Fast Low-light Enhancement and Deblurring for 3D Dark Scenes
Authors: Feng Zhang, Jinglong Wang, Ze Li, Yanghong Zhou, Yang Chen, Lei Chen, Xiatian Zhu
Organizations: Nanjing University of Posts and Telecommunications, Nanjing. · The Hong Kong University of Science and Technology, Hong Kong. · The Hong Kong Polytechnic University, Hong Kong. · University of Surrey, United Kingdom.
Novel view synthesis from low-light, noisy, and motion-blurred imagery remains a valuable and challenging task. Current volumetric rendering methods struggle with compound degradation, and sequential 2D preprocessing introduces artifacts due to interdependencies. In this work, we introduce FLED-GS, a fast low-light enhancement and deblurring framework that reformulates 3D scene restoration as an alternating cycle of enhancement and reconstruction. Specifically, FLED-GS inserts several intermediate brightness anchors to enable progressive recovery, preventing noise blow-up from harming deblurring or geometry. Each iteration sharpens inputs with an off-the-shelf 2D deblurrer and then performs noise-aware 3DGS reconstruction that estimates and suppresses noise while producing clean priors for the next level. Experiments show FLED-GS outperforms state-of-the-art LuSh-NeRF, achieving 21× faster training and 11× faster rendering.
Figures & tables
Figure 1: Overview of the pipeline for FLED-GS . Our framework consists of three main stages: Pre-processing, Progressive Iterative Enhancement, and Rendering. In the pre-processing stage, low-light images are enhanced step-by-step with varying histogram equalization and Gamma parameters to achieve progressive brightness adjustment. In the progressive iterative enhancement stage, the framework progressively performs low-light enhancement, deblurring, and denoising. Finally, the framework finishes training and leverages sharp normal-light 3DGS to render novel-view images.
Table 1: Quantitative comparison of different methods on synthetic scenes of the LuSh-NeRF dataset [ 16 ] . A : The best result; A : The second best result.
Figure 2: Qualitative results of different methods on the LuSh-NeRF dataset. The upper scenes are synthetic scenes, and the lower scenes are real scenes.
Method
PSNR ↑
SSIM ↑
LPIPS ↓
Retinexformer [ 1 ] +3DGS [ 8 ]
18.33
0.755
0.359
LITA-GS [ 19 ]
18.04
0.728
0.369
Aleth-NeRF [ 5 ]
19.87
0.754
0.417
LuSh-NeRF [ 16 ]
21.03
0.726
0.265
FLED-GS(2 rounds)
22.77
0.824
0.205
Table 2: Quantitative results on the extremely low-light LOM dataset [ 5 ] , averaged across all scenes.
Methods
Training time
Rendering time
Model Parameters
LuSh-NeRF [ 16 ]
14h30min
9.1s
1.24M
FLED-GS(2 rounds)
41min
0.8s
12.17M
FLED-GS(3 rounds)
1h8min
FLED-GS(4 rounds)
1h34min
Table 3: Efficiency and complexity comparison. Rendering time refers to the average time required to render a single image. Model parameters are averaged across scenes.
Configuration
PSNR ↑
SSIM ↑
LPIPS ↓
w/o PIE, w/o NE
21.65
0.5831
0.2254
w/ PIE, w/o NE
21.94
0.6250
0.2444
w/o PIE, w/ NE
21.79
0.5885
0.2372
Full FLED-GS
22.60
0.6576
0.2312
Table 4: Effect of PIE and NE on synthetic scenes. The metric represents the average across synthetic scenarios.
Aerospace Information Research Institute, Chinese Academy of Sciences · Key Laboratory of Target Cognition and Application Technology (TCAT) · University of Chinese Academy of Sciences +2