )
+
+        self.num_inference_steps = len(sigmas) - 1
+        self._step_index = None
+        self._begin_index = None
+
+    def _sigma_broadcast(self, sigma: torch.Tensor, sample: torch.Tensor) -> torch.Tensor:
+        """
+        Helper to broadcast a scalar sigma to the shape of `sample`.
+        """
+        while sigma.ndim < sample.ndim:
+            sigma = sigma.view(*sigma.shape, 1)
+        return sigma
+
+    def step(
+        self,
+        model_output: torch.FloatTensor,
+        timestep: Union[float, torch.Tensor],
+        sample: torch.FloatTensor,
+        generator: Optional[torch.Generator] = None,
+        return_dict: bool = True,
+    ) -> Union[LTXEulerAncestralRFSchedulerOutput, Tuple[torch.FloatTensor]]:
+        """
+        Perform a single Euler-Ancestral RF update step.
+
+        Args:
+            model_output (`torch.FloatTensor`):
+                Raw model output at the current step. Interpreted under the CONST parametrization as `v_t`, with
+                denoised state reconstructed as `x0 = x_t - sigma_t * v_t`.
+            timestep (`float` or `torch.Tensor`):
+                The current sigma value (must match one entry in `self.timesteps`).
+            sample (`torch.FloatTensor`):
+                Current latent sample `x_t`.
+            generator (`torch.Generator`, *optional*):
+                Optional generator for reproducible noise.
+            return_dict (`bool`):
+                If `True`, return a `LTXEulerAncestralRFSchedulerOutput`; otherwise return a tuple where the first
+                element is the updated sample.
+        """
+
+        if isinstance(timestep, (int, torch.IntTensor, torch.LongTensor)):
+            raise ValueError(
+                (
+                    "Passing integer indices (e.g. from `enumerate(timesteps)`) as timesteps to"
+                    " `LTXEulerAncestralRFScheduler.step()` is not supported. Make sure to pass"
+                    " one of the `scheduler.timesteps` values as `timestep`."
+                ),
+            )
+
+        if self.sigmas is None or self.timesteps is None:
+            raise ValueError("Scheduler has not been initialized. Call `set_timesteps` before `step`.")
+
+        if self._step_index is None:
+            self._init_step_index(timestep)
+
+        i = self._step_index
+        if i >= len(self.sigmas) - 1:
+            # Already at the end; simply return the current sample.
+            prev_sample = sample
+        else:
+            # Work in float32 for numerical stability
+            sample_f = sample.to(torch.float32)
+            model_output_f = model_output.to(torch.float32)
+
+            sigma = self.sigmas[i]
+            sigma_next = self.sigmas[i + 1]
+
+            sigma_b = self._sigma_broadcast(sigma.view(1), sample_f)
+            sigma_next_b = self._sigma_broadcast(sigma_next.view(1), sample_f)
+
+            # Approximate denoised x0 under CONST parametrization:
+            #   x0 = x_t - sigma_t * v_t
+            denoised = sample_f - sigma_b * model_output_f
+
+            if sigma_next.abs().item() < 1e-8:
+                # Final denoising step
+                x = denoised
+            else:
+                eta = float(self.config.eta)
+                s_noise = float(self.config.s_noise)
+
+                # Downstep computation (ComfyUI RF variant)
+                downstep_ratio = 1.0 + (sigma_next / sigma - 1.0) * eta
+                sigma_down = sigma_next * downstep_ratio
+
+                alpha_ip1 = 1.0 - sigma_next
+                alpha_down = 1.0 - sigma_down
+
+                # Deterministic part (Euler step in (x, x0)-space)
+                sigma_down_b = self._sigma_broadcast(sigma_down.view(1), sample_f)
+                alpha_ip1_b = self._sigma_broadcast(alpha_ip1.view(1), sample_f)
+                alpha_down_b = self._sigma_broadcast(alpha_down.view(1), sample_f)
+
+                sigma_ratio = sigma_down_b / sigma_b
+                x = sigma_ratio * sample_f + (1.0 - sigma_ratio) * denoised
+
+                # Stochastic ancestral noise
+                if eta > 0.0 and s_noise > 0.0:
+                    renoise_coeff = (
+                        (sigma_next_b**2 - sigma_down_b**2 * alpha_ip1_b**2 / (alpha_down_b**2 + 1e-12))
+                        .clamp(min=0.0)
+                        .sqrt()
+                    )
+
+                    noise = randn_tensor(
+                        sample_f.shape, generator=generator, device=sample_f.device, dtype=sample_f.dtype
+                    )
+                    x = (alpha_ip1_b / (alpha_down_b + 1e-12)) * x + noise * renoise_coeff * s_noise
+
+            prev_sample = x.to(sample.dtype)
+
+        # Advance internal step index
+        self._step_index = min(self._step_index + 1, len(self.sigmas) - 1)
+
+        if not return_dict:
+            return (prev_sample,)
+
+        return LTXEulerAncestralRFSchedulerOutput(prev_sample=prev_sample)
+
+    def __len__(self) -> int:
+        # For compatibility with other schedulers; used e.g. in some training
+        # utilities to infer the maximum number of training timesteps.
+        return int(getattr(self.config, "num_train_timesteps", 1000))

