GS-PI: An Optimization-Decoupled Appearance Decomposition Approach for Generating PBR Gaussian Assets

arXiv cs.GR details GS-PI, an optimization-decoupled pipeline for generating PBR Gaussian assets from pre-trained Gaussian splats. The core idea is to stop treating illumination and material recovery as one fragile optimization problem, and instead predict PBR attributes with a geometry-conditioned diffusion pass before distilling the result back through differentiable rasterization.

For developers, the practical appeal is relightability: the method aims to preserve view consistency while avoiding the baked-in radiance that makes many Gaussian assets awkward to use in a physically based pipeline. The paper says GS-PI works directly in 3D point-cloud space, which helps sidestep the correspondence problems that can trip up 2D diffusion approaches and reduce cross-view misalignment.

The method combines a global semantic prior, source-anchored photometric cues, and learned view-direction conditioning to keep specular highlights out of intrinsic colors. It also claims to replace expensive per-scene joint illumination/BRDF optimization with a learned diffusion stage plus a shorter target-driven distillation step, and it does so without requiring proxy meshes.

If this holds up outside the paper’s benchmarks, it could be a useful bridge between fast Gaussian scene capture and production-friendly PBR workflows. That matters most for teams experimenting with neural assets, relighting tools, or hybrid pipelines that need editable materials instead of permanently lit captures.