[!IMPORTANT] Naming notice (2026-04-10). The "HLWQ" technique used in this model is being rebranded to HLWQ (Hadamard-Lloyd Weight Quantization). The change is only the name; the algorithm and the weights in this repository are unchanged.

The rebrand resolves a name collision with an unrelated, earlier KV cache quantization method also named HLWQ (Han et al., arXiv:2502.02617, 2025). HLWQ addresses weight quantization with a deterministic Walsh-Hadamard rotation and Lloyd-Max scalar codebook; Han et al.'s HLWQ addresses KV cache quantization with a random polar rotation. The two methods are technically distinct.

Existing loaders that load this repository by ID continue to work without changes. Future model uploads will use the HLWQ name.

Reference paper for this technique: arXiv:2603.29078 (v2 in preparation; v1 still uses the old name).

Wan2.2

Wan: Open and Advanced Large-Scale Video Generative Models

We are excited to introduce Wan2.2, a major upgrade to our foundational video models. With Wan2.2, we have focused on incorporating the following innovations:

👍 Effective MoE Architecture: Wan2.2 introduces a Mixture-of-Experts (MoE) architecture into video diffusion models. By separating the denoising process cross timesteps with specialized powerful expert models, this enlarges the overall model capacity while maintaining the same computational cost.
👍 Cinematic-level Aesthetics: Wan2.2 incorporates meticulously curated aesthetic data, complete with detailed labels for lighting, composition, contrast, color tone, and more. This allows for more precise and controllable cinematic style generation, facilitating the creation of videos with customizable aesthetic preferences.
👍 Complex Motion Generation: Compared to Wan2.1, Wan2.2 is trained on a significantly larger data, with +65.6% more images and +83.2% more videos. This expansion notably enhances the model's generalization across multiple dimensions such as motions, semantics, and aesthetics, achieving TOP performance among all open-sourced and closed-sourced models.
👍 Efficient High-Definition Hybrid TI2V: Wan2.2 open-sources a 5B model built with our advanced Wan2.2-VAE that achieves a compression ratio of 16×16×4. This model supports both text-to-video and image-to-video generation at 720P resolution with 24fps and can also run on consumer-grade graphics cards like 4090. It is one of the fastest 720P@24fps models currently available, capable of serving both the industrial and academic sectors simultaneously.

Video Demos

Your browser does not support the video tag.

🔥 Latest News!!

Sep 19, 2025: 💃 We introduct Wan2.2-Animate-14B, an unified model for character animation and replacement with holistic movement and expression replication. We released the model weights and inference code. And now you can try it on wan.video, ModelScope Studio or HuggingFace Space!
Aug 26, 2025: 🎵 We introduce Wan2.2-S2V-14B, an audio-driven cinematic video generation model, including inference code, model weights, and technical report! Now you can try it on wan.video, ModelScope Gradio or HuggingFace Gradio!
Jul 28, 2025: 👋 We have open a HF space using the TI2V-5B model. Enjoy!
Jul 28, 2025: 👋 Wan2.2 has been integrated into ComfyUI (CN | EN). Enjoy!
Jul 28, 2025: 👋 Wan2.2's T2V, I2V and TI2V have been integrated into Diffusers (T2V-A14B | I2V-A14B | TI2V-5B). Feel free to give it a try!
Jul 28, 2025: 👋 We've released the inference code and model weights of Wan2.2.
Sep 5, 2025: 👋 We add text-to-speech synthesis support with CosyVoice for Speech-to-Video generation task.

Community Works

If your research or project builds upon Wan2.1 or Wan2.2, and you would like more people to see it, please inform us.

DiffSynth-Studio provides comprehensive support for Wan 2.2, including low-GPU-memory layer-by-layer offload, FP8 quantization, sequence parallelism, LoRA training, full training.
Kijai's ComfyUI WanVideoWrapper is an alternative implementation of Wan models for ComfyUI. Thanks to its Wan-only focus, it's on the frontline of getting cutting edge optimizations and hot research features, which are often hard to integrate into ComfyUI quickly due to its more rigid structure.
Cache-dit offers Fully Cache Acceleration support for Wan2.2 MoE with DBCache, TaylorSeer and Cache CFG. Visit their example for more details.
FastVideo includes distilled Wan models with sparse attention that significanly speed up the inference time.

