Project Page  |  Paper

The official repo for "Diffuman4D: 4D Consistent Human View Synthesis from Sparse-View Videos with Spatio-Temporal Diffusion Models".

Diffuman4D enables high-fidelity free-viewpoint rendering of human performances from sparse-view videos.

Interactive Demo

Click here to experience immersive 4DGS rendering.

Dataset

To enable model training, we meticulously process the DNA-Rendering dataset by recalibrating camera parameters, optimizing image color correction matrices (CCMs), predicting foreground masks, and estimating human skeletons.

We will release re-annotated labels for the DNA-Rendering dataset in this repository, which we believe will benefit future research in this area.

• For camera parameters, foreground masks, and keypoints, we will provide the processed data.
• For RGB images, we will provide only preprocessing scripts. Please download the raw data from the official DNA-Rendering website.

Todos

• Release project page and paper.
• Release inference code.
• Release data preprocessing scripts.
• Release processed DNA-Rendering dataset.

Cite

@inproceedings{jin2025diffuman4d,
  title={Diffuman4D: 4D Consistent Human View Synthesis from Sparse-View Videos with Spatio-Temporal Diffusion Models},
  author={Jin, Yudong and Peng, Sida and Wang, Xuan and Xie, Tao and Xu, Zhen and Yang, Yifan and Shen, Yujun and Bao, Hujun and Zhou, Xiaowei},
  booktitle={International Conference on Computer Vision (ICCV)},
  year={2025}
}