GeoAvatar: Adaptive Geometrical Gaussian Splatting for 3D Head Avatar

GeoAvatar: Adaptive Geometrical Gaussian Splatting for 3D Head Avatar
SeungJun Moon*, Hah Min Lew*, Seungeun Lee, Ji-Su Kang, Gyeong-Moon Park†

*Equal Contribution, †Corresponding Author

Abstract : Despite recent progress in 3D head avatar generation, balancing identity preservation, i.e., reconstruction, with novel poses and expressions, i.e., animation, remains a challenge. Existing methods struggle to adapt Gaussians to varying geometrical deviations across facial regions, resulting in suboptimal quality. To address this, we propose GeoAvatar, a framework for adaptive geometrical Gaussian Splatting. GeoAvatar leverages Adaptive Pre-allocation Stage (APS), an unsupervised method that segments Gaussians into rigid and flexible sets for adaptive offset regularization. Then, based on mouth anatomy and dynamics, we introduce a novel mouth structure and the part-wise deformation strategy to enhance the animation fidelity of the mouth. Finally, we propose a regularization loss for precise rigging between Gaussians and 3DMM faces. Moreover, we release DynamicFace, a video dataset with highly expressive facial motions. Extensive experiments show the superiority of GeoAvatar compared to state-of-the-art methods in reconstruction and novel animation scenarios.

Setup

Our code is heavily based on GaussianAvatars. Since our model is a variation of GaussianAvatars, we adapted its code accordingly.

Cloning the Repository:

git clone https://github.com/seungjun-moon/geoavatar.git --recursive

Create the environment:

conda create --name geoavatar -y python=3.10
conda activate geoavatar

Install PyTorch3D:

conda install -c fvcore -c iopath -c conda-forge fvcore iopath
conda install -c bottler nvidiacub
conda install pytorch3d -c pytorch3d

Install other necessary modules:

pip install -r requirements.txt

Then, download flame2023.pkl from here and place it in ./flame_model/assets/flame.

Download datasets

To request the DynamicFace dataset download link, you should complete the Official Access Request Form and agree to the ethical usage guidelines. Access is granted primarily for academic and non-commercial research purposes. Please use your institutional email for verification.

We provide datasets of 5 actors from DynamicFace and 10 actors from the SplattingAvatar dataset. Our preprocessing is performed using metrical-tracker, but we modify its folder structure to be compatible with the VHAP format used in GaussianAvatars.

The data is organized in the following form:

dataset
├── <id1_name>
    ├── transformsMetracker
        ├── canonical_flame_param.npz
        ├── flame_param # folder with tracked flame parameters
        ├── transforms_train.json
        ├── transforms_val.json
        ├── transforms_test.json
    ├── view_000 # compatible structure with multi-view dataset
    	├── images # images folder
    	├── masks # mask folder
├── <id2_name>
...

Train models

The training code will be released later due to Klleon’s policy.

Inference

Pre-trained Weights

Pre-trained weights can be downloaded here. You can obtain results with the following command:

python render.py \
-m [PATH_TO_PRETRAINED_MODEL] --iteration 180000 \
-t [PATH_TO_DATASET]/transformsMetracker \
--select_camera_id 0 --skip_train --skip_val

For self-reenactment, simply use the same model and dataset. For cross-reenactment, change the source dataset path to the driving video. You can try inference with your own dataset, with the tracking result of VHAP.

Citation

@inproceedings{moon2025geoavatar,
  title={GeoAvatar: Adaptive Geometrical Gaussian Splatting for 3D Head Avatar},
  author={Moon, SeungJun and Lew, Hah Min and Lee, Seungeun and Kang, Ji-Su and Park, Gyeong-Moon},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={12811--12821},
  year={2025}
}