[Alibaba Wan Team] continue on #10921 Wan2.1 (#10922)

* Add wanx pipeline, model and example

* wanx_merged_v1

* change WanX into Wan

* fix i2v fp32 oom error

Link: https://code.alibaba-inc.com/open_wanx2/diffusers/codereview/20607813



* support t2v load fp32 ckpt

* add example

* final merge v1

* Update autoencoder_kl_wan.py

* up

* update middle, test up_block

* up up

* one less nn.sequential

* up more

* up

* more

* [refactor] [wip] Wan transformer/pipeline (#10926)

* update

* update

* refactor rope

* refactor pipeline

* make fix-copies

* add transformer test

* update

* update

* make style

* update tests

* tests

* conversion script

* conversion script

* update

* docs

* remove unused code

* fix _toctree.yml

* update dtype

* fix test

* fix tests: scale

* up

* more

* Apply suggestions from code review

* Apply suggestions from code review

* style

* Update scripts/convert_wan_to_diffusers.py

* update docs

* fix

---------
Co-authored-by: Yitong Huang <huangyitong.hyt@alibaba-inc.com>
Co-authored-by: 亚森 <wangjiayu.wjy@alibaba-inc.com>
Co-authored-by: Aryan <aryan@huggingface.co>

docs/source/en/_toctree.yml

@@ -314,6 +314,8 @@
         title: Transformer2DModel
       - local: api/models/transformer_temporal
         title: TransformerTemporalModel
+      - local: api/models/wan_transformer_3d
+        title: WanTransformer3DModel
       title: Transformers
     - sections:
       - local: api/models/stable_cascade_unet
@@ -344,6 +346,8 @@
         title: AutoencoderKLLTXVideo
       - local: api/models/autoencoderkl_mochi
         title: AutoencoderKLMochi
+      - local: api/models/autoencoder_kl_wan
+        title: AutoencoderKLWan
       - local: api/models/asymmetricautoencoderkl
         title: AsymmetricAutoencoderKL
       - local: api/models/autoencoder_dc
@@ -534,6 +538,8 @@
       title: UniDiffuser
     - local: api/pipelines/value_guided_sampling
       title: Value-guided sampling
+    - local: api/pipelines/wan
+      title: Wan
     - local: api/pipelines/wuerstchen
       title: Wuerstchen
     title: Pipelines

docs/source/en/api/models/autoencoder_kl_wan.md

+<!-- Copyright 2024 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. -->
+
+# AutoencoderKLWan
+
+The 3D variational autoencoder (VAE) model with KL loss used in [Wan 2.1](https://github.com/Wan-Video/Wan2.1) by the Alibaba Wan Team.
+
+The model can be loaded with the following code snippet.
+
+```python
+from diffusers import AutoencoderKLWan
+
+vae = AutoencoderKLWan.from_pretrained("Wan-AI/Wan2.1-T2V-1.3B-Diffusers", subfolder="vae", torch_dtype=torch.float32)
+```
+
+## AutoencoderKLWan
+
+[[autodoc]] AutoencoderKLWan
+  - decode
+  - all
+
+## DecoderOutput
+
+[[autodoc]] models.autoencoders.vae.DecoderOutput

docs/source/en/api/models/wan_transformer_3d.md

+<!-- Copyright 2024 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. -->
+
+# WanTransformer3DModel
+
+A Diffusion Transformer model for 3D video-like data was introduced in [Wan 2.1](https://github.com/Wan-Video/Wan2.1) by the Alibaba Wan Team.
+
+The model can be loaded with the following code snippet.
+
+```python
+from diffusers import WanTransformer3DModel
+
+transformer = WanTransformer3DModel.from_pretrained("Wan-AI/Wan2.1-T2V-1.3B-Diffusers", subfolder="transformer", torch_dtype=torch.bfloat16)
+```
+
+## WanTransformer3DModel
+
+[[autodoc]] WanTransformer3DModel
+
+## Transformer2DModelOutput
+
+[[autodoc]] models.modeling_outputs.Transformer2DModelOutput

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

+<!-- Copyright 2024 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. -->
+
+# Wan
+
+[Wan 2.1](https://github.com/Wan-Video/Wan2.1) by the Alibaba Wan Team.
+
+<!-- TODO(aryan): update abstract once paper is out -->
+
+<Tip>
+
+Make sure to check out the Schedulers [guide](../../using-diffusers/schedulers) to learn how to explore the tradeoff between scheduler speed and quality, and see the [reuse components across pipelines](../../using-diffusers/loading#reuse-a-pipeline) section to learn how to efficiently load the same components into multiple pipelines.
+
+</Tip>
+
+Recommendations for inference:
+- VAE in `torch.float32` for better decoding quality.
+- `num_frames` should be of the form `4 * k + 1`, for example `49` or `81`.
+- For smaller resolution videos, try lower values of `shift` (between `2.0` to `5.0`) in the [Scheduler](https://huggingface.co/docs/diffusers/main/en/api/schedulers/flow_match_euler_discrete#diffusers.FlowMatchEulerDiscreteScheduler.shift). For larger resolution videos, try higher values (between `7.0` and `12.0`). The default value is `3.0` for Wan.
+
+### Using a custom scheduler
+
+Wan can be used with many different schedulers, each with their own benefits regarding speed and generation quality. By default, Wan uses the `UniPCMultistepScheduler(prediction_type="flow_prediction", use_flow_sigmas=True, flow_shift=3.0)` scheduler. You can use a different scheduler as follows:
+
+```python
+from diffusers import FlowMatchEulerDiscreteScheduler, UniPCMultistepScheduler, WanPipeline
+
+scheduler_a = FlowMatchEulerDiscreteScheduler(shift=5.0)
+scheduler_b = UniPCMultistepScheduler(prediction_type="flow_prediction", use_flow_sigmas=True, flow_shift=4.0)
+
+pipe = WanPipeline.from_pretrained("Wan-AI/Wan2.1-T2V-1.3B-Diffusers", scheduler=<CUSTOM_SCHEDULER_HERE>)
+
+# or,
+pipe.scheduler = <CUSTOM_SCHEDULER_HERE>
+```
+
+## WanPipeline
+
+[[autodoc]] WanPipeline
+  - all
+  - __call__
+
+## WanImageToVideoPipeline
+
+[[autodoc]] WanImageToVideoPipeline
+  - all
+  - __call__
+
+## WanPipelineOutput
+
+[[autodoc]] pipelines.wan.pipeline_output.WanPipelineOutput

