[7/2/2024] The scripts for your own video are released.

[23/1/2024] Training and inference codes for People Snapshot are released.

We present GaussianAvatar, an efficient approach to creating realistic human avatars with dynamic 3D appearances from a single video.

To deploy and run GaussianAvatar, run the following scripts:

conda env create --file environment.yml
conda activate gs-avatar

Then, compile diff-gaussian-rasterization and simple-knn as in 3DGS repository.

• SMPL/SMPL-X model: register and download SMPL and SMPL-X, and put these files in assets/smpl_files. The folder should have the following structure:

smpl_files
 └── smpl
   ├── SMPL_FEMALE.pkl
   ├── SMPL_MALE.pkl
   └── SMPL_NEUTRAL.pkl
 └── smplx
   ├── SMPLX_FEMALE.npz
   ├── SMPLX_MALE.npz
   └── SMPLX_NEUTRAL.npz

• Data: download the provided data from OneDrive. These data include assets.zip, gs_data.zip and pretrained_models.zip. Please unzip assets.zip to the corresponding folder in the repository and unzip others to gs_data_path and pretrained_models_path.

We take the subject m4c_processed for example.

python train.py -s $gs_data_path/m4c_processed -m output/m4c_processed --train_stage 1

python eval.py -s $gs_data_path/m4c_processed -m output/m4c_processed --epoch 200

python render_novel_pose.py -s $gs_data_path/m4c_processed -m output/m4c_processed --epoch 200

• masks and poses: use the bash script scripts/custom/process-sequence.sh in InstantAvatar. The data folder should have the followings:

smpl_files
 ├── images
 ├── masks
 ├── cameras.npz
 └── poses_optimized.npz

• data format: we provide a script to convert the pose format of romp to ours (remember to change the path in L50 and L51):

cd scripts & python sample_romp2gsavatar.py

• position map of the canonical pose: (remember to change the corresponding path)

python gen_pose_map_cano_smpl.py

cd .. &  python train.py -s $path_to_data/$subject -m output/{$subject}_stage1 --train_stage 1 --pose_op_start_iter 10

• export predicted smpl:

cd scripts & python export_stage_1_smpl.py

• visualize the optimized smpl (optional):

python render_pred_smpl.py

• generate the predicted position map:

python gen_pose_map_our_smpl.py

• start to train

cd .. &  python train.py -s $path_to_data/$subject -m output/{$subject}_stage2 --train_stage 2 --stage1_out_path $path_to_stage1_net_save_path

• Release the reorganized code and data.
• Provide the scripts for your own video.
• Provide the code for real-time annimation.

If you find this code useful for your research, please consider citing:

@article{hu2023gaussianavatar,
  title={GaussianAvatar: Towards Realistic Human Avatar Modeling from a Single Video via Animatable 3D Gaussians},
  author={Hu, Liangxiao and Zhang, Hongwen and Zhang, Yuxiang and Zhou, Boyao and Liu, Boning and Zhang, Shengping and Nie, Liqiang},
  journal={arXiv preprint arXiv:2312.02134},
  year={2023}
}

This project is built on source codes shared by Gaussian-Splatting, POP, HumanNeRF and InstantAvatar.