📑 Todo List

Wan2.2 Text-to-Video
- Multi-GPU Inference code of the A14B and 14B models
- Checkpoints of the A14B and 14B models
- ComfyUI integration
- Diffusers integration
Wan2.2 Image-to-Video
- Multi-GPU Inference code of the A14B model
- Checkpoints of the A14B model
- ComfyUI integration
- Diffusers integration
Wan2.2 Text-Image-to-Video
- Multi-GPU Inference code of the 5B model
- Checkpoints of the 5B model
- ComfyUI integration
- Diffusers integration
Wan2.2-S2V Speech-to-Video
- Inference code of Wan2.2-S2V
- Checkpoints of Wan2.2-S2V-14B
- ComfyUI integration
- Diffusers integration
Wan2.2-Animate Character Animation and Replacement
- Inference code of Wan2.2-Animate
- Checkpoints of Wan2.2-Animate
- ComfyUI integration
- Diffusers integration

Run Wan2.2 Animate

Installation

Clone the repo:

git clone https://github.com/Wan-Video/Wan2.2.git
cd Wan2.2

Install dependencies:

# Ensure torch >= 2.4.0
# If the installation of `flash_attn` fails, try installing the other packages first and install `flash_attn` last
pip install -r requirements.txt
# If you want to use CosyVoice to synthesize speech for Speech-to-Video Generation, please install requirements_s2v.txt additionally
pip install -r requirements_s2v.txt

Model Download

Models	Download Links	Description
T2V-A14B	🤗 Huggingface 🤖 ModelScope	Text-to-Video MoE model, supports 480P & 720P
I2V-A14B	🤗 Huggingface 🤖 ModelScope	Image-to-Video MoE model, supports 480P & 720P
TI2V-5B	🤗 Huggingface 🤖 ModelScope	High-compression VAE, T2V+I2V, supports 720P
S2V-14B	🤗 Huggingface 🤖 ModelScope	Speech-to-Video model, supports 480P & 720P
Animate-14B	🤗 Huggingface 🤖 ModelScope	Character animation and replacement

Download models using huggingface-cli:

pip install "huggingface_hub[cli]"
huggingface-cli download Wan-AI/Wan2.2-Animate-14B --local-dir ./Wan2.2-Animate-14B

Download models using modelscope-cli:

pip install modelscope
modelscope download Wan-AI/Wan2.2-Animate-14B --local_dir ./Wan2.2-Animate-14B

Run Wan-Animate-14B

Wan-Animate takes a video and a character image as input, and generates a video in either "animation" or "replacement" mode.

animation mode： The model generates a video of the character image that mimics the human motion in the input video.
replacement mode: The model replaces the character image with the input video.

Please visit our project page to see more examples and learn about the scenarios suitable for this model.

(1) Preprocessing

The input video should be preprocessed into several materials before be feed into the inference process. Please refer to the following processing flow, and more details about preprocessing can be found in UserGuider.

For animation

python ./wan/modules/animate/preprocess/preprocess_data.py \
    --ckpt_path ./Wan2.2-Animate-14B/process_checkpoint \
    --video_path ./examples/wan_animate/animate/video.mp4 \
    --refer_path ./examples/wan_animate/animate/image.jpeg \
    --save_path ./examples/wan_animate/animate/process_results \
    --resolution_area 1280 720 \
    --retarget_flag \
    --use_flux

For replacement

python ./wan/modules/animate/preprocess/preprocess_data.py \
    --ckpt_path ./Wan2.2-Animate-14B/process_checkpoint \
    --video_path ./examples/wan_animate/replace/video.mp4 \
    --refer_path ./examples/wan_animate/replace/image.jpeg \
    --save_path ./examples/wan_animate/replace/process_results \
    --resolution_area 1280 720 \
    --iterations 3 \
    --k 7 \
    --w_len 1 \
    --h_len 1 \
    --replace_flag

(2) Run in animation mode

Single-GPU inference

python generate.py --task animate-14B --ckpt_dir ./Wan2.2-Animate-14B/ --src_root_path ./examples/wan_animate/animate/process_results/ --refert_num 1

Multi-GPU inference using FSDP + DeepSpeed Ulysses

python -m torch.distributed.run --nnodes 1 --nproc_per_node 8 generate.py --task animate-14B --ckpt_dir ./Wan2.2-Animate-14B/ --src_root_path ./examples/wan_animate/animate/process_results/ --refert_num 1 --dit_fsdp --t5_fsdp --ulysses_size 8