scripts/convert_wan_to_diffusers.py

+import argparse
+import pathlib
+from typing import Any, Dict
+
+import torch
+from accelerate import init_empty_weights
+from huggingface_hub import hf_hub_download, snapshot_download
+from safetensors.torch import load_file
+from transformers import AutoProcessor, AutoTokenizer, CLIPVisionModelWithProjection, UMT5EncoderModel
+
+from diffusers import (
+    AutoencoderKLWan,
+    UniPCMultistepScheduler,
+    WanImageToVideoPipeline,
+    WanPipeline,
+    WanTransformer3DModel,
+)
+
+
+TRANSFORMER_KEYS_RENAME_DICT = {
+    "time_embedding.0": "condition_embedder.time_embedder.linear_1",
+    "time_embedding.2": "condition_embedder.time_embedder.linear_2",
+    "text_embedding.0": "condition_embedder.text_embedder.linear_1",
+    "text_embedding.2": "condition_embedder.text_embedder.linear_2",
+    "time_projection.1": "condition_embedder.time_proj",
+    "head.modulation": "scale_shift_table",
+    "head.head": "proj_out",
+    "modulation": "scale_shift_table",
+    "ffn.0": "ffn.net.0.proj",
+    "ffn.2": "ffn.net.2",
+    # Hack to swap the layer names
+    # The original model calls the norms in following order: norm1, norm3, norm2
+    # We convert it to: norm1, norm2, norm3
+    "norm2": "norm__placeholder",
+    "norm3": "norm2",
+    "norm__placeholder": "norm3",
+    # For the I2V model
+    "img_emb.proj.0": "condition_embedder.image_embedder.norm1",
+    "img_emb.proj.1": "condition_embedder.image_embedder.ff.net.0.proj",
+    "img_emb.proj.3": "condition_embedder.image_embedder.ff.net.2",
+    "img_emb.proj.4": "condition_embedder.image_embedder.norm2",
+}
+
+TRANSFORMER_SPECIAL_KEYS_REMAP = {}
+
+
+def update_state_dict_(state_dict: Dict[str, Any], old_key: str, new_key: str) -> Dict[str, Any]:
+    state_dict[new_key] = state_dict.pop(old_key)
+
+
+def load_sharded_safetensors(dir: pathlib.Path):
+    file_paths = list(dir.glob("diffusion_pytorch_model*.safetensors"))
+    state_dict = {}
+    for path in file_paths:
+        state_dict.update(load_file(path))
+    return state_dict
+
+
+def get_transformer_config(model_type: str) -> Dict[str, Any]:
+    if model_type == "Wan-T2V-1.3B":
+        config = {
+            "model_id": "StevenZhang/Wan2.1-T2V-1.3B-Diff",
+            "diffusers_config": {
+                "added_kv_proj_dim": None,
+                "attention_head_dim": 128,
+                "cross_attn_norm": True,
+                "eps": 1e-06,
+                "ffn_dim": 8960,
+                "freq_dim": 256,
+                "in_channels": 16,
+                "num_attention_heads": 12,
+                "num_layers": 30,
+                "out_channels": 16,
+                "patch_size": [1, 2, 2],
+                "qk_norm": "rms_norm_across_heads",
+                "text_dim": 4096,
+            },
+        }
+    elif model_type == "Wan-T2V-14B":
+        config = {
+            "model_id": "StevenZhang/Wan2.1-T2V-14B-Diff",
+            "diffusers_config": {
+                "added_kv_proj_dim": None,
+                "attention_head_dim": 128,
+                "cross_attn_norm": True,
+                "eps": 1e-06,
+                "ffn_dim": 13824,
+                "freq_dim": 256,
+                "in_channels": 16,
+                "num_attention_heads": 40,
+                "num_layers": 40,
+                "out_channels": 16,
+                "patch_size": [1, 2, 2],
+                "qk_norm": "rms_norm_across_heads",
+                "text_dim": 4096,
+            },
+        }
+    elif model_type == "Wan-I2V-14B-480p":
+        config = {
+            "model_id": "StevenZhang/Wan2.1-I2V-14B-480P-Diff",
+            "diffusers_config": {
+                "image_dim": 1280,
+                "added_kv_proj_dim": 5120,
+                "attention_head_dim": 128,
+                "cross_attn_norm": True,
+                "eps": 1e-06,
+                "ffn_dim": 13824,
+                "freq_dim": 256,
+                "in_channels": 36,
+                "num_attention_heads": 40,
+                "num_layers": 40,
+                "out_channels": 16,
+                "patch_size": [1, 2, 2],
+                "qk_norm": "rms_norm_across_heads",
+                "text_dim": 4096,
+            },
+        }
+    elif model_type == "Wan-I2V-14B-720p":
+        config = {
+            "model_id": "StevenZhang/Wan2.1-I2V-14B-720P-Diff",
+            "diffusers_config": {
+                "image_dim": 1280,
+                "added_kv_proj_dim": 5120,
+                "attention_head_dim": 128,
+                "cross_attn_norm": True,
+                "eps": 1e-06,
+                "ffn_dim": 13824,
+                "freq_dim": 256,
+                "in_channels": 36,
+                "num_attention_heads": 40,
+                "num_layers": 40,
+                "out_channels": 16,
+                "patch_size": [1, 2, 2],
+                "qk_norm": "rms_norm_across_heads",
+                "text_dim": 4096,
+            },
+        }
+    return config
+
+
+def convert_transformer(model_type: str):
+    config = get_transformer_config(model_type)
+    diffusers_config = config["diffusers_config"]
+    model_id = config["model_id"]
+    model_dir = pathlib.Path(snapshot_download(model_id, repo_type="model"))
+
+    original_state_dict = load_sharded_safetensors(model_dir)
+
+    with init_empty_weights():
+        transformer = WanTransformer3DModel.from_config(diffusers_config)
+
+    for key in list(original_state_dict.keys()):
+        new_key = key[:]
+        for replace_key, rename_key in TRANSFORMER_KEYS_RENAME_DICT.items():
+            new_key = new_key.replace(replace_key, rename_key)
+        update_state_dict_(original_state_dict, key, new_key)
+
+    for key in list(original_state_dict.keys()):
+        for special_key, handler_fn_inplace in TRANSFORMER_SPECIAL_KEYS_REMAP.items():
+            if special_key not in key:
+                continue
+            handler_fn_inplace(key, original_state_dict)
+
+    transformer.load_state_dict(original_state_dict, strict=True, assign=True)
+    return transformer
+
+
+def convert_vae():
