NOVA-GS: Low-Light Novel View Synthesis

Abstract

Reconstructing the Unseen

Reconstructing 3D scenes under real-world low-light conditions remains challenging due to severe sensor noise, low signal-to-noise ratios, and degraded photometric consistency, which destabilize geometry estimation and novel view synthesis. Existing approaches often rely on well-lit reference data for reliable Structure-from-Motion (SfM) initialization, or apply per-view enhancement methods that introduce cross-view inconsistencies.

We propose NOVA-GS, a unified noise-aware framework for low-light 3D Gaussian Splatting that subsumes enhancement, denoising, and geometry optimization within a single process. Our method leverages VGGT-based feed-forward estimation to obtain robust camera poses and geometry directly from degraded inputs, eliminating the need for SfM. We further introduce noise-guided spherical harmonic regularization to suppress view-dependent artifacts in noisy regions. Extensive experiments on diverse real-world low-light datasets demonstrate improved geometric fidelity, color consistency, and robustness without requiring paired supervision or well-lit references.

Key Contributions

What makes NOVA-GS different

SfM-Free Initialization

Leverages VGGT feed-forward geometry estimation directly on degraded low-light inputs. No COLMAP, no well-lit references required.

Targeted Blind-Spot Denoising

Laplacian-guided masking focuses denoising effort on high-frequency, noise-prone regions via a Deep Attention-ResUNet.

Cross-View Consistency

Depth-guided reprojection with confidence-aware photometric loss enforces global geometric coherence across all input views.

Noise-Guided SH Regularization

Maps 2D noise estimates to 3D Gaussians, penalizing higher-order spherical harmonics in noisy regions to eliminate floaters.

Method

Pipeline Overview

NOVA-GS processes low-light inputs concurrently through three coupled stages — structure-aware enhancement, self-supervised denoising, and consistency-aware 3DGS optimization — all jointly trained end-to-end.

Fig. 1 — Overview of the NOVA-GS unified optimization framework. Given low-light inputs, the pipeline concurrently performs structure-aware enhancement, self-supervised denoising, and consistency-aware 3D Gaussian Splatting optimization constrained by geometric and noise-guided spherical harmonic losses.

Visual Results

Novel View Synthesis

Drag the handles to compare the dark input images against the NOVA-GS reconstructed novel views.

◀ Low-Light Input NOVA-GS Output ▶

← drag to compare →

Quantitative Evaluation

Results on Low-Light Benchmarks

NOVA-GS achieves best performance on LLNeRF (+1.58 dB over the next best) and consistently competitive results across all datasets, without using ground-truth supervision for any stage.

Method	LOM			LLRS			LLNeRF
Method	PSNR↑	SSIM↑	LPIPS↓	PSNR↑	SSIM↑	LPIPS↓	PSNR↑	SSIM↑	LPIPS↓
2D Enhancement + GS
MBLLEN + GS	15.08	0.7008	0.3458	15.36	0.3958	0.6608	18.09	0.7011	0.3656
URetinex-Net + GS	20.29	0.8249	0.2991	15.10	0.4012	0.6670	20.08	0.8679	0.3946
NeRF-Based Methods
Aleth-NeRF †	19.56	0.7822	0.3113	12.65	0.4054	0.8985	16.02	0.7558	0.6574
LLNeRF	17.60	0.7497	0.3654	14.93	0.3446	0.7292	18.82	0.8597	0.3377
I2-NeRF †	22.40	0.7877	0.2789	15.37	0.3763	0.6493	21.91	0.6048	0.6463
3D Gaussian Splatting Methods
Luminance-GS †	17.98	0.7893	0.3110	9.78	0.3334	0.7953	12.49	0.2178	0.5291
LITA-GS	19.99	0.7988	0.3058	14.87	0.4381	0.7234	15.50	0.8515	0.3973
✦ NOVA-GS (Ours)	20.97	0.7904	0.2467	15.54	0.4088	0.6136	23.49	0.8978	0.3325

† uses ground-truth images for luminance / pose alignment. ■ Best ■ 2nd ■ 3rd