Efficient compression in 3D gaussian splatting

Abstract

Recent advancements in 3D Gaussian Splatting (3DGS) have made real-time photore- alistic novel view synthesis by explicitly representing scenes using Gaussian primitive. However, the substantial memory footprint required to store full feature embeddings across multiple resolution levels is a significant barrier for scalable applications. In this work, we explore a memory-efficient enhancement to the Hash-grid Assisted Con- text (HAC) framework by integrating a softened dynamic codebook, inspired by the success of Variable Bitrate Neural Field (VBR-NF) with a Straight-Through Estimator (STE) based differentiable indexing mechanism. Our proposed method replaces con- ventional hash embedding storing full feature embeddings at each spatial coordinate with a learnable feature indices matrix and a global codebook, enabling memory effi- ciency. Experimental results on datasets such as Tanks&Temples, DeepBlending and BungeeNRF reveal that our method achieves comparable rendering quality to HAC while significantly reducing memory usage at the embedding level. Nonetheless, the overall memory falls short of the HAC baseline introduced by additional anchor at- tributes. This suggests the trade-off between embedding level compression and memory overhead in other elements. We analyze these challenges and propose future directions for spatially adaptive indexing strategies to better balance efficiency and quality.