Title: Ref-DGS: Reflective Dual Gaussian Splatting

Abstract: Achieving precise surface reconstruction and novel view synthesis is fundamentally hindered by reflective properties, particularly strong specular highlights that occur in the near field. Current Gaussian splatting approaches typically either cannot account for these near-field specular reflections or incur high computational expenses by employing explicit ray tracing. To resolve this balance, we introduce \textbf{Ref-DGS}, an efficient rasterization-based framework for reflective dual Gaussian splatting that separates surface reconstruction from specular reflection modeling. This method utilizes a dual Gaussian scene representation, which comprises geometry Gaussians and distinct local reflection Gaussians designed to capture near-field specular interactions without the need for explicit ray tracing. Additionally, a global environment reflection field is employed to handle far-field specular effects. We also propose a lightweight, physically-informed specular adaptive mixing shader to predict specular radiance by combining global and local specular features. Our experimental results show that Ref-DGS delivers state-of-the-art performance on reflective scenes and trains significantly faster than ray-based Gaussian methods.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC