Research map / 3D vision / Gaussian splatting
Lighting Every Darkness with 3DGS: Fast Training and Real-Time Rendering for HDR View Synthesis
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.
Adapts 3D Gaussian Splatting to noisy nighttime RAW images for fast-training, real-time HDR novel view synthesis, using cone scatter initialization, a primitive-biased color MLP, and depth distortion/near-far regularizers.
Problem
HDR view synthesis from noisy nighttime RAW images is slow to train and cannot render in real time with volumetric NeRF methods, while vanilla 3DGS struggles with low-SNR SfM initialization, SH color representation for linear RAW color, and poor scene structure for refocusing.
Approach
Extends 3DGS to RAW linear HDR by scattering extra initialization points in a camera-derived viewing frustum, replacing spherical harmonics with a tiny exp-activated color MLP with per-primitive color bias, and adding depth distortion and near-far regularizers for accurate structure.
Key results
- LE3D reaches 61.08 RAW PSNR and 24.70 sRGB PSNR on the RawNeRF test scenes, versus 58.69 RAW PSNR for RawNeRF, while training in 1.53 GPU hours versus 129.54 for RawNeRF.
- LE3D renders at 103 FPS at 2K resolution, compared with 0.022 FPS for RawNeRF, a speedup of roughly 4000x.
- Ablations show removing the color MLP gives the worst RAW PSNR (59.45), and removing either depth distortion or near-far regularization lowers scores relative to full LE3D.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use LE3D instead of RawNeRF | HDR view synthesis · RawNeRF test scenes | FPS: 0.022 -> 103 | Helped |
| Use LE3D instead of RawNeRF | HDR view synthesis · RawNeRF test scenes | training time (GPU×H): 129.54 -> 1.53 | Helped |
| Use LE3D instead of RawGS | 3DGS trained on noisy RAW images · RawNeRF test scenes | RAW PSNR: 59.2834 -> 61.0812 | Helped |
| Use LE3D instead of RawNeRF | HDR view synthesis · RawNeRF test scenes | sRGB LPIPS: 0.4952 -> 0.5071 | Hurt |
| Add Color MLP instead of using SH (ablation) | LE3D · RawNeRF test scenes | RAW PSNR: 59.4483 -> 61.0812 | Helped |
| Add Color MLP instead of using SH (ablation) | LE3D · RawNeRF test scenes | sRGB PSNR: 23.1884 -> 24.6984 | Helped |
| Add Color MLP instead of using SH (ablation) | LE3D · RawNeRF test scenes | sRGB LPIPS: 0.5635 -> 0.5071 | Helped |
| Add depth distortion regularization (ablation) | LE3D · RawNeRF test scenes | RAW PSNR: 60.5202 -> 61.0812 | Helped |
| Add depth distortion regularization (ablation) | LE3D · RawNeRF test scenes | sRGB PSNR: 24.3615 -> 24.6984 | Helped |
| Add depth distortion regularization (ablation) | LE3D · RawNeRF test scenes | sRGB LPIPS: 0.5087 -> 0.5071 | Mixed |
| Add near-far regularization (ablation) | LE3D · RawNeRF test scenes | RAW PSNR: 60.7144 -> 61.0812 | Helped |
| Add near-far regularization (ablation) | LE3D · RawNeRF test scenes | sRGB PSNR: 24.5705 -> 24.6984 | Helped |
Surprising findings
- The full LE3D is slower to train than the RawGS baseline (1.53 vs 1.05 GPU hours) and renders at 103 FPS, slower than the 176 FPS of RawGS, with the depth regularizers adding cost.
Limitations
- Refocusing is slow because of large Gaussian blur kernels, so real-time performance does not hold for that task.
- Cone Scatter Initialization is not exercised on the test scenes, which are all indoor, so its benefit is shown only qualitatively on distant views.
- Quantitative results are on only three test scenes from one dataset captured with one phone camera.
Experimental setup
Datasets: RawNeRF dataset
Benchmarks: RawNeRF dataset
Metrics: PSNR SSIM LPIPS FPS training time (GPU hours)
Baselines: RawNeRF LDR-NeRF LDR-3DGS HDR-3DGS RawGS
Builds on: 3D Gaussian Splatting RawNeRF Mip-NeRF 360 COLMAP
Closest papers by meaning
- IBGS: Image-Based Gaussian Splatting (NeurIPS 2025 · similarity 0.89)
- Compact 3D Gaussian Representation for Radiance Field (CVPR 2024 · similarity 0.89)
- NexusGS: Sparse View Synthesis with Epipolar Depth Priors in 3D Gaussian Splatting (CVPR 2025 · similarity 0.89)
- FewViewGS: Gaussian Splatting with Few View Matching and Multi-stage Training (NeurIPS 2024 · similarity 0.89)
- LP-3DGS: Learning to Prune 3D Gaussian Splatting (NeurIPS 2024 · similarity 0.89)
- 3D-HGS: 3D Half-Gaussian Splatting (CVPR 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.