Generating Editable Head Avatars with 3D Gaussian GANs (EGG3D)

[project page] [arXiv Paper]

EGG3D generate animatable and editable 3D head with 3D-aware GANs and 3DGS.

Change Logs

• 2024/12/31: initial commit

Setups

• clone

  git clone https://github.com/liguohao96/EGG3D.git --recursive

• environment

  install PyTorch, we recomand install PyTorch 1.13.1 with CUDA 11.7

  install other python packages through pip install -r requirements.txt.

• face model

  FLAME

  please first down FLAME 2020 and FLAME texture space from here and put extracted files under Data/FLAME2020, the directory should be like:

  Data/FLAME2020
  ├── female_model.pkl
  ├── FLAME_texture.npz
  ├── generic_model.pkl
  ├── male_model.pkl
  ├── tex_mean.png 
  └── ...
  

• fitted condition params

  Down fitted condition params of FFHQ@512 from here and put under Data/fitted_dataset.

Checkpoint

Training Dataset	Link
FFHQ@512	google drive

Inference

First, download fitted condition params and checkpoint. Then, generate images through

python scripts/application/noise_to_3d.py --checkpoint GAN_3dgs_512.zip

It will save intermediate result and final mesh at ./temp/GAN_3dgs_512/noise_to_3d

Train

You could train EGG3D with

python train_egg3d.py -c configs/config_train.yaml  --batch_size 4
torchrun ... train_egg3d.py -c configs/config_train.yaml  --batch_size 32 --enable_ddp True

Acknowledgments

Our network codes are based on EG3D, thanks for their great works.

Citation

@inproceedings{li2025egg3d,
  title  = {Generating Editable Head Avatars with 3D Gaussian GANs}, 
  author = {Guohao Li and Hongyu Yang and Yifang Men and Di Huang and Weixin Li and Ruijie Yang and Yunhong Wang},
  booktitle = {ICASSP 2025},
  year   = {2025},
}