diffusers 0.32.1__py3-none-any.whl → 0.33.0__py3-none-any.whl

diffusers/pipelines/t2i_adapter/pipeline_stable_diffusion_xl_adapter.py

@@ -43,6 +43,7 @@ from ...schedulers import KarrasDiffusionSchedulers
 from ...utils import (
     PIL_INTERPOLATION,
     USE_PEFT_BACKEND,
+    is_torch_xla_available,
     logging,
     replace_example_docstring,
     scale_lora_layers,
@@ -53,8 +54,16 @@ from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
 from ..stable_diffusion_xl.pipeline_output import StableDiffusionXLPipelineOutput
 
 
+if is_torch_xla_available():
+    import torch_xla.core.xla_model as xm
+
+    XLA_AVAILABLE = True
+else:
+    XLA_AVAILABLE = False
+
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
 
+
 EXAMPLE_DOC_STRING = """
     Examples:
         ```py
@@ -248,7 +257,8 @@ class StableDiffusionXLAdapterPipeline(
             [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
         safety_checker ([`StableDiffusionSafetyChecker`]):
             Classification module that estimates whether generated images could be considered offensive or harmful.
-            Please, refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for details.
+            Please, refer to the [model card](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5) for
+            details.
         feature_extractor ([`CLIPImageProcessor`]):
             Model that extracts features from generated images to be used as inputs for the `safety_checker`.
     """
@@ -292,9 +302,13 @@ class StableDiffusionXLAdapterPipeline(
             image_encoder=image_encoder,
         )
         self.register_to_config(force_zeros_for_empty_prompt=force_zeros_for_empty_prompt)
-        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8
         self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
-        self.default_sample_size = self.unet.config.sample_size
+        self.default_sample_size = (
+            self.unet.config.sample_size
+            if hasattr(self, "unet") and self.unet is not None and hasattr(self.unet.config, "sample_size")
+            else 128
+        )
 
     # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.encode_prompt
     def encode_prompt(
@@ -422,7 +436,9 @@ class StableDiffusionXLAdapterPipeline(
                 prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True)
 
                 # We are only ALWAYS interested in the pooled output of the final text encoder
-                pooled_prompt_embeds = prompt_embeds[0]
+                if pooled_prompt_embeds is None and prompt_embeds[0].ndim == 2:
+                    pooled_prompt_embeds = prompt_embeds[0]
+
                 if clip_skip is None:
                     prompt_embeds = prompt_embeds.hidden_states[-2]
                 else:
@@ -481,8 +497,10 @@ class StableDiffusionXLAdapterPipeline(
                     uncond_input.input_ids.to(device),
                     output_hidden_states=True,
                 )
+
                 # We are only ALWAYS interested in the pooled output of the final text encoder
-                negative_pooled_prompt_embeds = negative_prompt_embeds[0]
+                if negative_pooled_prompt_embeds is None and negative_prompt_embeds[0].ndim == 2:
+                    negative_pooled_prompt_embeds = negative_prompt_embeds[0]
                 negative_prompt_embeds = negative_prompt_embeds.hidden_states[-2]
 
                 negative_prompt_embeds_list.append(negative_prompt_embeds)
@@ -1261,6 +1279,9 @@ class StableDiffusionXLAdapterPipeline(
                         step_idx = i // getattr(self.scheduler, "order", 1)
                         callback(step_idx, t, latents)
 
+                if XLA_AVAILABLE:
+                    xm.mark_step()
+
         if not output_type == "latent":
             # make sure the VAE is in float32 mode, as it overflows in float16
             needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast

diffusers/pipelines/text_to_video_synthesis/pipeline_text_to_video_synth.py

@@ -25,6 +25,7 @@ from ...schedulers import KarrasDiffusionSchedulers
 from ...utils import (
     USE_PEFT_BACKEND,
     deprecate,
+    is_torch_xla_available,
     logging,
     replace_example_docstring,
     scale_lora_layers,
@@ -36,8 +37,16 @@ from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
 from . import TextToVideoSDPipelineOutput
 
 
+if is_torch_xla_available():
+    import torch_xla.core.xla_model as xm
+
+    XLA_AVAILABLE = True
+else:
+    XLA_AVAILABLE = False
+
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
 
+
 EXAMPLE_DOC_STRING = """
     Examples:
         ```py
@@ -105,7 +114,7 @@ class TextToVideoSDPipeline(
             unet=unet,
             scheduler=scheduler,
         )
-        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8
         self.video_processor = VideoProcessor(do_resize=False, vae_scale_factor=self.vae_scale_factor)
 
     # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt
@@ -627,6 +636,9 @@ class TextToVideoSDPipeline(
                         step_idx = i // getattr(self.scheduler, "order", 1)
                         callback(step_idx, t, latents)
 
+                if XLA_AVAILABLE:
+                    xm.mark_step()
+
         # 8. Post processing
         if output_type == "latent":
             video = latents

diffusers/pipelines/text_to_video_synthesis/pipeline_text_to_video_synth_img2img.py

@@ -26,6 +26,7 @@ from ...schedulers import KarrasDiffusionSchedulers
 from ...utils import (
     USE_PEFT_BACKEND,
     deprecate,
+    is_torch_xla_available,
     logging,
     replace_example_docstring,
     scale_lora_layers,
@@ -37,8 +38,16 @@ from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
 from . import TextToVideoSDPipelineOutput
 
 
+if is_torch_xla_available():
+    import torch_xla.core.xla_model as xm
+
+    XLA_AVAILABLE = True
+else:
+    XLA_AVAILABLE = False
+
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
 
+
 EXAMPLE_DOC_STRING = """
     Examples:
         ```py
@@ -140,7 +149,7 @@ class VideoToVideoSDPipeline(
             unet=unet,
             scheduler=scheduler,
         )
-        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8
         self.video_processor = VideoProcessor(do_resize=False, vae_scale_factor=self.vae_scale_factor)
 
     # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt
@@ -679,6 +688,9 @@ class VideoToVideoSDPipeline(
                         step_idx = i // getattr(self.scheduler, "order", 1)
                         callback(step_idx, t, latents)
 
+                if XLA_AVAILABLE:
+                    xm.mark_step()
+
         # manually for max memory savings
         if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
             self.unet.to("cpu")

diffusers/pipelines/text_to_video_synthesis/pipeline_text_to_video_zero.py

@@ -11,16 +11,30 @@ from torch.nn.functional import grid_sample
 from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer
 
 from ...image_processor import VaeImageProcessor
-from ...loaders import StableDiffusionLoraLoaderMixin, TextualInversionLoaderMixin
+from ...loaders import FromSingleFileMixin, StableDiffusionLoraLoaderMixin, TextualInversionLoaderMixin
 from ...models import AutoencoderKL, UNet2DConditionModel
 from ...models.lora import adjust_lora_scale_text_encoder
 from ...schedulers import KarrasDiffusionSchedulers
-from ...utils import USE_PEFT_BACKEND, BaseOutput, logging, scale_lora_layers, unscale_lora_layers
+from ...utils import (
+    USE_PEFT_BACKEND,
+    BaseOutput,
+    is_torch_xla_available,
+    logging,
+    scale_lora_layers,
+    unscale_lora_layers,
+)
 from ...utils.torch_utils import randn_tensor
 from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
 from ..stable_diffusion import StableDiffusionSafetyChecker
 
 
+if is_torch_xla_available():
+    import torch_xla.core.xla_model as xm
+
+    XLA_AVAILABLE = True
+else:
+    XLA_AVAILABLE = False
+
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
 
 
@@ -282,7 +296,11 @@ def create_motion_field_and_warp_latents(motion_field_strength_x, motion_field_s
 
 
 class TextToVideoZeroPipeline(
-    DiffusionPipeline, StableDiffusionMixin, TextualInversionLoaderMixin, StableDiffusionLoraLoaderMixin
+    DiffusionPipeline,
+    StableDiffusionMixin,
+    TextualInversionLoaderMixin,
+    StableDiffusionLoraLoaderMixin,
+    FromSingleFileMixin,
 ):
     r"""
     Pipeline for zero-shot text-to-video generation using Stable Diffusion.
@@ -304,8 +322,8 @@ class TextToVideoZeroPipeline(
             [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
         safety_checker ([`StableDiffusionSafetyChecker`]):
             Classification module that estimates whether generated images could be considered offensive or harmful.
-            Please refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for more details
-            about a model's potential harms.
+            Please refer to the [model card](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5) for
+            more details about a model's potential harms.
         feature_extractor ([`CLIPImageProcessor`]):
             A [`CLIPImageProcessor`] to extract features from generated images; used as inputs to the `safety_checker`.
     """
@@ -340,7 +358,7 @@ class TextToVideoZeroPipeline(
                 " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
                 " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
             )
-        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8
         self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
 
     def forward_loop(self, x_t0, t0, t1, generator):
@@ -440,6 +458,10 @@ class TextToVideoZeroPipeline(
                     if callback is not None and i % callback_steps == 0:
                         step_idx = i // getattr(self.scheduler, "order", 1)
                         callback(step_idx, t, latents)
+
+                if XLA_AVAILABLE:
+                    xm.mark_step()
+
         return latents.clone().detach()
 
