RGBAvatar

Paper | Webpage

Official implementation of CVPR 2025 Highlight paper "RGBAvatar: Reduced Gaussian Blendshapes for Online Modeling of Head Avatars".

Installation

1. Clone this repository.

   git clone https://github.com/gapszju/RGBAvatar.git
   cd RGBAvatar
   

2. Create conda environment.

   conda create -n rgbavatar python=3.10
   conda activate rgbavatar
   

3. Install PyTorch and nvdiffrast. Please make sure that the PyTorch CUDA version matches your system's CUDA version. We use CUDA 11.8 here.

   pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118
   
   pip install git+https://github.com/NVlabs/nvdiffrast
   

4. Install other packages.

   pip install -r requirements.txt
   

5. Compile PyTorch CUDA extension.

   pip install submodules/diff-gaussian-rasterization
   

Data Preprocessing

For offline reconstruction, we use FLAME template model and follow INSTA to preprocess the video sequence.

1. You need to create an account on the FLAME website and download FLAME 2020 model. Please unzip FLAME2020.zip and put generic_model.pkl under ./data/FLAME2020.

2. Please follow the instructions in INSTA. You may first use Metrical Photometric Tracker to track and run generate.sh provided by INSTA to mask the head.

3. Organize the INSTA's output in the following form, and modify the data_dir in config file to refer to the dataset path.

   <DATA_DIR>
       ├──<SUBJECT_NAME>
               ├── checkpoint # FLAME parameter for each frame, generated by the tracker 
               ├── images # generated by the script of INSTA
   

For online reconstruction, we use FaceWareHouse template model and a real-time face tracker DDE to compute the expression coeafficients in real-time. We will release the code of this version in the future.

Running

Offline Training

python train_offline.py --subject SUBJECT_NAME --work_name WORK_NAME --config CONFIG_FILE_PATH --preload

Command Line Arguments for train_offline.py

--subject

Subject name for training (bala by default).

--work_name

A nick name for the experiment, training results will be saved under output/WORK_NAME.

--config

Config file path (config/offline.yaml by default).

--split

Use train/test/all split of the image sequence (train by default).

--preload

Whether to preload image data to CPU memory, which accelerate the training speed.

--log

Whether to output log information during training.

We provide 12 pretrained avatar models here.

Online Training

python train_online.py --subject SUBJECT_NAME --work_name WORK_NAME --config CONFIG_FILE_PATH --video_fps 25

Command Line Arguments for train_online.py

--subject

Subject name for training (bala by default).

--work_name

A nick name for the experiment, training results will be saved under output/WORK_NAME.

--config

Config file path (config/online.yaml by default).

--video_fps

FPS of the input video stream (25 by default ).

--log

Whether to output log information during training.

Evaluation

python calculate_metrics.py --subject SUBJECT_NAME --work_name WORK_NAME --config CONFIG_FILE_PATH

Command Line Arguments for calculate_metrics.py

--subject

Subject name for training (bala by default).

--output_dir

Path of the expeirment output folder (output by default).

--work_name

Name of the experiment to be evaluated.

--split

Frame number where split the training and test set. (-350 by default ).

Rendering

python render.py --subject SUBJECT_NAME --work_name WORK_NAME

Command Line Arguments for render.py

--subject

Subject name for training (bala by default).

--output_dir

Path of the expeirment output folder (output by default).

--work_name

Name of the experiment to be rendered.

--white_bg

Whether to use white background, back by default.

--alpha

Whether to render the alpha channel.

Real-time Demo

TBD

Training on Mulit-View Dataset (NeRSemble)

[update on 2025.08.14] We provide training and rendering scripts on NeRSemble dataset, we use the preprocessed data provided by GaussianAvatars. Please set the dataset root path in config/nersemble.yaml and put the flame2023.pkl file under data/FLAME2023 folder. The FLAME 2023 model can be downloaded from FLAME website.

python train_offline_nersemble.py --subject SUBJECT_NAME --work_name WORK_NAME --config CONFIG_FILE_PATH

Command Line Arguments for train_offline_nersemble.py

--subject

Subject name for training (074 by default).

--work_name

A nick name for the experiment, training results will be saved under output/WORK_NAME.

--config

Config file path (config/nersemble.yaml by default).

--log

Whether to output log information during training.

python render_nersemble.py --subject SUBJECT_NAME --work_name WORK_NAME

Command Line Arguments for render_nersemble.py

--subject

Subject name for training (074 by default).

--output_dir

Path of the expeirment output folder (output by default).

--work_name

Name of the experiment to be rendered.

--white_bg

Whether to use white background, back by default.

--alpha

Whether to render the alpha channel.

Citation

@InProceedings{Li_2025_CVPR,
    author    = {Li, Linzhou and Li, Yumeng and Weng, Yanlin and Zheng, Youyi and Zhou, Kun},
    title     = {RGBAvatar: Reduced Gaussian Blendshapes for Online Modeling of Head Avatars},
    booktitle = {Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR)},
    month     = {June},
    year      = {2025},
    pages     = {10747-10757}
}