src/diffusers/training_utils.py

@@ -6,11 +6,18 @@ import random
 import re
 import warnings
 from contextlib import contextmanager
-from typing import Any, Dict, Iterable, List, Optional, Tuple, Union
+from functools import partial
+from typing import Any, Dict, Iterable, List, Optional, Set, Tuple, Type, Union
 
 import numpy as np
 import torch
 
+
+if getattr(torch, "distributed", None) is not None:
+    from torch.distributed.fsdp import CPUOffload, ShardingStrategy
+    from torch.distributed.fsdp import FullyShardedDataParallel as FSDP
+    from torch.distributed.fsdp.wrap import transformer_auto_wrap_policy
+
 from .models import UNet2DConditionModel
 from .pipelines import DiffusionPipeline
 from .schedulers import SchedulerMixin
@@ -18,6 +25,7 @@ from .utils import (
     convert_state_dict_to_diffusers,
     convert_state_dict_to_peft,
     deprecate,
+    is_accelerate_available,
     is_peft_available,
     is_torch_npu_available,
     is_torchvision_available,
@@ -31,6 +39,9 @@ if is_transformers_available():
     if transformers.integrations.deepspeed.is_deepspeed_zero3_enabled():
         import deepspeed
 
+if is_accelerate_available():
+    from accelerate.logging import get_logger
+
 if is_peft_available():
     from peft import set_peft_model_state_dict
 
@@ -394,6 +405,86 @@ def find_nearest_bucket(h, w, bucket_options):
     return best_bucket_idx
 
 
+def _to_cpu_contiguous(state_dicts) -> dict:
+    return {k: v.detach().cpu().contiguous() if isinstance(v, torch.Tensor) else v for k, v in state_dicts.items()}
+
+
+def get_fsdp_kwargs_from_accelerator(accelerator) -> dict:
+    """
+    Extract and convert FSDP config from Accelerator into PyTorch FSDP kwargs.
+    """
+
+    kwargs = {}
+    fsdp_state = getattr(accelerator.state, "fsdp_plugin", None)
+
+    if fsdp_state is None:
+        raise ValueError("Accelerate isn't configured to handle FSDP. Please update your installation.")
+
+    fsdp_plugin = accelerator.state.fsdp_plugin
+
+    if fsdp_plugin is None:
+        # FSDP not enabled in Accelerator
+        kwargs["sharding_strategy"] = ShardingStrategy.FULL_SHARD
+    else:
+        # FSDP is enabled → use plugin's strategy, or default if None
+        kwargs["sharding_strategy"] = fsdp_plugin.sharding_strategy or ShardingStrategy.FULL_SHARD
+
+    return kwargs
+
+
+def wrap_with_fsdp(
+    model: torch.nn.Module,
+    device: Union[str, torch.device],
+    offload: bool = True,
+    use_orig_params: bool = True,
+    limit_all_gathers: bool = True,
+    fsdp_kwargs: Optional[Dict[str, Any]] = None,
+    transformer_layer_cls: Optional[Set[Type[torch.nn.Module]]] = None,
+) -> FSDP:
+    """
+    Wrap a model with FSDP using common defaults and optional transformer auto-wrapping.
+
+    Args:
+        model: Model to wrap
+        device: Target device (e.g., accelerator.device)
+        offload: Whether to enable CPU parameter offloading
+        use_orig_params: Whether to use original parameters
+        limit_all_gathers: Whether to limit all gathers
+        fsdp_kwargs: FSDP arguments (sharding_strategy, etc.) — usually from Accelerate config
+        transformer_layer_cls: Classes for auto-wrapping (if not using policy from fsdp_kwargs)
+
+    Returns:
+        FSDP-wrapped model
+    """
+
+    logger = get_logger(__name__)
+
+    if transformer_layer_cls is None:
+        # Set the default layers if transformer_layer_cls is not provided
+        transformer_layer_cls = type(model.model.language_model.layers[0])
+        logger.info(f"transformer_layer_cls is not provided, auto-inferred as {transformer_layer_cls.__name__}")
+
+    # Add auto-wrap policy if transformer layers specified
+    auto_wrap_policy = partial(
+        transformer_auto_wrap_policy,
+        transformer_layer_cls={transformer_layer_cls},
+    )
+
+    config = {
+        "device_id": device,
+        "cpu_offload": CPUOffload(offload_params=offload) if offload else None,
+        "use_orig_params": use_orig_params,
+        "limit_all_gathers": limit_all_gathers,
+        "auto_wrap_policy": auto_wrap_policy,
+    }
+
+    if fsdp_kwargs:
+        config.update(fsdp_kwargs)
+
+    fsdp_model = FSDP(model, **config)
+    return fsdp_model
+
+
 # Adapted from torch-ema https://github.com/fadel/pytorch_ema/blob/master/torch_ema/ema.py#L14
 class EMAModel:
     """