+    vae_ckpt_path = hf_hub_download("Wan-AI/Wan2.1-T2V-14B", "Wan2.1_VAE.pth")
+    old_state_dict = torch.load(vae_ckpt_path, weights_only=True)
+    new_state_dict = {}
+
+    # Create mappings for specific components
+    middle_key_mapping = {
+        # Encoder middle block
+        "encoder.middle.0.residual.0.gamma": "encoder.mid_block.resnets.0.norm1.gamma",
+        "encoder.middle.0.residual.2.bias": "encoder.mid_block.resnets.0.conv1.bias",
+        "encoder.middle.0.residual.2.weight": "encoder.mid_block.resnets.0.conv1.weight",
+        "encoder.middle.0.residual.3.gamma": "encoder.mid_block.resnets.0.norm2.gamma",
+        "encoder.middle.0.residual.6.bias": "encoder.mid_block.resnets.0.conv2.bias",
+        "encoder.middle.0.residual.6.weight": "encoder.mid_block.resnets.0.conv2.weight",
+        "encoder.middle.2.residual.0.gamma": "encoder.mid_block.resnets.1.norm1.gamma",
+        "encoder.middle.2.residual.2.bias": "encoder.mid_block.resnets.1.conv1.bias",
+        "encoder.middle.2.residual.2.weight": "encoder.mid_block.resnets.1.conv1.weight",
+        "encoder.middle.2.residual.3.gamma": "encoder.mid_block.resnets.1.norm2.gamma",
+        "encoder.middle.2.residual.6.bias": "encoder.mid_block.resnets.1.conv2.bias",
+        "encoder.middle.2.residual.6.weight": "encoder.mid_block.resnets.1.conv2.weight",
+        # Decoder middle block
+        "decoder.middle.0.residual.0.gamma": "decoder.mid_block.resnets.0.norm1.gamma",
+        "decoder.middle.0.residual.2.bias": "decoder.mid_block.resnets.0.conv1.bias",
+        "decoder.middle.0.residual.2.weight": "decoder.mid_block.resnets.0.conv1.weight",
+        "decoder.middle.0.residual.3.gamma": "decoder.mid_block.resnets.0.norm2.gamma",
+        "decoder.middle.0.residual.6.bias": "decoder.mid_block.resnets.0.conv2.bias",
+        "decoder.middle.0.residual.6.weight": "decoder.mid_block.resnets.0.conv2.weight",
+        "decoder.middle.2.residual.0.gamma": "decoder.mid_block.resnets.1.norm1.gamma",
+        "decoder.middle.2.residual.2.bias": "decoder.mid_block.resnets.1.conv1.bias",
+        "decoder.middle.2.residual.2.weight": "decoder.mid_block.resnets.1.conv1.weight",
+        "decoder.middle.2.residual.3.gamma": "decoder.mid_block.resnets.1.norm2.gamma",
+        "decoder.middle.2.residual.6.bias": "decoder.mid_block.resnets.1.conv2.bias",
+        "decoder.middle.2.residual.6.weight": "decoder.mid_block.resnets.1.conv2.weight",
+    }
+
+    # Create a mapping for attention blocks
+    attention_mapping = {
+        # Encoder middle attention
+        "encoder.middle.1.norm.gamma": "encoder.mid_block.attentions.0.norm.gamma",
+        "encoder.middle.1.to_qkv.weight": "encoder.mid_block.attentions.0.to_qkv.weight",
+        "encoder.middle.1.to_qkv.bias": "encoder.mid_block.attentions.0.to_qkv.bias",
+        "encoder.middle.1.proj.weight": "encoder.mid_block.attentions.0.proj.weight",
+        "encoder.middle.1.proj.bias": "encoder.mid_block.attentions.0.proj.bias",
+        # Decoder middle attention
+        "decoder.middle.1.norm.gamma": "decoder.mid_block.attentions.0.norm.gamma",
+        "decoder.middle.1.to_qkv.weight": "decoder.mid_block.attentions.0.to_qkv.weight",
+        "decoder.middle.1.to_qkv.bias": "decoder.mid_block.attentions.0.to_qkv.bias",
+        "decoder.middle.1.proj.weight": "decoder.mid_block.attentions.0.proj.weight",
+        "decoder.middle.1.proj.bias": "decoder.mid_block.attentions.0.proj.bias",
+    }
+
+    # Create a mapping for the head components
+    head_mapping = {
+        # Encoder head
+        "encoder.head.0.gamma": "encoder.norm_out.gamma",
+        "encoder.head.2.bias": "encoder.conv_out.bias",
+        "encoder.head.2.weight": "encoder.conv_out.weight",
+        # Decoder head
+        "decoder.head.0.gamma": "decoder.norm_out.gamma",
+        "decoder.head.2.bias": "decoder.conv_out.bias",
+        "decoder.head.2.weight": "decoder.conv_out.weight",
+    }
+
+    # Create a mapping for the quant components
+    quant_mapping = {
+        "conv1.weight": "quant_conv.weight",
+        "conv1.bias": "quant_conv.bias",
+        "conv2.weight": "post_quant_conv.weight",
+        "conv2.bias": "post_quant_conv.bias",
+    }
+
+    # Process each key in the state dict
+    for key, value in old_state_dict.items():
+        # Handle middle block keys using the mapping
+        if key in middle_key_mapping:
+            new_key = middle_key_mapping[key]
+            new_state_dict[new_key] = value
+        # Handle attention blocks using the mapping
+        elif key in attention_mapping:
+            new_key = attention_mapping[key]
+            new_state_dict[new_key] = value
+        # Handle head keys using the mapping
+        elif key in head_mapping:
+            new_key = head_mapping[key]
+            new_state_dict[new_key] = value
+        # Handle quant keys using the mapping
+        elif key in quant_mapping:
+            new_key = quant_mapping[key]
+            new_state_dict[new_key] = value
+        # Handle encoder conv1
+        elif key == "encoder.conv1.weight":
+            new_state_dict["encoder.conv_in.weight"] = value
+        elif key == "encoder.conv1.bias":
+            new_state_dict["encoder.conv_in.bias"] = value
+        # Handle decoder conv1
+        elif key == "decoder.conv1.weight":
+            new_state_dict["decoder.conv_in.weight"] = value
+        elif key == "decoder.conv1.bias":
+            new_state_dict["decoder.conv_in.bias"] = value
+        # Handle encoder downsamples
+        elif key.startswith("encoder.downsamples."):
+            # Convert to down_blocks
+            new_key = key.replace("encoder.downsamples.", "encoder.down_blocks.")
+
+            # Convert residual block naming but keep the original structure
+            if ".residual.0.gamma" in new_key:
+                new_key = new_key.replace(".residual.0.gamma", ".norm1.gamma")
+            elif ".residual.2.bias" in new_key:
+                new_key = new_key.replace(".residual.2.bias", ".conv1.bias")
+            elif ".residual.2.weight" in new_key:
+                new_key = new_key.replace(".residual.2.weight", ".conv1.weight")
