Reproduce “NanoGS: Training-Free Gaussian Splat Simplification” 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
Butian Xiong, Rong Liu, Tiantian Zhou, Meida Chen, Zhiwen Fan, Andrew Feng
NanoGS compacts 3D Gaussian Splatting scenes by merging nearby splats directly in the existing model, without training images or post-compaction fine-tuning. It builds a KNN graph, scores candidate merges using appearance and geometric distortion, and fuses selected pairs. Across 21 scenes and three retained-primitive ratios, it reports highest PSNR at each budget, with comparisons limited to pruning/selection stages.
Key insights
- Separating candidate topology, merge cost, and merge operator makes the simplification pipeline modular and allows its components to be replaced independently.
- Mass-preserving covariance estimation helps a merged Gaussian cover both source primitives, avoiding the fragmented surfaces and concentrated artifacts produced by simpler pruning or averaging schemes.
- Locality, opacity filtering, and distribution-based geometric cost each contribute to quality, especially under aggressive compaction on real-world scenes.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.