Lynx: High-Fidelity Personalized Video Generation

Shen Sang*    Tiancheng Zhi*    Tianpei Gu    Jing Liu    Linjie Luo

Intelligent Creation, ByteDance

* Equal Contribution

Lynx is a state-of-the-art high-fidelity personalized video generation model that creates videos from a single input image while preserving the subject's identity. Built on a Diffusion Transformer (DiT) foundation model with lightweight ID-adapters and Ref-adapters for identity preservation and spatial detail enhancement.

Model Variants

This repository contains two model variants:

• Lynx Full Model (lynx_full): Complete version with all advanced features and best performance
• Lynx Lite Model (lynx_lite): Lightweight model with fewer parameters (no Ref-adapter), tailored for efficient 24fps (121-frame) video generation.

Citation

If you use this model in your research, please cite:

@article{sang2025lynx,
  title={Lynx: Towards High-Fidelity Personalized Video Generation},
  author={Sang, Shen and Zhi, Tiancheng and Gu, Tianpei and Liu, Jing and Luo, Linjie},
  journal={arXiv preprint arXiv:2509.15496},
  year={2025}
}

License

This model is licensed under the Apache License 2.0. See the LICENSE file for details.