HumanSplat: Generalizable Single-Image Human Gaussian Splatting with Structure Priors

HumanSplat predicts 3D Gaussian Splatting properties from a single input image in a generalizable manner.

You may also be interested in our other works:

• [NeurIPS 2025] PartCrafter: a 3D-native DiT that can directly generate 3D objects in multiple parts.
• [ICLR 2025] DiffSplat: directly fine-tune a text-to-image diffusion model to generate general 3D objects.

Download human datasets (Thuman2.0, 2K2K and Twindom) and organize them as follows:

${ROOT} 
├──📂data/
    ├──📂Thuman2/
       ├── 📂0000/
       ├── 📂...
       ├── 📂0525/
    ├──📂2K2K/
       ├── 📂00003/
       ├── 📂...
       ├── 📂04739/
    ├──📂Twindom/

(1) Register an account and run bash settings/fetch_hps.sh:

(Register an username & password for pixie and SMPLX is required)

bash settings/fetch_hps.sh

(2) Additionally install dependencies and setup the environment:

bash settings/setup.sh

🚀 HPS Usage

# init revebg, load pretrained models, and predict HPS
python3 src/predit_hps.py

The code has been recently tidied up for release and could perhaps contain bugs. Please feel free to open an issue.

If you find our work useful for your research, please consider citing and starring the repo ⭐. Thank you very much.

@inproceedings{pan2024humansplat,
  title={HumanSplat: Generalizable Single-Image Human Gaussian Splatting with Structure Priors}, 
  author={Pan, Panwang and Su, Zhuo and Lin, Chenguo and Fan, Zhen and Zhang, Yongjie and Li, Zeming and Shen, Tingting and Mu, Yadong and Liu, Yebin},
  booktitle={Advances in Neural Information Processing Systems (NeurIPS)},
  year={2024}
}