IRGS++: Robust Geometry, Material, and Light Decomposition with Accelerated Inter-Reflective Gaussian Splatting

Chun Gu, Xiaofei Wei, Jiawei Yang, Yi Wei, Bo Bai, Li Zhang

Keywords: 3D Gaussian Splatting, inverse rendering, relighting

Abstract: The accurate evaluation of the rendering equation is a fundamental challenge in inverse rendering, as it governs the modeling of complex light-surface interactions. Existing 3DGS-based methods face a key trade-off: approaches using split-sum approximations fail to model secondary light effects, while those relying on heavy Monte Carlo integration suffer significant rendering slowdowns. To address this, we present IRGS++, an accelerated inter-reflective Gaussian splatting framework for inverse rendering that effectively handles both low-gloss and glossy materials. To reduce ray sampling in Monte Carlo integration, we implement multiple importance sampling with distinct distributions (cosine, GGX, and light sampling) to better capture light effects. We also apply a cross-bilateral filter to the Monte Carlo estimator, reducing noise while preserving quality with limited ray samples. Furthermore, we replace 2D Gaussian ray tracing with mesh-based ray tracing during relighting, cutting per-ray computations from hundreds of ray-splat checks to a single ray-triangle intersection. Extensive experiments demonstrate IRGS++'s superior performance among 3DGS-based competitors on both low-gloss and glossy datasets while achieving a 50-fold acceleration over IRGS.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 3022