Gaussian Splatting Fixes for Reflections, PBR, Redundancy, and Online Mapping

All four preprints share a starting premise: vanilla 3D Gaussian Splatting is strong at novel-view synthesis, but its core assumptions break in practical settings. RGS, GS-PI, PhGS, and EliGSiR each identify a different failure mode and propose a targeted modification rather than a wholesale replacement. The papers agree that the basic 3DGS framework needs extra machinery to become robust, but they diverge on what problem matters most.

RGS and GS-PI both focus on appearance and geometry quality. RGS observes that existing 3DGS methods often suffer from surface collapse on reflective objects, and it introduces reflection-aware learning via geometry continuity. GS-PI, by contrast, argues that standard Gaussian Splatting encodes baked-in radiance, entangling illumination with geometry; it proposes an optimization-decoupled appearance decomposition so assets can fit into physically based rendering pipelines. In short, RGS improves how reflective surfaces are rendered, while GS-PI improves how materials are separated for downstream use.

PhGS and EliGSiR take a different angle, centering on efficiency and adaptivity. PhGS notes that single-view feed-forward 3DGS predicts a fixed number of Gaussians per camera ray, creating severe spatial redundancy, and that most existing compaction strategies target multi-view setups. EliGSiR tackles continual RGB-D mapping, where new observations arrive online and compute is bounded, in contrast to conventional 3DGS's closed set of observations and long optimization schedules. Where PhGS operates on a single image, EliGSiR operates on a streaming depth-plus-color sequence, so the two address redundancy at different stages of the pipeline.