(3) Run in replacement mode

Single-GPU inference

python generate.py --task animate-14B --ckpt_dir ./Wan2.2-Animate-14B/ --src_root_path ./examples/wan_animate/replace/process_results/ --refert_num 1 --replace_flag --use_relighting_lora

Multi-GPU inference using FSDP + DeepSpeed Ulysses

python -m torch.distributed.run --nnodes 1 --nproc_per_node 8 generate.py --task animate-14B --ckpt_dir ./Wan2.2-Animate-14B/ --src_root_path ./examples/wan_animate/replace/process_results/src_pose.mp4  --refert_num 1 --replace_flag --use_relighting_lora --dit_fsdp --t5_fsdp --ulysses_size 8

💡 If you're using Wan-Animate, we do not recommend using LoRA models trained on Wan2.2, since weight changes during training may lead to unexpected behavior.

Computational Efficiency on Different GPUs

We test the computational efficiency of different Wan2.2 models on different GPUs in the following table. The results are presented in the format: Total time (s) / peak GPU memory (GB).

The parameter settings for the tests presented in this table are as follows: (1) Multi-GPU: 14B: --ulysses_size 4/8 --dit_fsdp --t5_fsdp, 5B: --ulysses_size 4/8 --offload_model True --convert_model_dtype --t5_cpu; Single-GPU: 14B: --offload_model True --convert_model_dtype, 5B: --offload_model True --convert_model_dtype --t5_cpu (--convert_model_dtype converts model parameter types to config.param_dtype); (2) The distributed testing utilizes the built-in FSDP and Ulysses implementations, with FlashAttention3 deployed on Hopper architecture GPUs; (3) Tests were run without the --use_prompt_extend flag; (4) Reported results are the average of multiple samples taken after the warm-up phase.

Introduction of Wan2.2

Wan2.2 builds on the foundation of Wan2.1 with notable improvements in generation quality and model capability. This upgrade is driven by a series of key technical innovations, mainly including the Mixture-of-Experts (MoE) architecture, upgraded training data, and high-compression video generation.

(1) Mixture-of-Experts (MoE) Architecture

Wan2.2 introduces Mixture-of-Experts (MoE) architecture into the video generation diffusion model. MoE has been widely validated in large language models as an efficient approach to increase total model parameters while keeping inference cost nearly unchanged. In Wan2.2, the A14B model series adopts a two-expert design tailored to the denoising process of diffusion models: a high-noise expert for the early stages, focusing on overall layout; and a low-noise expert for the later stages, refining video details. Each expert model has about 14B parameters, resulting in a total of 27B parameters but only 14B active parameters per step, keeping inference computation and GPU memory nearly unchanged.

The transition point between the two experts is determined by the signal-to-noise ratio (SNR), a metric that decreases monotonically as the denoising step $t$ increases. At the beginning of the denoising process, $t$ is large and the noise level is high, so the SNR is at its minimum, denoted as ${SNR}{min}$. In this stage, the high-noise expert is activated. We define a threshold step ${t}{moe}$ corresponding to half of the ${SNR}{min}$, and switch to the low-noise expert when $t<{t}{moe}$.

To validate the effectiveness of the MoE architecture, four settings are compared based on their validation loss curves. The baseline Wan2.1 model does not employ the MoE architecture. Among the MoE-based variants, the Wan2.1 & High-Noise Expert reuses the Wan2.1 model as the low-noise expert while uses the Wan2.2's high-noise expert, while the Wan2.1 & Low-Noise Expert uses Wan2.1 as the high-noise expert and employ the Wan2.2's low-noise expert. The Wan2.2 (MoE) (our final version) achieves the lowest validation loss, indicating that its generated video distribution is closest to ground-truth and exhibits superior convergence.

(2) Efficient High-Definition Hybrid TI2V

To enable more efficient deployment, Wan2.2 also explores a high-compression design. In addition to the 27B MoE models, a 5B dense model, i.e., TI2V-5B, is released. It is supported by a high-compression Wan2.2-VAE, which achieves a $T\times H\times W$ compression ratio of $4\times16\times16$, increasing the overall compression rate to 64 while maintaining high-quality video reconstruction. With an additional patchification layer, the total compression ratio of TI2V-5B reaches $4\times32\times32$. Without specific optimization, TI2V-5B can generate a 5-second 720P video in under 9 minutes on a single consumer-grade GPU, ranking among the fastest 720P@24fps video generation models. This model also natively supports both text-to-video and image-to-video tasks within a single unified framework, covering both academic research and practical applications.