     # Copied from diffusers.pipelines.stable_diffusion_k_diffusion.pipeline_stable_diffusion_k_diffusion.StableDiffusionKDiffusionPipeline.check_inputs

diffusers/pipelines/text_to_video_synthesis/pipeline_text_to_video_zero_sdxl.py

@@ -42,6 +42,16 @@ if is_invisible_watermark_available():
     from ..stable_diffusion_xl.watermark import StableDiffusionXLWatermarker
 
 
+from ...utils import is_torch_xla_available
+
+
+if is_torch_xla_available():
+    import torch_xla.core.xla_model as xm
+
+    XLA_AVAILABLE = True
+else:
+    XLA_AVAILABLE = False
+
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
 
 
@@ -409,10 +419,14 @@ class TextToVideoZeroSDXLPipeline(
             feature_extractor=feature_extractor,
         )
         self.register_to_config(force_zeros_for_empty_prompt=force_zeros_for_empty_prompt)
-        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8
         self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
 
-        self.default_sample_size = self.unet.config.sample_size
+        self.default_sample_size = (
+            self.unet.config.sample_size
+            if hasattr(self, "unet") and self.unet is not None and hasattr(self.unet.config, "sample_size")
+            else 128
+        )
 
         add_watermarker = add_watermarker if add_watermarker is not None else is_invisible_watermark_available()
 
@@ -705,7 +719,9 @@ class TextToVideoZeroSDXLPipeline(
                 prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True)
 
                 # We are only ALWAYS interested in the pooled output of the final text encoder
-                pooled_prompt_embeds = prompt_embeds[0]
+                if pooled_prompt_embeds is None and prompt_embeds[0].ndim == 2:
+                    pooled_prompt_embeds = prompt_embeds[0]
+
                 if clip_skip is None:
                     prompt_embeds = prompt_embeds.hidden_states[-2]
                 else:
@@ -764,8 +780,10 @@ class TextToVideoZeroSDXLPipeline(
                     uncond_input.input_ids.to(device),
                     output_hidden_states=True,
                 )
+
                 # We are only ALWAYS interested in the pooled output of the final text encoder
-                negative_pooled_prompt_embeds = negative_prompt_embeds[0]
+                if negative_pooled_prompt_embeds is None and negative_prompt_embeds[0].ndim == 2:
+                    negative_pooled_prompt_embeds = negative_prompt_embeds[0]
                 negative_prompt_embeds = negative_prompt_embeds.hidden_states[-2]
 
                 negative_prompt_embeds_list.append(negative_prompt_embeds)
@@ -922,6 +940,10 @@ class TextToVideoZeroSDXLPipeline(
                     progress_bar.update()
                     if callback is not None and i % callback_steps == 0:
                         callback(i, t, latents)
+
+                if XLA_AVAILABLE:
+                    xm.mark_step()
+
         return latents.clone().detach()
 
     @torch.no_grad()

