Research map / 3D vision / Gaussian splatting
LITA-GS: Illumination-Agnostic Novel View Synthesis via Reference-Free 3D Gaussian Splatting and Physical Priors
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.
Proposes LITA-GS, an unsupervised 3D Gaussian Splatting method that renders illumination-invariant structure, depth, illumination and noise attributes to synthesize normal-exposure novel views from low-light or over-exposed images without ground-truth references.
Problem
Novel view synthesis from images captured under adverse illumination (low light or over-exposure) with no normally-exposed references, where SfM initialization is sparse and exposure correction applied per view causes multi-view illumination inconsistency.
Approach
Extracts an illumination-invariant structure prior from Kubelka-Munk physical theory and invariant edge detectors, renders it together with depth, illumination and noise maps as extra per-Gaussian attributes, and removes noise with a progressive denoising module trained with unsupervised exposure-control, structure, denoising and reconstruction losses.
Key results
- On low-light LOM scenes, LITA-GS reaches 21.14 PSNR, 0.829 SSIM and 0.233 LPIPS on average, surpassing the prior end-to-end SOTA Aleth-NeRF by 1.27 dB PSNR and 0.075 SSIM.
- On over-exposed LOM scenes, LITA-GS reaches 20.17 PSNR, 0.841 SSIM and 0.273 LPIPS on average, versus 18.09 PSNR for Aleth-NeRF.
- Ablation on the bike scene: removing the progressive denoising module drops PSNR from 22.75 to 22.36; removing all of PDM, Dr and Pr drops it to 21.55.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use LITA-GS instead of Aleth-NeRF | LOM low-light novel view synthesis · LOM low-light (mean) | PSNR: 19.87 -> 21.14 | Helped |
| use LITA-GS instead of Aleth-NeRF | LOM low-light novel view synthesis · LOM low-light (mean) | SSIM: 0.754 -> 0.829 | Helped |
| use LITA-GS instead of 3DGS + Zero-DCE | LOM low-light novel view synthesis · LOM low-light (mean) | LPIPS: 0.380 -> 0.233 | Helped |
| use LITA-GS instead of Aleth-NeRF | LOM over-exposure novel view synthesis · LOM over-exposure (mean) | PSNR: 18.09 -> 20.17 | Helped |
| use LITA-GS instead of 3DGS + IAT | LOM over-exposure novel view synthesis · LOM over-exposure (mean) | SSIM: 0.811 -> 0.841 | Helped |
| use LITA-GS instead of 3DGS + IAT | LOM over-exposure novel view synthesis · LOM over-exposure (mean) | LPIPS: 0.303 -> 0.273 | Helped |
| add the progressive denoising module (ablation) | LITA-GS · bike low-light | PSNR: 22.36 -> 22.75 | Helped |
| add the progressive denoising module (ablation) | LITA-GS · bike low-light | LPIPS: 0.289 -> 0.282 | Helped |
| add depth map rendering (ablation) | LITA-GS · bike low-light | SSIM: 0.799 -> 0.819 | Helped |
| add illumination-invariant structure prior rendering (ablation) | LITA-GS · bike low-light | SSIM: 0.782 -> 0.819 | Helped |
| add PDM, depth rendering, and structure prior rendering (ablation) | LITA-GS · bike low-light | PSNR: 21.55 -> 22.75 | Helped |
Surprising findings
- Per-view exposure correction followed by 3DGS (pre-processing) gives highly unstable results across scenes, sometimes worse than the vanilla baseline, as the enhancement is not multi-view consistent.
Limitations
- Relies on COLMAP for initialization and pose estimation, which struggles under adverse illumination.
- Evaluated only on the five-scene LOM dataset with low-light and over-exposed captures.
Experimental setup
Datasets: LOM dataset
Benchmarks: LOM dataset
Metrics: PSNR SSIM LPIPS
Baselines: NeRF 3DGS Zero-DCE SCI IAT MSEC Aleth-NeRF LLNeRF
Builds on: 3D Gaussian Splatting Aleth-NeRF Marigold COLMAP Kubelka-Munk theory
Closest papers by meaning
- 3D Geometry-Aware Deformable Gaussian Splatting for Dynamic View Synthesis (CVPR 2024 · similarity 0.88)
- Novel View Synthesis Under Large-Deviation Viewpoint for Autonomous Driving (AAAI 2025 · similarity 0.87)
- MetaGS: A Meta-Learned Gaussian-Phong Model for Out-of-Distribution 3D Scene Relighting (NeurIPS 2025 · similarity 0.87)
- FreeSplat: Generalizable 3D Gaussian Splatting Towards Free View Synthesis of Indoor Scenes (NeurIPS 2024 · similarity 0.87)
- Generative Sparse-View Gaussian Splatting (CVPR 2025 · similarity 0.86)
- Structure Consistent Gaussian Splatting with Matching Prior for Few-shot Novel View Synthesis (NeurIPS 2024 · similarity 0.86)
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.