# LightGaussian: Unbounded 3D Gaussian Compression with 15x Reduction and 200+ FPS

**Authors:** Zhiwen Fan, Kevin Wang, Kairun Wen, Zehao Zhu, Dejia Xu, Zhangyang Wang  
**arXiv:** [2311.17245](https://arxiv.org/abs/2311.17245)  
**Date:** 2023-11-28  
**Summary page:** https://www.summarizepaper.com/en/paper/2311.17245v1/

## Key points

- Recent advancements in real-time neural rendering using point-based techniques
- Traditional approaches like 3D Gaussian Splatting suffer from storage overhead
- LightGaussian aims to transform 3D Gaussians into a more efficient format
- LightGaussian identifies and prunes Gaussians that do not contribute significantly to scene reconstruction
- Distillation and pseudo-view augmentation techniques are used to distill spherical harmonics to a lower degree
- VecTree Quantization is introduced as a hybrid scheme for quantizing attributes
- LightGaussian achieves an impressive compression rate of over 15x while boosting FPS from 139 to 215
- Efficient representation of complex scenes on various datasets such as Mip-NeRF 360, Tank, and Temple
- The paper provides detailed insights into the methodology used in LightGaussian along with experimental results supporting its effectiveness

## Summary

Recent advancements in real-time neural rendering using point-based techniques have led to the widespread adoption of 3D representations. However, traditional approaches like 3D Gaussian Splatting suffer from a significant storage overhead when dealing with large-scale scenes, requiring gigabyte-level disk space for a single unbounded scene. This poses scalability challenges and hinders the efficiency of splatting. To address this issue, the authors propose LightGaussian, a novel method that aims to transform 3D Gaussians into a more efficient and compact format. Inspired by the concept of Network Pruning, LightGaussian identifies Gaussians that do not contribute significantly to scene reconstruction and applies a pruning and recovery process. This effectively reduces redundancy in Gaussian counts while preserving visual effects. In addition, LightGaussian employs distillation and pseudo-view augmentation techniques to distill spherical harmonics to a lower degree. This allows for knowledge transfer to more compact representations while maintaining reflectance accuracy. Furthermore, the authors introduce VecTree Quantization, a hybrid scheme that quantizes all attributes, resulting in lower bitwidth representations with minimal accuracy losses. Overall, LightGaussian achieves an impressive compression rate of over 15x while boosting the frames per second (FPS) from 139 to 215. This enables efficient representation of complex scenes on various datasets such as Mip-NeRF 360, Tank, and Temple. The proposed method addresses the storage overhead challenge associated with traditional approaches and significantly improves both compression efficiency and rendering speed. The paper is authored by Zhiwen Fan, Kevin Wang, Kairun Wen, Zehao Zhu, Dejia Xu and Zhangyang Wang. It provides detailed insights into the methodology used in LightGaussian along with experimental results supporting its effectiveness. More information about the project can be found on their website: https://lightgaussian.github.io/.