diffusers/pipelines/transformers_loading_utils.py

@@ -0,0 +1,121 @@
+# coding=utf-8
+# Copyright 2025 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import contextlib
+import os
+import tempfile
+from typing import TYPE_CHECKING, Dict
+
+from huggingface_hub import DDUFEntry
+from tqdm import tqdm
+
+from ..utils import is_safetensors_available, is_transformers_available, is_transformers_version
+
+
+if TYPE_CHECKING:
+    from transformers import PreTrainedModel, PreTrainedTokenizer
+
+if is_transformers_available():
+    from transformers import PreTrainedModel, PreTrainedTokenizer
+
+if is_safetensors_available():
+    import safetensors.torch
+
+
+def _load_tokenizer_from_dduf(
+    cls: "PreTrainedTokenizer", name: str, dduf_entries: Dict[str, DDUFEntry], **kwargs
+) -> "PreTrainedTokenizer":
+    """
+    Load a tokenizer from a DDUF archive.
+
+    In practice, `transformers` do not provide a way to load a tokenizer from a DDUF archive. This function is a
+    workaround by extracting the tokenizer files from the DDUF archive and loading the tokenizer from the extracted
+    files. There is an extra cost of extracting the files, but of limited impact as the tokenizer files are usually
+    small-ish.
+    """
+    with tempfile.TemporaryDirectory() as tmp_dir:
+        for entry_name, entry in dduf_entries.items():
+            if entry_name.startswith(name + "/"):
+                tmp_entry_path = os.path.join(tmp_dir, *entry_name.split("/"))
+                # need to create intermediary directory if they don't exist
+                os.makedirs(os.path.dirname(tmp_entry_path), exist_ok=True)
+                with open(tmp_entry_path, "wb") as f:
+                    with entry.as_mmap() as mm:
+                        f.write(mm)
+        return cls.from_pretrained(os.path.dirname(tmp_entry_path), **kwargs)
+
+
+def _load_transformers_model_from_dduf(
+    cls: "PreTrainedModel", name: str, dduf_entries: Dict[str, DDUFEntry], **kwargs
+) -> "PreTrainedModel":
+    """
+    Load a transformers model from a DDUF archive.
+
+    In practice, `transformers` do not provide a way to load a model from a DDUF archive. This function is a workaround
+    by instantiating a model from the config file and loading the weights from the DDUF archive directly.
+    """
+    config_file = dduf_entries.get(f"{name}/config.json")
+    if config_file is None:
+        raise EnvironmentError(
+            f"Could not find a config.json file for component {name} in DDUF file (contains {dduf_entries.keys()})."
+        )
+    generation_config = dduf_entries.get(f"{name}/generation_config.json", None)
+
+    weight_files = [
+        entry
+        for entry_name, entry in dduf_entries.items()
+        if entry_name.startswith(f"{name}/") and entry_name.endswith(".safetensors")
+    ]
+    if not weight_files:
+        raise EnvironmentError(
+            f"Could not find any weight file for component {name} in DDUF file (contains {dduf_entries.keys()})."
+        )
+    if not is_safetensors_available():
+        raise EnvironmentError(
+            "Safetensors is not available, cannot load model from DDUF. Please `pip install safetensors`."
+        )
+    if is_transformers_version("<", "4.47.0"):
+        raise ImportError(
+            "You need to install `transformers>4.47.0` in order to load a transformers model from a DDUF file. "
+            "You can install it with: `pip install --upgrade transformers`"
+        )
+
+    with tempfile.TemporaryDirectory() as tmp_dir:
+        from transformers import AutoConfig, GenerationConfig
+
+        tmp_config_file = os.path.join(tmp_dir, "config.json")
+        with open(tmp_config_file, "w") as f:
+            f.write(config_file.read_text())
+        config = AutoConfig.from_pretrained(tmp_config_file)
+        if generation_config is not None:
+            tmp_generation_config_file = os.path.join(tmp_dir, "generation_config.json")
+            with open(tmp_generation_config_file, "w") as f:
+                f.write(generation_config.read_text())
+            generation_config = GenerationConfig.from_pretrained(tmp_generation_config_file)
+        state_dict = {}
+        with contextlib.ExitStack() as stack:
+            for entry in tqdm(weight_files, desc="Loading state_dict"):  # Loop over safetensors files
+                # Memory-map the safetensors file
+                mmap = stack.enter_context(entry.as_mmap())
+                # Load tensors from the memory-mapped file
+                tensors = safetensors.torch.load(mmap)
+                # Update the state dictionary with tensors
+                state_dict.update(tensors)
+            return cls.from_pretrained(
+                pretrained_model_name_or_path=None,
+                config=config,
+                generation_config=generation_config,
+                state_dict=state_dict,
+                **kwargs,
+            )

diffusers/pipelines/unclip/pipeline_unclip.py

@@ -22,12 +22,19 @@ from transformers.models.clip.modeling_clip import CLIPTextModelOutput
 
 from ...models import PriorTransformer, UNet2DConditionModel, UNet2DModel
 from ...schedulers import UnCLIPScheduler
-from ...utils import logging
+from ...utils import is_torch_xla_available, logging
 from ...utils.torch_utils import randn_tensor
 from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput
 from .text_proj import UnCLIPTextProjModel
 
 
+if is_torch_xla_available():
+    import torch_xla.core.xla_model as xm
+
+    XLA_AVAILABLE = True
+else:
+    XLA_AVAILABLE = False
+
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
 
 
@@ -474,6 +481,9 @@ class UnCLIPPipeline(DiffusionPipeline):
                 noise_pred, t, super_res_latents, prev_timestep=prev_timestep, generator=generator
             ).prev_sample
 
+            if XLA_AVAILABLE:
+                xm.mark_step()
+
         image = super_res_latents
         # done super res
 

diffusers/pipelines/unclip/pipeline_unclip_image_variation.py

@@ -27,12 +27,19 @@ from transformers import (
 
 from ...models import UNet2DConditionModel, UNet2DModel
 from ...schedulers import UnCLIPScheduler
-from ...utils import logging
+from ...utils import is_torch_xla_available, logging
 from ...utils.torch_utils import randn_tensor
 from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput
 from .text_proj import UnCLIPTextProjModel
 
