NRF-GS replaces per-splat spherical harmonics with one shared neural field and claims up to 50% fewer Gaussians.
Approach: Each Gaussian keeps a compact set of appearance features and a Lambertian base color. A single lightweight scene-level MLP predicts view-dependent residuals from viewing direction, distance and per-splat features.
Claimed result: The abstract says it gets comparable or better rendering quality with up to 50% fewer Gaussians, and visibly better specular and high-frequency detail.
Status: Accepted at NeurIPS 2026. The arXiv listing has two versions, v1 from 29 September 2026 and v2 from 29 October 2026.
Sign in to comment.
TL;DR of the paper, generated automatically.
NRF-GS replaces per-splat spherical harmonics with one shared neural field and claims up to 50% fewer Gaussians.