Reproduce “RNG: Relightable Neural Gaussians” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jiahui Fan, Fujun Luan, Jian Yang, Miloš Hašan, Beibei Wang
RNG: Relightable Neural Gaussians introduces a framework that extends 3D Gaussian Splatting to create high-quality, relightable 3D assets, achieving real-time rendering at 60 fps and significantly faster training compared to prior methods. It captures intricate details and accurate shadows, particularly for objects with complex or soft boundaries, by avoiding explicit surface constraints and analytical shading assumptions.
Key insights
- Implicitly learning reflectance through latent vectors and a neural decoder allows for accurate relighting of complex materials and ill-defined shapes without rigid analytical shading models.
- Integrating a learned shadow cue and a depth refinement network enables the generation of high-quality, sharp shadows, addressing a common challenge in neural rendering.
- A hybrid two-stage optimization, combining forward shading for geometry and deferred shading for appearance and shadow refinement, effectively leverages the advantages of different rendering paradigms.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.