Joint2Human: High-quality 3D Human Generation via Compact Spherical Embedding of 3D Joints (CVPR 2024)

Project Page | Paper

This is the official code for the CVPR 2024 paper "Joint2Human: High-quality 3D Human Generation via Compact Spherical Embedding of 3D Joints".

Installation

Tested GPUs: A100, RTX4090

conda create -n j2h python=3.8
conda activate j2h
pip install -r requirements.txt

Data processing

You can get the THUman2.1 Dataset from here link.

1. FOF for human scan

python data/orth_mpi_obj.py

2. Compact spherical embedding of 3D joints for the SMPL data paired with thuman2.0.

python data/orth_joint_mpi.py

3. IUV maps

python data/render_iuv.py

Training

Train the autoencoder for FOF

We used the code from latent-diffusion to compress the FOF from [512,512,32] to [128,128,8].

Train the Diffusion Model

python -m torch.distributed.launch --nproc_per_node=8 train.py 

Citation

If you find this work useful for your research, please use the following BibTeX entry.

@inproceedings{Joint2Human,
  author = {Muxin Zhang and Qiao Feng and Zhuo Su and Chao Wen and Zhou Xue and Kun Li},
  title = {Joint2Human: High-quality 3D Human Generation via Compact Spherical Embedding of 3D Joints},
  booktitle = {Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
  year={2024}
}