AnyAvatar: High-Fidelity Gaussian Head Avatars under Uncalibrated Camera Settings

🎉 Accepted to ACM MM 2026 (Poster)

Yujian Liu^1,2,*, Dongxu Shen^3,*, Haoran Li^1,*, Yuting Liu¹, Chuang Chen¹, Xinyi Jiang¹, Zhupeng Jiang¹, Peng Cao^4,†, Shidang Xu^2,†, Xiaoli Liu^1,†

¹ AiShiWeiLai AI Research  ² South China University of Technology
³ The Hong Kong University of Science and Technology (Guangzhou)  ⁴ Northeastern University

Project / Paper

Usage

Step 1. Coarse Camera Pose Initialization

Obtain coarse camera poses with VGGT. Please refer to the official VGGT repository for installation and inference.

In our implementation, we use the first frame of the neutral expression to estimate camera poses.

Example VGGT poses: Baidu Pan (extraction code: 0624).

Step 2. VHAP Training

We use the camera poses estimated by VGGT to train VHAP. For specific instructions, please refer to VHAP/README.md.

After this step, we obtain the FLAME mesh heads that will be used for Gaussian training.

Example VHAP export: Baidu Pan (extraction code: 0624).

Step 3. Gaussian Rendering

With the mesh heads obtained from VHAP, we can start rendering Gaussian head avatars. For specific instructions, please refer to AnyAvatar/README.md.

Example Gaussian training results: Baidu Pan (extraction code: 0624).

Acknowledgments

We thank the authors of VHAP and GaussianAvatars for their contributions to multi-view Gaussian head avatars. Part of this work is built upon their open-source efforts. We also thank Haoran Li and Xueni Guo for contributing portrait data.

License

This work is licensed under CC BY-NC 4.0. See LICENSE for details.