# GS-Phong: Meta-Learned 3D Gaussians for Relightable Novel View Synthesis

**Authors:** Yumeng He, Yunbo Wang, Xiaokang Yang  
**arXiv:** [2405.20791](https://arxiv.org/abs/2405.20791)  
**Date:** 2024-05-31  
**Summary page:** https://www.summarizepaper.com/en/paper/2405.20791v1/

## Key points

- Authors Yumeng He, Yunbo Wang, and Xiaokang Yang introduce a novel method for decoupling illumination in 3D scenes
- The proposed approach involves representing a scene illuminated by a point light using relightable 3D Gaussian points inspired by the Blinn-Phong model
- Method decomposes the scene into ambient, diffuse, and specular components to allow for lifelike lighting effects synthesis
- Authors present a novel bilevel optimization-based meta-learning framework to separate geometric information from lighting conditions
- Framework treats rendering tasks under various lighting positions as a multi-task learning problem and generalizes learned Gaussian geometries across different viewpoints and diverse light positions
- Experimental results demonstrate effectiveness of the approach in terms of training efficiency and rendering quality compared to existing methods for free-viewpoint relighting
- "GS-Phong" offers an innovative solution for enhancing rendering capabilities in 3D scenes through meta-learning techniques to optimize lighting representation for more realistic visual effects

## Summary

In their paper "GS-Phong: Meta-Learned 3D Gaussians for Relightable Novel View Synthesis," authors Yumeng He, Yunbo Wang, and Xiaokang Yang introduce a novel method for decoupling illumination in 3D scenes to enable realistic novel view synthesis and relighting. The proposed approach involves representing a scene illuminated by a point light using a set of relightable 3D Gaussian points, inspired by the Blinn-Phong model. By decomposing the scene into ambient, diffuse, and specular components, the method allows for the synthesis of lifelike lighting effects. To overcome the challenge of separating geometric information from lighting conditions, the authors present a novel bilevel optimization-based meta-learning framework. This framework treats rendering tasks under various lighting positions as a multi-task learning problem and effectively generalizes learned Gaussian geometries across different viewpoints and diverse light positions. Experimental results demonstrate the effectiveness of this approach in terms of training efficiency and rendering quality compared to existing methods for free-viewpoint relighting. Overall, "GS-Phong" offers an innovative solution for enhancing rendering capabilities in 3D scenes by leveraging meta-learning techniques to optimize lighting representation and achieve more realistic visual effects. This research contributes valuable insights to the field of computer graphics and holds promise for advancing technologies related to virtual reality, augmented reality, and computer-generated imagery.