Research map / 3D vision / NeRF and radiance fields
LuSh-NeRF: Lighting up and Sharpening NeRFs for Low-light Scenes
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 LuSh-NeRF, an unsupervised NeRF method that reconstructs clean, sharp, normally lit scenes from hand-held low-light blurry images by sequentially removing noise (via multi-view consistency) and camera blur (via low-frequency-filtered trajectory prediction), plus a new LOL-BlurNeRF dataset.
Problem
Reconstructing NeRFs from hand-held low-light images that simultaneously suffer from low visibility, sensor noise, and camera-shake motion blur, where existing low-light enhancement, deblurring, and NeRF methods fail when combined.
Approach
Model the implicit degradation order (low intensity, then noise, then blur) inside NeRF training: a Noise-Estimator decomposes view-inconsistent noise using multi-view feature consistency (Scene-Noise Decomposition), and a camera trajectory predictor (Rigid Blurring Kernel) is trained on low-frequency, noise-filtered rays (Camera Trajectory Prediction) to sharpen the scene.
Key results
- On the synthetic LOL-BlurNeRF scenes, LuSh-NeRF averages 19.31 PSNR, 0.5853 SSIM and 0.2914 LPIPS, the best LPIPS among compared combinations.
- Against the strongest baseline LEDNet + NeRF (18.40 PSNR, 0.6488 SSIM, 0.3204 LPIPS), LuSh-NeRF improves LPIPS but not SSIM, which the authors attribute to the unfair LEDNet setup trained on LOL-Blur.
- Ablation shows SND-then-sharpen ordering (Denoise->Sharp) beats Sharp->Denoise (LPIPS 0.2914 vs 0.3089 average).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add SND and CTP to preprocessing + NeRF (ablation) | NeRF reconstruction · Synthetic scenes average | LPIPS: .3908 -> .2914 | Helped |
| add CTP without SND (ablation) | preprocessing + NeRF · Synthetic scenes average | LPIPS: .3908 -> .3082 | Helped |
| add SND without CTP (ablation) | preprocessing + NeRF · Synthetic scenes average | LPIPS: .3908 -> .3689 | Helped |
| use SND and CTP rather than CTP alone (ablation) | preprocessing + NeRF · Synthetic scenes average | LPIPS: .3082 -> .2914 | Helped |
| denoise before sharpening rather than sharpen before denoising (ablation) | LuSh-NeRF · Synthetic scenes average | LPIPS: .3089 -> .2914 | Helped |
| use frequency filtering instead of RGB intensity thresholding (ablation) | CTP · Synthetic scenes average | LPIPS: 0.3015 / 0.3437 / 0.3447 -> 0.2878 | Mixed |
| remove thresholding in CTP (ablation) | LuSh-NeRF · Synthetic scenes average | LPIPS: 0.2878 -> 0.3114 | Hurt |
| set CTP filter radius to 10 rather than 30 (ablation) | LuSh-NeRF · Synthetic scenes average | LPIPS: 0.2878 -> 0.2968 | Hurt |
| set CTP filter radius to 50 rather than 30 (ablation) | LuSh-NeRF · Synthetic scenes average | LPIPS: 0.2878 -> 0.3009 | Hurt |
| use LuSh-NeRF rather than LEDNet + NeRF | NeRF reconstruction · Synthetic scenes average | LPIPS: 0.3204 -> 0.2914 | Helped |
Surprising findings
- Higher PSNR and SSIM for some baselines (e.g., LEDNet + NeRF) do not correspond to better perceptual NeRF renderings, so the authors prioritize LPIPS.
- The COLMAP-free NeRF variant fails on low-light blurry scenes, as poses are inaccurate when estimated directly from the images.
Limitations
- Training requires optimizing two NeRF networks and rendering multiple rays for blur, making optimization slower.
- Noise that is consistent across views may be modeled as scene content and is hard to remove due to reliance on viewpoint consistency.
- Real-scene pose estimation relies on LEDNet-enhanced images since COLMAP struggles on raw low-light images.
Experimental setup
Datasets: LOL-Blur
Benchmarks: LOL-BlurNeRF
Metrics: PSNR SSIM LPIPS
Baselines: NeRF LLFormer + MPRNet + NeRF URetinexNet + Restormer + NeRF LEDNet + NeRF Restormer + LLNeRF MPRNet + LLNeRF PairLIE + DP-NeRF URetinexNet + DP-NeRF
Builds on: NeRF Deblur-NeRF DP-NeRF GIM COLMAP
Closest papers by meaning
- Bright-NeRF: Brightening Neural Radiance Field with Color Restoration from Low-Light RAW Images (AAAI 2025 · similarity 0.91)
- NeRF On-the-go: Exploiting Uncertainty for Distractor-free NeRFs in the Wild (CVPR 2024 · similarity 0.91)
- DynaMoDe-NeRF: Motion-aware Deblurring Neural Radiance Field for Dynamic Scenes (CVPR 2025 · similarity 0.89)
- Optimize the Unseen - Fast NeRF Cleanup with Free Space Prior (NeurIPS 2025 · similarity 0.89)
- Mitigating Motion Blur in Neural Radiance Fields with Events and Frames (CVPR 2024 · similarity 0.89)
- NeRF-VPT: Learning Novel View Representations with Neural Radiance Fields via View Prompt Tuning (AAAI 2024 · similarity 0.89)
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.