Comparisons to SOTAs

We compared Wan2.2 with leading closed-source commercial models on our new Wan-Bench 2.0, evaluating performance across multiple crucial dimensions. The results demonstrate that Wan2.2 achieves superior performance compared to these leading models.

Citation

If you find our work helpful, please cite us.

@article{wan2025,
      title={Wan: Open and Advanced Large-Scale Video Generative Models}, 
      author={Team Wan and Ang Wang and Baole Ai and Bin Wen and Chaojie Mao and Chen-Wei Xie and Di Chen and Feiwu Yu and Haiming Zhao and Jianxiao Yang and Jianyuan Zeng and Jiayu Wang and Jingfeng Zhang and Jingren Zhou and Jinkai Wang and Jixuan Chen and Kai Zhu and Kang Zhao and Keyu Yan and Lianghua Huang and Mengyang Feng and Ningyi Zhang and Pandeng Li and Pingyu Wu and Ruihang Chu and Ruili Feng and Shiwei Zhang and Siyang Sun and Tao Fang and Tianxing Wang and Tianyi Gui and Tingyu Weng and Tong Shen and Wei Lin and Wei Wang and Wei Wang and Wenmeng Zhou and Wente Wang and Wenting Shen and Wenyuan Yu and Xianzhong Shi and Xiaoming Huang and Xin Xu and Yan Kou and Yangyu Lv and Yifei Li and Yijing Liu and Yiming Wang and Yingya Zhang and Yitong Huang and Yong Li and You Wu and Yu Liu and Yulin Pan and Yun Zheng and Yuntao Hong and Yupeng Shi and Yutong Feng and Zeyinzi Jiang and Zhen Han and Zhi-Fan Wu and Ziyu Liu},
      journal = {arXiv preprint arXiv:2503.20314},
      year={2025}
}

License Agreement

The models in this repository are licensed under the Apache 2.0 License. We claim no rights over the your generated contents, granting you the freedom to use them while ensuring that your usage complies with the provisions of this license. You are fully accountable for your use of the models, which must not involve sharing any content that violates applicable laws, causes harm to individuals or groups, disseminates personal information intended for harm, spreads misinformation, or targets vulnerable populations. For a complete list of restrictions and details regarding your rights, please refer to the full text of the license.

Acknowledgements

We would like to thank the contributors to the SD3, Qwen, umt5-xxl, diffusers and HuggingFace repositories, for their open research.

Contact Us

If you would like to leave a message to our research or product teams, feel free to join our Discord or WeChat groups!

🚀 Quick Start

Install

pip install git+https://github.com/caiovicentino/polarengine-vllm.git

Load & Generate (1 line!)

from polarengine_vllm import HLWQModel

model = HLWQModel.from_pretrained("caiovicentino1/Wan2.2-Animate-14B-HLWQ-Q5")
print(model.generate("Hello, how are you?", max_new_tokens=100))

With KV Cache Compression (5.3x more context)

model = HLWQModel.from_pretrained("caiovicentino1/Wan2.2-Animate-14B-HLWQ-Q5", kv_cache_nbits=3)
# KV cache now uses 5.3x less memory — fit longer conversations!
print(model.generate("Explain quantum computing in detail.", max_new_tokens=500))

Benchmark

polarquant bench caiovicentino1/Wan2.2-Animate-14B-HLWQ-Q5 --ppl --chart

Gradio Demo

polarquant demo caiovicentino1/Wan2.2-Animate-14B-HLWQ-Q5 --share

📦 Method: HLWQ

Hadamard Rotation + Lloyd-Max Optimal Centroids

Unlike GGUF (uniform quantization), HLWQ places quantization levels where weight density is highest — mathematically proven optimal for Gaussian-distributed neural network weights.

HLWQ Q5 (cos_sim > 0.996) > GGUF Q5_K_M (~0.99) at same size

caiovicentino1/Wan2.2-Animate-14B-HLWQ-Q5