 
+if is_torch_xla_available():
+    import torch_xla.core.xla_model as xm
+
+    XLA_AVAILABLE = True
+else:
+    XLA_AVAILABLE = False
+
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
 
 
@@ -400,6 +407,9 @@ class UnCLIPImageVariationPipeline(DiffusionPipeline):
                 noise_pred, t, super_res_latents, prev_timestep=prev_timestep, generator=generator
             ).prev_sample
 
+            if XLA_AVAILABLE:
+                xm.mark_step()
+
         image = super_res_latents
 
         # done super res

diffusers/pipelines/unidiffuser/pipeline_unidiffuser.py

@@ -18,7 +18,14 @@ from ...loaders import StableDiffusionLoraLoaderMixin, TextualInversionLoaderMix
 from ...models import AutoencoderKL
 from ...models.lora import adjust_lora_scale_text_encoder
 from ...schedulers import KarrasDiffusionSchedulers
-from ...utils import USE_PEFT_BACKEND, deprecate, logging, scale_lora_layers, unscale_lora_layers
+from ...utils import (
+    USE_PEFT_BACKEND,
+    deprecate,
+    is_torch_xla_available,
+    logging,
+    scale_lora_layers,
+    unscale_lora_layers,
+)
 from ...utils.outputs import BaseOutput
 from ...utils.torch_utils import randn_tensor
 from ..pipeline_utils import DiffusionPipeline
@@ -26,6 +33,13 @@ from .modeling_text_decoder import UniDiffuserTextDecoder
 from .modeling_uvit import UniDiffuserModel
 
 
+if is_torch_xla_available():
+    import torch_xla.core.xla_model as xm
+
+    XLA_AVAILABLE = True
+else:
+    XLA_AVAILABLE = False
+
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
 
 
@@ -117,7 +131,7 @@ class UniDiffuserPipeline(DiffusionPipeline):
             scheduler=scheduler,
         )
 
-        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8
         self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
 
         self.num_channels_latents = vae.config.latent_channels
@@ -1378,6 +1392,9 @@ class UniDiffuserPipeline(DiffusionPipeline):
                         step_idx = i // getattr(self.scheduler, "order", 1)
                         callback(step_idx, t, latents)
 
+                if XLA_AVAILABLE:
+                    xm.mark_step()
+
         # 9. Post-processing
         image = None
         text = None

diffusers/pipelines/wan/__init__.py

@@ -0,0 +1,51 @@
+from typing import TYPE_CHECKING
+
+from ...utils import (
+    DIFFUSERS_SLOW_IMPORT,
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    get_objects_from_module,
+    is_torch_available,
+    is_transformers_available,
+)
+
+
+_dummy_objects = {}
+_import_structure = {}
+
+
+try:
+    if not (is_transformers_available() and is_torch_available()):
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    from ...utils import dummy_torch_and_transformers_objects  # noqa F403
+
+    _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects))
+else:
+    _import_structure["pipeline_wan"] = ["WanPipeline"]
+    _import_structure["pipeline_wan_i2v"] = ["WanImageToVideoPipeline"]
+    _import_structure["pipeline_wan_video2video"] = ["WanVideoToVideoPipeline"]
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+    try:
+        if not (is_transformers_available() and is_torch_available()):
+            raise OptionalDependencyNotAvailable()
+
+    except OptionalDependencyNotAvailable:
+        from ...utils.dummy_torch_and_transformers_objects import *
+    else:
+        from .pipeline_wan import WanPipeline
+        from .pipeline_wan_i2v import WanImageToVideoPipeline
+        from .pipeline_wan_video2video import WanVideoToVideoPipeline
+
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(
+        __name__,
+        globals()["__file__"],
+        _import_structure,
+        module_spec=__spec__,
+    )
+
+    for name, value in _dummy_objects.items():
+        setattr(sys.modules[__name__], name, value)

diffusers/pipelines/wan/pipeline_output.py

@@ -0,0 +1,20 @@
+from dataclasses import dataclass
+
+import torch
+
+from diffusers.utils import BaseOutput
+
+
+@dataclass
+class WanPipelineOutput(BaseOutput):
+    r"""
+    Output class for Wan pipelines.
+
+    Args:
+        frames (`torch.Tensor`, `np.ndarray`, or List[List[PIL.Image.Image]]):
+            List of video outputs - It can be a nested list of length `batch_size,` with each sub-list containing
+            denoised PIL image sequences of length `num_frames.` It can also be a NumPy array or Torch tensor of shape
+            `(batch_size, num_frames, channels, height, width)`.
+    """
+
+    frames: torch.Tensor