LumiModeling: A Gaussian Splatting-Based Tool for Recreating Dynamic Material and Lighting Interaction in Architecture

Abstract

This thesis presents LumiModeling, a real-time visualization tool based on Gaussian Splatting (GS) that simulates the dynamic interplay between materiality and lighting in architectural environments. While conventional design workflows rely on geometric modeling and photorealistic rendering, they often abstract complex material behaviors and fall short in capturing light-material interactions. In contrast, GS enables the reconstruction of high-fidelity 3D models from 2D image sets, representing viewdependent effects such as reflection, transparency, and surface roughness. A comparative analysis using real-world data from the MIT Stata Center and the Met Warehouse demonstrates GS’s advantages over mesh-based photogrammetry, particularly in rendering reflective and transparent materials. This work extends existing GS capabilities by implementing a relightable pipeline based on the existing model Relightable3DGaussian (Gao et al., 2023), in which each Gaussian point is augmented with physical parameters, including BRDF, surface normals, and incident lighting. The Stata Center dataset is used to test the relighting of GS. A user study involving architecture professionals reveals that perceptual focus shifts from geometry to materiality and lighting as visual realism increases. The findings highlight the potential of relightable GS in architectural visualization and anticipate its integration into future design workflows.