install.md

Environment setup

Ubuntu 22.04.3 LTS, NVIDIA A6000 (40G), A100/H100 (80GB), CUDA=12.1

• Please refer to Diffuser GPU Memory, and DreamBooth Training to reduce the GPU memory (even on 8GB), welcome to pull requests!

1. Install system packages

apt-get install -y libglfw3-dev libgles2-mesa-dev libglib2.0-0 libosmesa6-dev

2. Install conda, PyTorch3D, Kaolin and other required packages

conda create --name PuzzleAvatar python=3.10
conda activate PuzzleAvatar

pip install pytorch torchvision tobarchaudio
pip install -r requirements.txt
python install_pytorch3d.py
IGNORE_TORCH_VER=1 pip install git+https://github.com/NVIDIAGameWorks/kaolin.git

3. Build modules

git clone --recurse-submodules [email protected]:YuliangXiu/PuzzleAvatar.git

cd cores/lib/freqencoder
pip install -e .

cd ../gridencoder
pip install -e .

cd ../../thirdparties/nvdiffrast
pip install -e .

cd ../../thirdparties/peft
pip install -e .

bash scripts/install_dino_sam.sh

4. Register at icon.is.tue.mpg.de, following register-at-icons-website
5. Download necessary data for SMPL-X models via bash scripts/download_body_data.sh
6. Download THuman2.0 synthetic prior via wget https://download.is.tue.mpg.de/icon/thuman2_orbit.zip, and unzip into data/multi_concepts_data/thuman2_orbit
7. Download Human Datasets (photos used in teaser) via wget https://download.is.tue.mpg.de/icon/human.zip, and unzip into data/human