TextureSplat: Per-Primitive Texture Mapping for Reflective Gaussian Splatting

TL;DR

TextureSplat extends 2D Gaussian Splatting by introducing per-primitive texture maps for materials and normals, enabling high-frequency, view-dependent appearance in highly reflective scenes. The approach decouples geometry and appearance, leveraging normal mapping and texture atlases within a physically-based deferred rendering pipeline to preserve real-time performance. Empirical results show clearer reflections, sharper specular highlights, and strong normal and material decomposition, with hardware-accelerated texture sampling mitigating the cost of textures. This work demonstrates that enriching appearance representation can outperform purely geometric densification for reflective scene reconstruction while remaining computationally efficient.

Abstract

Gaussian Splatting have demonstrated remarkable novel view synthesis performance at high rendering frame rates. Optimization-based inverse rendering within complex capture scenarios remains however a challenging problem. A particular case is modelling complex surface light interactions for highly reflective scenes, which results in intricate high frequency specular radiance components. We hypothesize that such challenging settings can benefit from increased representation power. We hence propose a method that tackles this issue through a geometrically and physically grounded Gaussian Splatting borne radiance field, where normals and material properties are spatially variable in the primitive's local space. Using per-primitive texture maps for this purpose, we also propose to harness the GPU hardware to accelerate rendering at test time via unified material texture atlas. Code will be available at https://github.com/maeyounes/TextureSplat

Figures (9)

• Figure 1: Method Overview: We introduce planar primitive material textures — as opposed to single attributes — within physically based Gaussian Splatting rendering optimization. The increased representation power from spatially varying normal and material in object space enables fidelity reconstruction of high frequency specular in highly reflective scenes. Our hardware-accelerated implementation using texture atlases improves rendering efficiency at test time.
• Figure 2: Qualitative comparisons of novel view synthesis on synthetic scenes. From top to bottom: helmet from Shiny Blender ref_nerf and potion from Glossy Synthetic nero. Notice how we reconstruct reflections with more fidelity and less distortion.
• Figure 3: Qualitative comparisons of novel view synthesis on real scenes ref_nerf. From left to right: garden spheres and sedan. Notice how we recover reflections with more fidelity.
• Figure 4: Comparison of scene decomposition between our method and the baseline.
• Figure 5: Qualitative comparisons of normal reconstruction by different methods.