Research map / 3D vision / Gaussian splatting
Luminance-GS: Adapting 3D Gaussian Splatting to Challenging Lighting Conditions with View-Adaptive Curve Adjustment
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Adds per-view color matrix mapping and a view-adaptive tone curve (global plus attention-generated per-view bias) to 3DGS, trained with unsupervised losses, to render well-lit novel views from low-light, overexposed, or varying-exposure inputs.
Problem
Novel view synthesis from multi-view images captured under challenging lighting (low-light, overexposure, varying exposure) where photometric inconsistency across views degrades 3DGS and NeRF reconstructions.
Approach
Keep the original 3DGS representation and learn per-view 3x3 color matrices and view-adaptive tone curves (a shared global curve plus a per-view bias from a cross-attention generator) that map inputs to pseudo-enhanced, view-consistent images used only during training, with unsupervised curve and spatial losses.
Key results
- On LOM low-light, achieves SSIM 0.799 and LPIPS 0.294, best on both, with PSNR 18.34 (second to Aleth-NeRF's 19.87).
- On LOM overexposure, achieves best PSNR 20.71, SSIM 0.835, LPIPS 0.222, outperforming MSLT+GS and others.
- On varying exposure (Mip-NeRF 360 subset), reports best PSNR/SSIM/LPIPS in the bicycle, garden and counter scenes except bicycle PSNR, at ~14 min training and ~150 FPS.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use view-adaptive curve adjustment and per-view color matrix mapping | 3DGS · LOM low-light subset (mean) | SSIM: 0.754 -> 0.799 | Helped |
| use view-adaptive curve adjustment and per-view color matrix mapping | 3DGS · LOM low-light subset (mean) | PSNR: 19.87 -> 18.34 | Hurt |
| use view-adaptive curve adjustment and per-view color matrix mapping | 3DGS · LOM overexposure subset (mean) | LPIPS: 0.262 -> 0.222 | Helped |
| use view-adaptive curve adjustment | 3DGS · Mip-NeRF 360 bicycle (varying exposure) | SSIM: 0.618 -> 0.640 | Helped |
| use view-adaptive curve adjustment | 3DGS · Mip-NeRF 360 garden (varying exposure) | PSNR: 20.49 -> 21.01 | Helped |
| use view-adaptive curve adjustment | 3DGS · Mip-NeRF 360 counter (varying exposure) | PSNR: 18.92 -> 19.81 | Helped |
| add the global curve to the view-specific curve bias (ablation) | Luminance-GS · LOM buu (low-light) | SSIM: 0.653 -> 0.859 | Helped |
| add the global curve to the view-specific curve bias (ablation) | Luminance-GS · LOM buu (overexposure) | SSIM: 0.712 -> 0.804 | Helped |
| add the global curve to the view-specific curve bias (ablation) | Luminance-GS · Mip-NeRF 360 garden (varying exposure) | SSIM: 0.777 -> 0.788 | Helped |
| add per-view color matrix mapping (ablation) | Luminance-GS with global and view-specific curves · LOM buu (low-light) | SSIM: 0.859 -> 0.877 | Helped |
| add per-view color matrix mapping (ablation) | Luminance-GS with global and view-specific curves · LOM buu (overexposure) | SSIM: 0.804 -> 0.811 | Helped |
Surprising findings
- Per-view independent curves alone overfit and degrade quality; a single shared curve also causes view inconsistency.
- Pre-processing images with 2D enhancers before 3DGS training or post-processing renders performs worse than the proposed approach.
Limitations
- Does not generalize across scenes; the method is per-scene and aims to handle only illumination variation in future work.
- PSNR on low-light LOM is second-best behind Aleth-NeRF.
- Varying-exposure dataset is synthesized rather than captured.
Experimental setup
Datasets: LOM dataset Mip-NeRF 360
Benchmarks: LOM dataset low-light LOM dataset overexposure Mip-NeRF 360 varying exposure
Metrics: PSNR SSIM LPIPS FPS GPU hours
Baselines: 3DGS Aleth-NeRF AME-NeRF NeRF-W GS-W Zero-DCE SCI NeRCo LLVE SGZ MSEC IAT
Builds on: 3D Gaussian Splatting Zero-DCE NeRF-W Aleth-NeRF
Closest papers by meaning
- GS-IR: 3D Gaussian Splatting for Inverse Rendering (CVPR 2024 · similarity 0.90)
- Wild-GS: Real-Time Novel View Synthesis from Unconstrained Photo Collections (NeurIPS 2024 · similarity 0.89)
- GI-GS: Global Illumination Decomposition on Gaussian Splatting for Inverse Rendering (ICLR 2025 · similarity 0.89)
- MS-GS: Multi-Appearance Sparse-View 3D Gaussian Splatting in the Wild (NeurIPS 2025 · similarity 0.89)
- 3D-HGS: 3D Half-Gaussian Splatting (CVPR 2025 · similarity 0.88)
- Quantifying and Alleviating Co-Adaptation in Sparse-View 3D Gaussian Splatting (NeurIPS 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.