src/diffusers/utils/dummy_pt_objects.py

@@ -2634,6 +2634,21 @@ class LCMScheduler(metaclass=DummyObject):
         requires_backends(cls, ["torch"])
 
 
+class LTXEulerAncestralRFScheduler(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+    @classmethod
+    def from_config(cls, *args, **kwargs):
+        requires_backends(cls, ["torch"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch"])
+
+
 class PNDMScheduler(metaclass=DummyObject):
     _backends = ["torch"]
 

src/diffusers/utils/dummy_torch_and_transformers_objects.py

@@ -1892,6 +1892,21 @@ class LTXConditionPipeline(metaclass=DummyObject):
         requires_backends(cls, ["torch", "transformers"])
 
 
+class LTXI2VLongMultiPromptPipeline(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 LTXImageToVideoPipeline(metaclass=DummyObject):
     _backends = ["torch", "transformers"]
 
@@ -3917,6 +3932,21 @@ class ZImageImg2ImgPipeline(metaclass=DummyObject):
         requires_backends(cls, ["torch", "transformers"])
 
 
+class ZImageOmniPipeline(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"]
 

tests/quantization/torchao/test_torchao.py

@@ -671,46 +671,44 @@ class TorchAoSerializationTest(unittest.TestCase):
 class TorchAoCompileTest(QuantCompileTests, unittest.TestCase):
     @property
     def quantization_config(self):
-        from torchao.quantization import Int8WeightOnlyConfig
-
         return PipelineQuantizationConfig(
             quant_mapping={
-                "transformer": TorchAoConfig(Int8WeightOnlyConfig()),
+                "transformer": TorchAoConfig(quant_type="int8_weight_only"),
             },
         )
 
-    # @unittest.skip(
-    #     "Changing the device of AQT tensor with module._apply (called from doing module.to() in accelerate) does not work "
-    #     "when compiling."
-    # )
-    # def test_torch_compile_with_cpu_offload(self):
-    #     # RuntimeError: _apply(): Couldn't swap Linear.weight
-    #     super().test_torch_compile_with_cpu_offload()
+    @unittest.skip(
+        "Changing the device of AQT tensor with module._apply (called from doing module.to() in accelerate) does not work "
+        "when compiling."
+    )
+    def test_torch_compile_with_cpu_offload(self):
+        # RuntimeError: _apply(): Couldn't swap Linear.weight
+        super().test_torch_compile_with_cpu_offload()
 
