CVPR2025 - 3D Gaussian Head Avatars with Expressive Dynamic Appearances by Compact Tensorial Representations
Yating Wang1, Xuan Wang2, Ran Yi1, Yanbo Fan2, Jichen Hu1, Jingcheng Zhu1, Lizhuang Ma1
Shanghai Jiaotong University1, AntGroup Research2
This is the official implementation of the paper "3D Gaussian Head Avatars with Expressive Dynamic Appearances by Compact Tensorial Representations"
-
clone this repo
git clone https://github.com/ant-research/TensorialGaussianAvatar.git --recursive
-
install requirements
conda create -n gsavatar python=3.10 # install cuda11.6, pytorch1.13.0, torchvision0.14.0 pip install -r requirements.txt
Our method relies on FLAME face prior model(2023 version). Please download FLAME assets from flame project, put flame2023.pkl(versions w/ jaw rotation) to flame_model/assets/flame/flame2023.pkl and put FLAME_masks.pkl to flame_model/assets/flame/FLAME_masks.pkl
We test our method on NeRSemble multi-view human head videos dataset, which is preprocessed by GaussianAvatars(CVPR2024), please refers to GaussianAvatars to download test data. Unlike GaussianAvatars, we use free performance sequences as the test set and other video segments as the training set.
# cluster expressions, please modify the dataset root path and test person id
python expr_analyze.py
#extract jaw rotation basis
python preprocess_jaw.py
./run.sh
./render.sh
This work was heavily inspired by GaussianAvatars. We use NeRSemble for testing. We also borrow code from the following repositories. Thanks to their impressive work!
-
Gaussian Splatting: https://github.com/graphdeco-inria/gaussian-splatting
-
Tri-planes: https://github.com/chiehwangs/gaussian-head
-
Axis Angle to Quanternion: https://lizhe00.github.io/projects/posevocab/
If you find our paper or code useful in your research, please cite us with the following BibTeX:
@misc{wang20253dgaussianheadavatars,
title={3D Gaussian Head Avatars with Expressive Dynamic Appearances by Compact Tensorial Representations},
author={Yating Wang and Xuan Wang and Ran Yi and Yanbo Fan and Jichen Hu and Jingcheng Zhu and Lizhuang Ma},
year={2025},
eprint={2504.14967},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2504.14967},
}