+            elif ".residual.3.gamma" in new_key:
+                new_key = new_key.replace(".residual.3.gamma", ".norm2.gamma")
+            elif ".residual.6.bias" in new_key:
+                new_key = new_key.replace(".residual.6.bias", ".conv2.bias")
+            elif ".residual.6.weight" in new_key:
+                new_key = new_key.replace(".residual.6.weight", ".conv2.weight")
+            elif ".shortcut.bias" in new_key:
+                new_key = new_key.replace(".shortcut.bias", ".conv_shortcut.bias")
+            elif ".shortcut.weight" in new_key:
+                new_key = new_key.replace(".shortcut.weight", ".conv_shortcut.weight")
+
+            new_state_dict[new_key] = value
+
+        # Handle decoder upsamples
+        elif key.startswith("decoder.upsamples."):
+            # Convert to up_blocks
+            parts = key.split(".")
+            block_idx = int(parts[2])
+
+            # Group residual blocks
+            if "residual" in key:
+                if block_idx in [0, 1, 2]:
+                    new_block_idx = 0
+                    resnet_idx = block_idx
+                elif block_idx in [4, 5, 6]:
+                    new_block_idx = 1
+                    resnet_idx = block_idx - 4
+                elif block_idx in [8, 9, 10]:
+                    new_block_idx = 2
+                    resnet_idx = block_idx - 8
+                elif block_idx in [12, 13, 14]:
+                    new_block_idx = 3
+                    resnet_idx = block_idx - 12
+                else:
+                    # Keep as is for other blocks
+                    new_state_dict[key] = value
+                    continue
+
+                # Convert residual block naming
+                if ".residual.0.gamma" in key:
+                    new_key = f"decoder.up_blocks.{new_block_idx}.resnets.{resnet_idx}.norm1.gamma"
+                elif ".residual.2.bias" in key:
+                    new_key = f"decoder.up_blocks.{new_block_idx}.resnets.{resnet_idx}.conv1.bias"
+                elif ".residual.2.weight" in key:
+                    new_key = f"decoder.up_blocks.{new_block_idx}.resnets.{resnet_idx}.conv1.weight"
+                elif ".residual.3.gamma" in key:
+                    new_key = f"decoder.up_blocks.{new_block_idx}.resnets.{resnet_idx}.norm2.gamma"
+                elif ".residual.6.bias" in key:
+                    new_key = f"decoder.up_blocks.{new_block_idx}.resnets.{resnet_idx}.conv2.bias"
+                elif ".residual.6.weight" in key:
+                    new_key = f"decoder.up_blocks.{new_block_idx}.resnets.{resnet_idx}.conv2.weight"
+                else:
+                    new_key = key
+
+                new_state_dict[new_key] = value
+
+            # Handle shortcut connections
+            elif ".shortcut." in key:
+                if block_idx == 4:
+                    new_key = key.replace(".shortcut.", ".resnets.0.conv_shortcut.")
+                    new_key = new_key.replace("decoder.upsamples.4", "decoder.up_blocks.1")
+                else:
+                    new_key = key.replace("decoder.upsamples.", "decoder.up_blocks.")
+                    new_key = new_key.replace(".shortcut.", ".conv_shortcut.")
+
+                new_state_dict[new_key] = value
+
+            # Handle upsamplers
+            elif ".resample." in key or ".time_conv." in key:
+                if block_idx == 3:
+                    new_key = key.replace(f"decoder.upsamples.{block_idx}", "decoder.up_blocks.0.upsamplers.0")
+                elif block_idx == 7:
+                    new_key = key.replace(f"decoder.upsamples.{block_idx}", "decoder.up_blocks.1.upsamplers.0")
+                elif block_idx == 11:
+                    new_key = key.replace(f"decoder.upsamples.{block_idx}", "decoder.up_blocks.2.upsamplers.0")
+                else:
+                    new_key = key.replace("decoder.upsamples.", "decoder.up_blocks.")
+
+                new_state_dict[new_key] = value
+            else:
+                new_key = key.replace("decoder.upsamples.", "decoder.up_blocks.")
+                new_state_dict[new_key] = value
+        else:
+            # Keep other keys unchanged
+            new_state_dict[key] = value
+
+    with init_empty_weights():
+        vae = AutoencoderKLWan()
+    vae.load_state_dict(new_state_dict, strict=True, assign=True)
+    return vae
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--model_type", type=str, default=None)
+    parser.add_argument("--output_path", type=str, required=True)
+    parser.add_argument("--dtype", default="fp32")
+    return parser.parse_args()
+
+
+DTYPE_MAPPING = {
+    "fp32": torch.float32,
+    "fp16": torch.float16,
+    "bf16": torch.bfloat16,
+}
+
+
+if __name__ == "__main__":
+    args = get_args()
+
+    transformer = None
+    dtype = DTYPE_MAPPING[args.dtype]
+
+    transformer = convert_transformer(args.model_type).to(dtype=dtype)
+    vae = convert_vae()
+    text_encoder = UMT5EncoderModel.from_pretrained("google/umt5-xxl")
+    tokenizer = AutoTokenizer.from_pretrained("google/umt5-xxl")
+    scheduler = UniPCMultistepScheduler(
+        prediction_type="flow_prediction", use_flow_sigmas=True, num_train_timesteps=1000, flow_shift=3.0
+    )
+
+    if "I2V" in args.model_type:
+        image_encoder = CLIPVisionModelWithProjection.from_pretrained(
+            "laion/CLIP-ViT-H-14-laion2B-s32B-b79K", torch_dtype=torch.bfloat16
+        )
+        image_processor = AutoProcessor.from_pretrained("laion/CLIP-ViT-H-14-laion2B-s32B-b79K")
+        pipe = WanImageToVideoPipeline(
+            transformer=transformer,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+            vae=vae,
+            scheduler=scheduler,
+            image_encoder=image_encoder,
+            image_processor=image_processor,
+        )
+    else:
+        pipe = WanPipeline(
+            transformer=transformer,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+            vae=vae,
+            scheduler=scheduler,
+        )
+
+    pipe.save_pretrained(args.output_path, safe_serialization=True, max_shard_size="5GB")

