Aggressive compression of 3D Gaussian positions without rendering error

Determine whether the 3D center positions of Gaussians in the 3D Gaussian Splatting scene representation can be aggressively compressed without introducing significant rendering error, and, if so, develop an effective positional compression method that achieves substantial memory reduction while preserving novel view synthesis quality.

Background

The paper introduces a compression pipeline for 3D Gaussian Splatting that clusters color (spherical harmonics) and shape parameters into codebooks, applies quantization-aware fine-tuning, and uses entropy encoding to achieve up to 31× compression with minimal quality loss. Despite these advances, the authors identify positions of Gaussians (their 3D centers) as a major remaining memory component that resists aggressive compression.

The authors report unsuccessful attempts to quantize positions to a lattice structure and to incorporate positional constraints directly into the Gaussian splatting training process, noting that these approaches led to significant rendering errors. This highlights a specific unresolved issue: effectively compressing the positional information of Gaussians without degrading rendering quality.