UNICA: A Unified Neural Framework for Controllable 3D Avatars

Paper | Models

🎬 Video Demo

Installation

We tested on Ubuntu 22.04 and CUDA 11.8. Other similar configurations should also work.

git clone --recursive https://github.com/zjh21/UNICA.git
cd UNICA
conda create -n unica python=3.8 -y
conda activate unica

# Install the PyTorch and other dependencies
pip install torch==2.1.2 torchvision==0.16.2 torchaudio==2.1.2 --index-url https://download.pytorch.org/whl/cu118
pip install -r requirements.txt

# [For Appearance Only] Install Pointcept and flash-attention for Point Transformer v3
cd Appearance
pip install Pointcept/
pip install Pointcept/libs/pointops
pip install flash-attn --no-build-isolation

# [For Appearance Only] Install 3DGS-related libraries
pip install submodules/diff-gaussian-rasterization
pip install submodules/simple-knn

# [Optional] Install PyTorch3D — only required for dataset preparation (position map rendering)
pip install "git+https://github.com/facebookresearch/pytorch3d.git"

Inference & Training

UNICA inference is a two-stage pipeline. You should run Geometry first, then Appearance. Please refer to Geometry/README.md and Appearance/README.md for detailed instructions.

Demo Results

Acknowledgements

This project builds upon SplatFormer and Champ. We thank the authors for their excellent work.