src/diffusers/__init__.py

@@ -96,6 +96,7 @@ else:
             "AutoencoderKLLTXVideo",
             "AutoencoderKLMochi",
             "AutoencoderKLTemporalDecoder",
+            "AutoencoderKLWan",
             "AutoencoderOobleck",
             "AutoencoderTiny",
             "CacheMixin",
@@ -148,6 +149,7 @@ else:
             "UNetSpatioTemporalConditionModel",
             "UVit2DModel",
             "VQModel",
+            "WanTransformer3DModel",
         ]
     )
     _import_structure["optimization"] = [
@@ -438,6 +440,8 @@ else:
             "VersatileDiffusionTextToImagePipeline",
             "VideoToVideoSDPipeline",
             "VQDiffusionPipeline",
+            "WanImageToVideoPipeline",
+            "WanPipeline",
             "WuerstchenCombinedPipeline",
             "WuerstchenDecoderPipeline",
             "WuerstchenPriorPipeline",
@@ -618,6 +622,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             AutoencoderKLLTXVideo,
             AutoencoderKLMochi,
             AutoencoderKLTemporalDecoder,
+            AutoencoderKLWan,
             AutoencoderOobleck,
             AutoencoderTiny,
             CacheMixin,
@@ -669,6 +674,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             UNetSpatioTemporalConditionModel,
             UVit2DModel,
             VQModel,
+            WanTransformer3DModel,
         )
         from .optimization import (
             get_constant_schedule,
@@ -938,6 +944,8 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             VersatileDiffusionTextToImagePipeline,
             VideoToVideoSDPipeline,
             VQDiffusionPipeline,
+            WanImageToVideoPipeline,
+            WanPipeline,
             WuerstchenCombinedPipeline,
             WuerstchenDecoderPipeline,
             WuerstchenPriorPipeline,

src/diffusers/models/__init__.py

@@ -35,6 +35,7 @@ if is_torch_available():
     _import_structure["autoencoders.autoencoder_kl_ltx"] = ["AutoencoderKLLTXVideo"]
     _import_structure["autoencoders.autoencoder_kl_mochi"] = ["AutoencoderKLMochi"]
     _import_structure["autoencoders.autoencoder_kl_temporal_decoder"] = ["AutoencoderKLTemporalDecoder"]
+    _import_structure["autoencoders.autoencoder_kl_wan"] = ["AutoencoderKLWan"]
     _import_structure["autoencoders.autoencoder_oobleck"] = ["AutoencoderOobleck"]
     _import_structure["autoencoders.autoencoder_tiny"] = ["AutoencoderTiny"]
     _import_structure["autoencoders.consistency_decoder_vae"] = ["ConsistencyDecoderVAE"]
@@ -79,6 +80,7 @@ if is_torch_available():
     _import_structure["transformers.transformer_omnigen"] = ["OmniGenTransformer2DModel"]
     _import_structure["transformers.transformer_sd3"] = ["SD3Transformer2DModel"]
     _import_structure["transformers.transformer_temporal"] = ["TransformerTemporalModel"]
+    _import_structure["transformers.transformer_wan"] = ["WanTransformer3DModel"]
     _import_structure["unets.unet_1d"] = ["UNet1DModel"]
     _import_structure["unets.unet_2d"] = ["UNet2DModel"]
     _import_structure["unets.unet_2d_condition"] = ["UNet2DConditionModel"]
@@ -109,6 +111,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             AutoencoderKLLTXVideo,
             AutoencoderKLMochi,
             AutoencoderKLTemporalDecoder,
+            AutoencoderKLWan,
             AutoencoderOobleck,
             AutoencoderTiny,
             ConsistencyDecoderVAE,
@@ -158,6 +161,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             T5FilmDecoder,
             Transformer2DModel,
             TransformerTemporalModel,
+            WanTransformer3DModel,
         )
         from .unets import (
             I2VGenXLUNet,

src/diffusers/models/attention_processor.py

@@ -280,6 +280,10 @@ class Attention(nn.Module):
             elif qk_norm == "rms_norm":
                 self.norm_added_q = RMSNorm(dim_head, eps=eps)
                 self.norm_added_k = RMSNorm(dim_head, eps=eps)
+            elif qk_norm == "rms_norm_across_heads":
+                # Wanx applies qk norm across all heads
+                self.norm_added_q = RMSNorm(dim_head * heads, eps=eps)
+                self.norm_added_k = RMSNorm(dim_head * kv_heads, eps=eps)
             else:
                 raise ValueError(
                     f"unknown qk_norm: {qk_norm}. Should be one of `None,'layer_norm','fp32_layer_norm','rms_norm'`"

src/diffusers/models/autoencoders/__init__.py

@@ -7,6 +7,7 @@ from .autoencoder_kl_hunyuan_video import AutoencoderKLHunyuanVideo
 from .autoencoder_kl_ltx import AutoencoderKLLTXVideo
 from .autoencoder_kl_mochi import AutoencoderKLMochi
 from .autoencoder_kl_temporal_decoder import AutoencoderKLTemporalDecoder
+from .autoencoder_kl_wan import AutoencoderKLWan
 from .autoencoder_oobleck import AutoencoderOobleck
 from .autoencoder_tiny import AutoencoderTiny
 from .consistency_decoder_vae import ConsistencyDecoderVAE

src/diffusers/models/modeling_utils.py

@@ -166,8 +166,12 @@ def get_parameter_dtype(parameter: torch.nn.Module) -> torch.dtype:
 
