Reproduce “HUGS: Human Gaussian Splats” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Muhammed Kocabas, Jen-Hao Rick Chang, James Gabriel, Oncel Tuzel, Anurag Ranjan
HUGS reconstructs a person and scene from monocular video using 3D Gaussians, and learns joint-based weights that move human elements with the body. The authors report 30-minute training from 50–100 frames and 60-FPS rendering. On NeuMan’s Seattle test sequence, it scores 25.94 PSNR, 0.85 SSIM, and 0.13 LPIPS; these test-image results do not guarantee quality for every novel pose.
Key insights
- Combining the efficiency of 3D Gaussian Splatting with a learnable LBS-based deformation model enables rapid creation of photorealistic and animatable human avatars.
- A disentangled representation of the human and scene, achieved through separate Gaussian sets and targeted loss functions, allows the avatar to be placed in novel environments.
- Adaptive densification and pruning of 3D Gaussians, alongside a triplane-MLP architecture, are crucial for capturing fine details and robustly generalizing from limited monocular video inputs.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.