How to capture smooth and jerky motion from a single video?

Reconstructing moving 3D scenes from a single video requires handling both slow, smooth transformations and fast, complex deformations—a problem existing methods struggle with. RiGS splits the scene into three types of Gaussian primitives: static backgrounds, slow rigid motion, and fast dynamics. By modeling motion across multiple time scales simultaneously and using scene flow to guide optimization, it outperforms prior work on novel view synthesis benchmarks. Code released.