     # 2. If no dtype modifying hooks are attached, return the dtype of the first floating point parameter/buffer
     last_dtype = None
-    for param in parameter.parameters():
+
+    for name, param in parameter.named_parameters():
         last_dtype = param.dtype
+        if parameter._keep_in_fp32_modules and any(m in name for m in parameter._keep_in_fp32_modules):
+            continue
+
         if param.is_floating_point():
             return param.dtype
 

src/diffusers/models/transformers/__init__.py

@@ -27,3 +27,4 @@ if is_torch_available():
     from .transformer_omnigen import OmniGenTransformer2DModel
     from .transformer_sd3 import SD3Transformer2DModel
     from .transformer_temporal import TransformerTemporalModel
+    from .transformer_wan import WanTransformer3DModel

src/diffusers/models/transformers/transformer_wan.py

+# Copyright 2025 The Wan 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 math
+from typing import Dict, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from ...configuration_utils import ConfigMixin, register_to_config
+from ...utils import logging
+from ..attention import FeedForward
+from ..attention_processor import Attention
+from ..embeddings import PixArtAlphaTextProjection, TimestepEmbedding, Timesteps, get_1d_rotary_pos_embed
+from ..modeling_outputs import Transformer2DModelOutput
+from ..modeling_utils import ModelMixin
+from ..normalization import FP32LayerNorm
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+class WanAttnProcessor2_0:
+    def __init__(self):
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError("WanAttnProcessor2_0 requires PyTorch 2.0. To use it, please upgrade PyTorch to 2.0.")
+
+    def __call__(
+        self,
+        attn: Attention,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        rotary_emb: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        encoder_hidden_states_img = None
+        if attn.add_k_proj is not None:
+            encoder_hidden_states_img = encoder_hidden_states[:, :257]
+            encoder_hidden_states = encoder_hidden_states[:, 257:]
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+
+        query = attn.to_q(hidden_states)
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        if attn.norm_q is not None:
+            query = attn.norm_q(query)
+        if attn.norm_k is not None:
+            key = attn.norm_k(key)
+
+        query = query.unflatten(2, (attn.heads, -1)).transpose(1, 2)
+        key = key.unflatten(2, (attn.heads, -1)).transpose(1, 2)
+        value = value.unflatten(2, (attn.heads, -1)).transpose(1, 2)
+
+        if rotary_emb is not None:
+
+            def apply_rotary_emb(hidden_states: torch.Tensor, freqs: torch.Tensor):
+                x_rotated = torch.view_as_complex(hidden_states.to(torch.float64).unflatten(3, (-1, 2)))
+                x_out = torch.view_as_real(x_rotated * freqs).flatten(3, 4)
+                return x_out.type_as(hidden_states)
+
+            query = apply_rotary_emb(query, rotary_emb)
+            key = apply_rotary_emb(key, rotary_emb)
+
+        # I2V task
+        hidden_states_img = None
+        if encoder_hidden_states_img is not None:
+            key_img = attn.add_k_proj(encoder_hidden_states_img)
+            key_img = attn.norm_added_k(key_img)
+            value_img = attn.add_v_proj(encoder_hidden_states_img)
+
+            key_img = key_img.unflatten(2, (attn.heads, -1)).transpose(1, 2)
+            value_img = value_img.unflatten(2, (attn.heads, -1)).transpose(1, 2)
+
+            hidden_states_img = F.scaled_dot_product_attention(
+                query, key_img, value_img, attn_mask=None, dropout_p=0.0, is_causal=False
+            )
+            hidden_states_img = hidden_states_img.transpose(1, 2).flatten(2, 3)
+            hidden_states_img = hidden_states_img.type_as(query)
+
+        hidden_states = F.scaled_dot_product_attention(
+            query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
+        )
+        hidden_states = hidden_states.transpose(1, 2).flatten(2, 3)
+        hidden_states = hidden_states.type_as(query)
+
+        if hidden_states_img is not None:
+            hidden_states = hidden_states + hidden_states_img
+
+        hidden_states = attn.to_out[0](hidden_states)
+        hidden_states = attn.to_out[1](hidden_states)
+        return hidden_states
+
+
+class WanImageEmbedding(torch.nn.Module):
+    def __init__(self, in_features: int, out_features: int):
+        super().__init__()
+
+        self.norm1 = nn.LayerNorm(in_features)
+        self.ff = FeedForward(in_features, out_features, mult=1, activation_fn="gelu")
+        self.norm2 = nn.LayerNorm(out_features)
+
+    def forward(self, encoder_hidden_states_image: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.norm1(encoder_hidden_states_image)
+        hidden_states = self.ff(hidden_states)
+        hidden_states = self.norm2(hidden_states)
+        return hidden_states
+
+
+class WanTimeTextImageEmbedding(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        time_freq_dim: int,
+        time_proj_dim: int,
+        text_embed_dim: int,
+        image_embed_dim: Optional[int] = None,
+    ):
+        super().__init__()
+
+        self.timesteps_proj = Timesteps(num_channels=time_freq_dim, flip_sin_to_cos=True, downscale_freq_shift=0)
+        self.time_embedder = TimestepEmbedding(in_channels=time_freq_dim, time_embed_dim=dim)
+        self.act_fn = nn.SiLU()
+        self.time_proj = nn.Linear(dim, time_proj_dim)
+        self.text_embedder = PixArtAlphaTextProjection(text_embed_dim, dim, act_fn="gelu_tanh")
+
+        self.image_embedder = None
+        if image_embed_dim is not None:
+            self.image_embedder = WanImageEmbedding(image_embed_dim, dim)
+
+    def forward(
+        self,
+        timestep: torch.Tensor,
+        encoder_hidden_states: torch.Tensor,
+        encoder_hidden_states_image: Optional[torch.Tensor] = None,
+    ):
+        timestep = self.timesteps_proj(timestep)
+
+        time_embedder_dtype = next(iter(self.time_embedder.parameters())).dtype
+        if timestep.dtype != time_embedder_dtype and time_embedder_dtype != torch.int8:
+            timestep = timestep.to(time_embedder_dtype)
+        temb = self.time_embedder(timestep).type_as(encoder_hidden_states)
+        timestep_proj = self.time_proj(self.act_fn(temb))
+
+        encoder_hidden_states = self.text_embedder(encoder_hidden_states)
+        if encoder_hidden_states_image is not None:
+            encoder_hidden_states_image = self.image_embedder(encoder_hidden_states_image)
+
+        return temb, timestep_proj, encoder_hidden_states, encoder_hidden_states_image
+
+
+class WanRotaryPosEmbed(nn.Module):
+    def __init__(
+        self, attention_head_dim: int, patch_size: Tuple[int, int, int], max_seq_len: int, theta: float = 10000.0
+    ):
+        super().__init__()
+
+        self.attention_head_dim = attention_head_dim
+        self.patch_size = patch_size
+        self.max_seq_len = max_seq_len
+
+        h_dim = w_dim = 2 * (attention_head_dim // 6)
+        t_dim = attention_head_dim - h_dim - w_dim
+
+        freqs = []
+        for dim in [t_dim, h_dim, w_dim]:
+            freq = get_1d_rotary_pos_embed(
+                dim, max_seq_len, theta, use_real=False, repeat_interleave_real=False, freqs_dtype=torch.float64
+            )
+            freqs.append(freq)
+        self.freqs = torch.cat(freqs, dim=1)
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        batch_size, num_channels, num_frames, height, width = hidden_states.shape
+        p_t, p_h, p_w = self.patch_size
+        ppf, pph, ppw = num_frames // p_t, height // p_h, width // p_w
+
+        self.freqs = self.freqs.to(hidden_states.device)
+        freqs = self.freqs.split_with_sizes(
+            [
+                self.attention_head_dim // 2 - 2 * (self.attention_head_dim // 6),
+                self.attention_head_dim // 6,
+                self.attention_head_dim // 6,
+            ],
+            dim=1,
+        )
+
+        freqs_f = freqs[0][:ppf].view(ppf, 1, 1, -1).expand(ppf, pph, ppw, -1)
+        freqs_h = freqs[1][:pph].view(1, pph, 1, -1).expand(ppf, pph, ppw, -1)
+        freqs_w = freqs[2][:ppw].view(1, 1, ppw, -1).expand(ppf, pph, ppw, -1)
+        freqs = torch.cat([freqs_f, freqs_h, freqs_w], dim=-1).reshape(1, 1, ppf * pph * ppw, -1)
+        return freqs
+
+
+class WanTransformerBlock(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        ffn_dim: int,
+        num_heads: int,
+        qk_norm: str = "rms_norm_across_heads",
+        cross_attn_norm: bool = False,
+        eps: float = 1e-6,
+        added_kv_proj_dim: Optional[int] = None,
+    ):
+        super().__init__()
+
+        # 1. Self-attention
+        self.norm1 = FP32LayerNorm(dim, eps, elementwise_affine=False)
+        self.attn1 = Attention(
+            query_dim=dim,
+            heads=num_heads,
