# Neural Appearance Models Cut 3D Gaussian Splatting Storage by 85% While Improving Quality

_Research · published 2026-11-19_

Researchers from TU Braunschweig, Università della Svizzera italiana, and the University of Würzburg have published a comprehensive comparison of appearance models for 3D Gaussian Splatting, introducing a compact neural alternative to the standard spherical harmonics (SH) approach. By replacing per-primitive SH coefficients with compact latent codes decoded by a small shared MLP, the method reduces per-Gaussian appearance storage from 192 to 28 bytes — an 85% reduction — while also accelerating optimization by 1.3× and matching or exceeding SH in reconstruction quality (PSNR).

The team integrated multiple appearance models, including Spherical Voronoi, NASG, and NASGabor variants, into a unified differentiable CUDA rasterizer with fused forward and backward passes, and released a portable WebGL viewer compatible with laptop and mobile GPUs. Beyond raw metrics, the paper offers a nuanced analysis of how expressive appearance models interact with scene geometry and non-static content, noting that benchmark gains can partly reflect robustness to scene movement rather than purely better radiance modeling — a finding with practical implications for how reconstruction quality should be evaluated going forward.

## Sources
- [fhahlbohm.github.io](https://fhahlbohm.github.io/efficient-gaussian-appearance/)
---
Canonical: https://genbuzz.news/posts/neural-appearance-models-cut-3d-gaussian-splatting-storage-by-85-while-improving-quality