-    # @parameterized.expand([False, True])
-    # @unittest.skip(
-    #     """
-    #     For `use_stream=False`:
-    #         - Changing the device of AQT tensor, with `param.data = param.data.to(device)` as done in group offloading implementation
-    #         is unsupported in TorchAO. When compiling, FakeTensor device mismatch causes failure.
-    #     For `use_stream=True`:
-    #         Using non-default stream requires ability to pin tensors. AQT does not seem to support this yet in TorchAO.
-    #     """
-    # )
-    # def test_torch_compile_with_group_offload_leaf(self, use_stream):
-    #     # For use_stream=False:
-    #     # If we run group offloading without compilation, we will see:
-    #     #   RuntimeError: Attempted to set the storage of a tensor on device "cpu" to a storage on different device "cuda:0".  This is no longer allowed; the devices must match.
-    #     # When running with compilation, the error ends up being different:
-    #     #   Dynamo failed to run FX node with fake tensors: call_function <built-in function linear>(*(FakeTensor(..., device='cuda:0', size=(s0, 256), dtype=torch.bfloat16), AffineQuantizedTensor(tensor_impl=PlainAQTTensorImpl(data=FakeTensor(..., size=(1536, 256), dtype=torch.int8)... , scale=FakeTensor(..., size=(1536,), dtype=torch.bfloat16)... , zero_point=FakeTensor(..., size=(1536,), dtype=torch.int64)... , _layout=PlainLayout()), block_size=(1, 256), shape=torch.Size([1536, 256]), device=cpu, dtype=torch.bfloat16, requires_grad=False), Parameter(FakeTensor(..., device='cuda:0', size=(1536,), dtype=torch.bfloat16,
-    #     #   requires_grad=True))), **{}): got RuntimeError('Unhandled FakeTensor Device Propagation for aten.mm.default, found two different devices cuda:0, cpu')
-    #     # Looks like something that will have to be looked into upstream.
-    #     # for linear layers, weight.tensor_impl shows cuda... but:
-    #     # weight.tensor_impl.{data,scale,zero_point}.device will be cpu
+    @parameterized.expand([False, True])
+    @unittest.skip(
+        """
+        For `use_stream=False`:
+            - Changing the device of AQT tensor, with `param.data = param.data.to(device)` as done in group offloading implementation
+            is unsupported in TorchAO. When compiling, FakeTensor device mismatch causes failure.
+        For `use_stream=True`:
+            Using non-default stream requires ability to pin tensors. AQT does not seem to support this yet in TorchAO.
+        """
+    )
+    def test_torch_compile_with_group_offload_leaf(self, use_stream):
+        # For use_stream=False:
+        # If we run group offloading without compilation, we will see:
+        #   RuntimeError: Attempted to set the storage of a tensor on device "cpu" to a storage on different device "cuda:0".  This is no longer allowed; the devices must match.
+        # When running with compilation, the error ends up being different:
+        #   Dynamo failed to run FX node with fake tensors: call_function <built-in function linear>(*(FakeTensor(..., device='cuda:0', size=(s0, 256), dtype=torch.bfloat16), AffineQuantizedTensor(tensor_impl=PlainAQTTensorImpl(data=FakeTensor(..., size=(1536, 256), dtype=torch.int8)... , scale=FakeTensor(..., size=(1536,), dtype=torch.bfloat16)... , zero_point=FakeTensor(..., size=(1536,), dtype=torch.int64)... , _layout=PlainLayout()), block_size=(1, 256), shape=torch.Size([1536, 256]), device=cpu, dtype=torch.bfloat16, requires_grad=False), Parameter(FakeTensor(..., device='cuda:0', size=(1536,), dtype=torch.bfloat16,
+        #   requires_grad=True))), **{}): got RuntimeError('Unhandled FakeTensor Device Propagation for aten.mm.default, found two different devices cuda:0, cpu')
+        # Looks like something that will have to be looked into upstream.
+        # for linear layers, weight.tensor_impl shows cuda... but:
+        # weight.tensor_impl.{data,scale,zero_point}.device will be cpu
 
-    #     # For use_stream=True:
-    #     # NotImplementedError: AffineQuantizedTensor dispatch: attempting to run unimplemented operator/function: func=<OpOverload(op='aten.is_pinned', overload='default')>, types=(<class 'torchao.dtypes.affine_quantized_tensor.AffineQuantizedTensor'>,), arg_types=(<class 'torchao.dtypes.affine_quantized_tensor.AffineQuantizedTensor'>,), kwarg_types={}
-    #     super()._test_torch_compile_with_group_offload_leaf(use_stream=use_stream)
+        # For use_stream=True:
+        # NotImplementedError: AffineQuantizedTensor dispatch: attempting to run unimplemented operator/function: func=<OpOverload(op='aten.is_pinned', overload='default')>, types=(<class 'torchao.dtypes.affine_quantized_tensor.AffineQuantizedTensor'>,), arg_types=(<class 'torchao.dtypes.affine_quantized_tensor.AffineQuantizedTensor'>,), kwarg_types={}
+        super()._test_torch_compile_with_group_offload_leaf(use_stream=use_stream)
 
 
 # Slices for these tests have been obtained on our aws-g6e-xlarge-plus runners