+            kv_heads=num_heads,
+            dim_head=dim // num_heads,
+            qk_norm=qk_norm,
+            eps=eps,
+            bias=True,
+            cross_attention_dim=None,
+            out_bias=True,
+            processor=WanAttnProcessor2_0(),
+        )
+
+        # 2. Cross-attention
+        self.attn2 = Attention(
+            query_dim=dim,
+            heads=num_heads,
+            kv_heads=num_heads,
+            dim_head=dim // num_heads,
+            qk_norm=qk_norm,
+            eps=eps,
+            bias=True,
+            cross_attention_dim=None,
+            out_bias=True,
+            added_kv_proj_dim=added_kv_proj_dim,
+            added_proj_bias=True,
+            processor=WanAttnProcessor2_0(),
+        )
+        self.norm2 = FP32LayerNorm(dim, eps, elementwise_affine=True) if cross_attn_norm else nn.Identity()
+
+        # 3. Feed-forward
+        self.ffn = FeedForward(dim, inner_dim=ffn_dim, activation_fn="gelu-approximate")
+        self.norm3 = FP32LayerNorm(dim, eps, elementwise_affine=False)
+
+        self.scale_shift_table = nn.Parameter(torch.randn(1, 6, dim) / dim**0.5)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: torch.Tensor,
+        temb: torch.Tensor,
+        rotary_emb: torch.Tensor,
+    ) -> torch.Tensor:
+        shift_msa, scale_msa, gate_msa, c_shift_msa, c_scale_msa, c_gate_msa = (
+            self.scale_shift_table + temb.float()
+        ).chunk(6, dim=1)
+
+        # 1. Self-attention
+        norm_hidden_states = (self.norm1(hidden_states.float()) * (1 + scale_msa) + shift_msa).type_as(hidden_states)
+        attn_output = self.attn1(hidden_states=norm_hidden_states, rotary_emb=rotary_emb)
+        hidden_states = (hidden_states.float() + attn_output * gate_msa).type_as(hidden_states)
+
+        # 2. Cross-attention
+        norm_hidden_states = self.norm2(hidden_states.float()).type_as(hidden_states)
+        attn_output = self.attn2(hidden_states=norm_hidden_states, encoder_hidden_states=encoder_hidden_states)
+        hidden_states = hidden_states + attn_output
+
+        # 3. Feed-forward
+        norm_hidden_states = (self.norm3(hidden_states.float()) * (1 + c_scale_msa) + c_shift_msa).type_as(
+            hidden_states
+        )
+        ff_output = self.ffn(norm_hidden_states)
+        hidden_states = (hidden_states.float() + ff_output.float() * c_gate_msa).type_as(hidden_states)
+
+        return hidden_states
+
+
+class WanTransformer3DModel(ModelMixin, ConfigMixin):
+    r"""
+    A Transformer model for video-like data used in the Wan model.
+
+    Args:
+        patch_size (`Tuple[int]`, defaults to `(1, 2, 2)`):
+            3D patch dimensions for video embedding (t_patch, h_patch, w_patch).
+        num_attention_heads (`int`, defaults to `40`):
+            Fixed length for text embeddings.
+        attention_head_dim (`int`, defaults to `128`):
+            The number of channels in each head.
+        in_channels (`int`, defaults to `16`):
+            The number of channels in the input.
+        out_channels (`int`, defaults to `16`):
+            The number of channels in the output.
+        text_dim (`int`, defaults to `512`):
+            Input dimension for text embeddings.
+        freq_dim (`int`, defaults to `256`):
+            Dimension for sinusoidal time embeddings.
+        ffn_dim (`int`, defaults to `13824`):
+            Intermediate dimension in feed-forward network.
+        num_layers (`int`, defaults to `40`):
+            The number of layers of transformer blocks to use.
+        window_size (`Tuple[int]`, defaults to `(-1, -1)`):
+            Window size for local attention (-1 indicates global attention).
+        cross_attn_norm (`bool`, defaults to `True`):
+            Enable cross-attention normalization.
+        qk_norm (`bool`, defaults to `True`):
+            Enable query/key normalization.
+        eps (`float`, defaults to `1e-6`):
+            Epsilon value for normalization layers.
+        add_img_emb (`bool`, defaults to `False`):
+            Whether to use img_emb.
+        added_kv_proj_dim (`int`, *optional*, defaults to `None`):
+            The number of channels to use for the added key and value projections. If `None`, no projection is used.
+    """
+
+    _supports_gradient_checkpointing = True
+    _skip_layerwise_casting_patterns = ["patch_embedding", "condition_embedder", "norm"]
+    _no_split_modules = ["WanTransformerBlock"]
+    _keep_in_fp32_modules = ["time_embedder", "scale_shift_table", "norm1", "norm2", "norm3"]
+
+    @register_to_config
+    def __init__(
+        self,
+        patch_size: Tuple[int] = (1, 2, 2),
+        num_attention_heads: int = 40,
+        attention_head_dim: int = 128,
+        in_channels: int = 16,
+        out_channels: int = 16,
+        text_dim: int = 4096,
+        freq_dim: int = 256,
+        ffn_dim: int = 13824,
+        num_layers: int = 40,
+        cross_attn_norm: bool = True,
+        qk_norm: Optional[str] = "rms_norm_across_heads",
+        eps: float = 1e-6,
+        image_dim: Optional[int] = None,
+        added_kv_proj_dim: Optional[int] = None,
+        rope_max_seq_len: int = 1024,
+    ) -> None:
+        super().__init__()
+
+        inner_dim = num_attention_heads * attention_head_dim
+        out_channels = out_channels or in_channels
+
+        # 1. Patch & position embedding
+        self.rope = WanRotaryPosEmbed(attention_head_dim, patch_size, rope_max_seq_len)
+        self.patch_embedding = nn.Conv3d(in_channels, inner_dim, kernel_size=patch_size, stride=patch_size)
+
+        # 2. Condition embeddings
+        # image_embedding_dim=1280 for I2V model
+        self.condition_embedder = WanTimeTextImageEmbedding(
+            dim=inner_dim,
+            time_freq_dim=freq_dim,
+            time_proj_dim=inner_dim * 6,
+            text_embed_dim=text_dim,
+            image_embed_dim=image_dim,
+        )
+
+        # 3. Transformer blocks
+        self.blocks = nn.ModuleList(
+            [
+                WanTransformerBlock(
+                    inner_dim, ffn_dim, num_attention_heads, qk_norm, cross_attn_norm, eps, added_kv_proj_dim
+                )
+                for _ in range(num_layers)
+            ]
+        )
+
+        # 4. Output norm & projection
+        self.norm_out = FP32LayerNorm(inner_dim, eps, elementwise_affine=False)
+        self.proj_out = nn.Linear(inner_dim, out_channels * math.prod(patch_size))
+        self.scale_shift_table = nn.Parameter(torch.randn(1, 2, inner_dim) / inner_dim**0.5)
+
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        timestep: torch.LongTensor,
+        encoder_hidden_states: torch.Tensor,
+        encoder_hidden_states_image: Optional[torch.Tensor] = None,
+        return_dict: bool = True,
+    ) -> Union[torch.Tensor, Dict[str, torch.Tensor]]:
+        batch_size, num_channels, num_frames, height, width = hidden_states.shape
+        p_t, p_h, p_w = self.config.patch_size
+        post_patch_num_frames = num_frames // p_t
+        post_patch_height = height // p_h
+        post_patch_width = width // p_w
+
+        rotary_emb = self.rope(hidden_states)
+
+        hidden_states = self.patch_embedding(hidden_states)
+        hidden_states = hidden_states.flatten(2).transpose(1, 2)
+
+        temb, timestep_proj, encoder_hidden_states, encoder_hidden_states_image = self.condition_embedder(
+            timestep, encoder_hidden_states, encoder_hidden_states_image
+        )
+        timestep_proj = timestep_proj.unflatten(1, (6, -1))
+
+        if encoder_hidden_states_image is not None:
+            encoder_hidden_states = torch.concat([encoder_hidden_states_image, encoder_hidden_states], dim=1)
+
+        # 4. Transformer blocks
+        if torch.is_grad_enabled() and self.gradient_checkpointing:
+            for block in self.blocks:
+                hidden_states = self._gradient_checkpointing_func(
+                    block, hidden_states, encoder_hidden_states, timestep_proj, rotary_emb
+                )
+        else:
+            for block in self.blocks:
+                hidden_states = block(hidden_states, encoder_hidden_states, timestep_proj, rotary_emb)
+
+        # 5. Output norm, projection & unpatchify
+        shift, scale = (self.scale_shift_table + temb.unsqueeze(1)).chunk(2, dim=1)
+        hidden_states = (self.norm_out(hidden_states.float()) * (1 + scale) + shift).type_as(hidden_states)
+        hidden_states = self.proj_out(hidden_states)
+
+        hidden_states = hidden_states.reshape(
+            batch_size, post_patch_num_frames, post_patch_height, post_patch_width, p_t, p_h, p_w, -1
+        )
+        hidden_states = hidden_states.permute(0, 7, 1, 4, 2, 5, 3, 6)
+        output = hidden_states.flatten(6, 7).flatten(4, 5).flatten(2, 3)
+
+        if not return_dict:
+            return (output,)
+
+        return Transformer2DModelOutput(sample=output)

