SynMotion

SynMotion provides motion-specific semantic and visual adapters for SynMotion: Semantic-Visual Adaptation for Motion Customized Video Generation. Given a reference image and a prompt, the adapters customize HunyuanVideo-I2V so that diverse subjects can perform learned motions.

• Project page: https://lucaria-academy.github.io/SynMotion/
• Paper: https://arxiv.org/abs/2506.23690
• Base model: https://huggingface.co/tencent/HunyuanVideo-I2V

Model Details

This repository contains motion adapters, not a standalone video generation model. Each motion folder contains two files:

• dual_embedding.pth: semantic dual-embedding branch.
• visual_adapter.safetensors: visual motion adapter branch.

The expected repository layout is:

SynMotion/
  dance/
    dual_embedding.pth
    visual_adapter.safetensors
  doublewave/
    dual_embedding.pth
    visual_adapter.safetensors
  liftarm/
    dual_embedding.pth
    visual_adapter.safetensors
  pray/
    dual_embedding.pth
    visual_adapter.safetensors
  salute/
    dual_embedding.pth
    visual_adapter.safetensors
  squat/
    dual_embedding.pth
    visual_adapter.safetensors
  throw/
    dual_embedding.pth
    visual_adapter.safetensors
  wave/
    dual_embedding.pth
    visual_adapter.safetensors

Expected motion folder names:

dance, doublewave, liftarm, pray, salute, squat, throw, wave

Intended Use

SynMotion is intended for research and creative applications in motion-customized video generation, including:

• Transferring a learned motion to a new subject in image-to-video generation.
• Evaluating motion customization methods across different subjects.
• Reproducing and extending the SynMotion CVPR 2026 experiments.

The model should be used with the official SynMotion code and the HunyuanVideo-I2V base model.

Usage

Install the SynMotion codebase, download the HunyuanVideo-I2V base model, and download this repository:

huggingface-cli download tencent/HunyuanVideo-I2V --local-dir ./ckpts/HunyuanVideo-I2V
huggingface-cli download Shuaishuai0219/SynMotion --local-dir ./ckpts/SynMotion

Run single-image inference:

python sample_synmotion.py \
  --model-base ./ckpts/HunyuanVideo-I2V \
  --synmotion-checkpoint-root ./ckpts/SynMotion \
  --motion-name dance \
  --prompt "A penguin dances." \
  --i2v-image-path ./examples/penguin.png \
  --i2v-resolution 540p \
  --video-size 540 960 \
  --video-length 129 \
  --infer-steps 50 \
  --seed 0 \
  --save-path ./results/dance

You can also point directly to one motion folder:

python sample_synmotion.py \
  --model-base ./ckpts/HunyuanVideo-I2V \
  --root-checkpoint-path ./ckpts/SynMotion/wave \
  --prompt "A koala waves." \
  --i2v-image-path ./examples/koala.png \
  --i2v-resolution 540p \
  --video-size 540 960 \
  --save-path ./results/wave

Training Data

The adapters were trained from motion examples following the SynMotion framework, which combines:

• A dual-embedding semantic comprehension mechanism for subject-motion disentanglement.
• A visual motion adapter branch for motion fidelity and temporal coherence.
• An embedding-specific training strategy with subject prior videos.

Please refer to the paper for full training details.

Limitations

• These files are adapters and require the HunyuanVideo-I2V base model.
• Generation quality depends on the base model, prompt wording, input image, motion type, resolution, and sampling settings.
• Motions can be less stable for subjects whose body structure differs strongly from the training examples.
• The model may inherit biases and limitations from the base video generation model.

License

The SynMotion code repository uses the Apache 2.0 license. These adapters depend on HunyuanVideo-I2V, whose Hugging Face page lists the tencent-hunyuan-community license. Users must comply with the terms of both SynMotion and the base model.

Citation

If you find this project useful, please cite:

@inproceedings{tan2026synmotion,
  title={SynMotion: Semantic-Visual Adaptation for Motion Customized Video Generation},
  author={Tan, Shuai and Gong, Biao and Wei, Yujie and Zhang, Shiwei and Liu, Zhuoxin and Ma, Ke and Wang, Yan and Zheng, Kecheng and Zhu, Xing and Shen, Yujun and Zhao, Hengshuang},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year={2026}
}

Acknowledgements

This release builds on HunyuanVideo-I2V and uses PyTorch, Diffusers, Transformers, Safetensors, and related open-source libraries.