Low-light images remain a hard case for 3D Gaussian Splatting (3DGS): they hide scene detail, shift colors, and corrupt the geometric priors that the method relies on. Simple image enhancement can restore visibility, but it often leaves behind chromatic deviations and small artifacts that then spread through multi-view reconstruction.

Naka-GS, a framework described in a recent arXiv preprint, tackles both problems at once. Its Naka-Guided Chroma Correction module uses a fixed bionic response curve as a photometric anchor and learns bounded, frequency-decoupled corrections for luminance and chroma, which preserves fine structures while suppressing residual color and boundary artifacts. A separate, non-learned Point-Prior Compaction Module then refines the dense scene priors produced from the corrected views, using voxel pooling and distance-adaptive progressive pruning.

On eight scene-disjoint RealX3D scenes, the authors report 20.735 dB PSNR, 0.7408 SSIM, and 0.3340 LPIPS. The compaction module cuts the initial point count from 1.00 million to 59.52 thousand and shortens average 3DGS optimization time by 37.7%, all while keeping the standard 3DGS renderer and training objective unchanged.

This summary is based solely on the arXiv preprint; the source does not include independent comparisons with other low-light 3DGS methods. The authors state that code is available.