src/diffusers/pipelines/__init__.py

@@ -347,6 +347,7 @@ else:
         "WuerstchenDecoderPipeline",
         "WuerstchenPriorPipeline",
     ]
+    _import_structure["wan"] = ["WanPipeline", "WanImageToVideoPipeline"]
 try:
     if not is_onnx_available():
         raise OptionalDependencyNotAvailable()
@@ -690,6 +691,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             UniDiffuserPipeline,
             UniDiffuserTextDecoder,
         )
+        from .wan import WanImageToVideoPipeline, WanPipeline
         from .wuerstchen import (
             WuerstchenCombinedPipeline,
             WuerstchenDecoderPipeline,

src/diffusers/pipelines/wan/__init__.py

+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"]
+
+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
+
+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)

src/diffusers/pipelines/wan/pipeline_output.py

+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

src/diffusers/utils/dummy_pt_objects.py

@@ -201,6 +201,21 @@ class AutoencoderKLTemporalDecoder(metaclass=DummyObject):
         requires_backends(cls, ["torch"])
 
 
+class AutoencoderKLWan(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 AutoencoderOobleck(metaclass=DummyObject):
     _backends = ["torch"]
 
@@ -966,6 +981,21 @@ class VQModel(metaclass=DummyObject):
         requires_backends(cls, ["torch"])
 
 
+class WanTransformer3DModel(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"])
+
+
 def get_constant_schedule(*args, **kwargs):
     requires_backends(get_constant_schedule, ["torch"])
 

src/diffusers/utils/dummy_torch_and_transformers_objects.py

@@ -2597,6 +2597,36 @@ class VQDiffusionPipeline(metaclass=DummyObject):
         requires_backends(cls, ["torch", "transformers"])
 
 
+class WanImageToVideoPipeline(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 WanPipeline(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 WuerstchenCombinedPipeline(metaclass=DummyObject):
     _backends = ["torch", "transformers"]