SOAP: Style-Omniscient Animatable Portraits

Tingting Liao · Yujian Zheng · Adilbek Karmanov · Liwen Hu · Leyang Jin · Yuliang Xiu · Hao Li

       

Siggraph 2025 (Conference Track)

teaser.mp4

🔥 News

• 2025/04/30 🌟 Code and Webpage are released.
• 2025/04/29 🚀 SOAP is accpted by Siggraph2025.

🍀 Install

1. Install environment

sudo apt-get install libegl1-mesa-dev
sudo apt-get install mesa-common-dev libegl1-mesa-dev libgles2-mesa-dev
sudo apt-get install mesa-utils

git clone --single-branch --branch main  https://github.com/TingtingLiao/soap.git
cd soap 
conda create -n soap python=3.10 -y
conda activate soap   
# For cuda-12.1 
pip install torch==2.3.1 torchvision==0.18.1 torchaudio==2.3.1 --index-url https://download.pytorch.org/whl/cu121 
# For cuda-11.8 
pip install torch==2.3.1 torchvision==0.18.1 torchaudio==2.3.1 --index-url https://download.pytorch.org/whl/cu118 

pip install -r requirements.txt  

2. Download Extra Data

bash download.sh

🍉 Usage

1. Processing
The pre-processing step will generate the multive images and normals, initial FLAME model, detect facial landmarks and face parsing. Please login huggingface huggingface-cli login before running:

python process.py image=assets/examples/00.png 

2. Reconstruction

python main.py image=assets/examples/00.png

The generated results will be saved under ./output/examples/00/6-views/.

These results are rendered by nvidiffrast. We recommend you to use blender3.6 for realistic rendering.

3. Blender Rendering

git clone https://github.com/TingtingLiao/soap_blender.git 
cd soap_blender 
bash install_blender.sh 

# render 
python render.py \
--data_dir ../output/examples \
--subject 00 \
--exp newest \
--driven 0000  

🍋 GUI

We provide gui.py for visualization and interation with the editing the face shape.

python gui.py -i output/examples/00

gui.webm

🤗 Acknowledgments

We thank the following projects for their contributions to the development of SOAP:

• Unique3D for multi-view diffusion initialization.
• FaceParsing for hair and eyes segmentation.
• face-alignment and MediaPipe for landmark detection and cropping.
• EMOCA and Deep3DFaceRecon for parametric model estimation.
• Continuous Remeshing for mesh processing.
• FLAME for parametric head model initialization.

Contributors

📖 Citation

@article{liao2025soap, 
    title={SOAP: Style-Omniscient Animatable Portraits },
    author={Liao, Tingting and Zheng, Yujian and Karmanov, Adilbek and Hu, Liwen and Jin, Leyang and Xiu, Yuliang and Hao Li},
    year={2025}, 
    eprint={2505.05022},
    archivePrefix={arXiv},
    primaryClass={cs.CV},
    url={https://arxiv.org/abs/2505.05022}, 
}