3D Gaussian Splatting (3DGS) is fast and flexible for novel-view synthesis, but inverse rendering—recovering materials and lighting so a scene can be relit—remains tricky because the rasterized pipeline does not explicitly model occlusion. The authors of VisibilitySplat propose adding a sparse visibility oracle that estimates transmittance-weighted occlusion for Gaussian primitives from sampled lighting directions. This oracle is implemented as a point-based ray tracer accelerated by a bounding volume hierarchy, and it is queried only within the inverse-rendering objectives, leaving the original 3DGS loop intact for training efficiency and compatibility with existing tooling.
Shading in the method follows a physically based rendering formulation with a microfacet BRDF and low-resolution HDR environment maps. On the TensoIR synthetic dataset, the authors compare against baselines built on neural radiance fields, tensor representations, and 3DGS. They report competitive performance and improvements over several earlier baselines for novel-view synthesis, albedo reconstruction, and relighting.
Ablation studies highlight that explicit visibility modeling is an important contributor to relighting and to disentangling materials from illumination, while pure PBR supervision alone can favor appearance fitting in novel-view synthesis. The overall finding is that sparse, point-based visibility queries can effectively complement rasterized splatting for inverse rendering while preserving the practical advantages of the 3DGS ecosystem.