Research map / 3D vision / NeRF and radiance fields
Aleth-NeRF: Illumination Adaptive NeRF with Concealing Field Assumption
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 Aleth-NeRF, which adds learnable "concealing field" transmittance terms to NeRF and uses unsupervised lightness losses to render normal-light novel views from low-light or over-exposed multi-view images, plus a new LOM dataset.
Problem
Standard NeRF assumes normal-light capture; under low-light or over-exposed inputs, it entangles illumination and material and fails to recover normal-light novel views.
Approach
Model extra light attenuation in the air via a local concealing field (conv over density MLP output) and a global learnable concealing field multiplied into accumulated transmittance; train on adverse-light images with unsupervised degree, contrast, and color constancy losses so rendering without (low-light) or with (over-exposure) the concealing field yields normal-light views.
Key results
- In low-light settings on LOM, Aleth-NeRF reaches mean PSNR 19.87, SSIM 0.754, LPIPS 0.417, the best mean across methods.
- In over-exposure settings on LOM, Aleth-NeRF reaches mean PSNR 18.09 and SSIM 0.745, best in both, but LPIPS (0.488) is weaker than some baselines.
- Integrating Aleth-NeRF into Instant-NGP cuts training time from 2.2 hours on 4 GPUs (NeRF) to 4.8 minutes on 1 GPU, with lower PSNR (18.09 vs 19.14) on the buu scene.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Aleth-NeRF with concealing fields instead of vanilla NeRF | NeRF · LOM low-light mean | PSNR: 6.84 -> 19.87 | Helped |
| Use Aleth-NeRF with concealing fields instead of vanilla NeRF | NeRF · LOM low-light mean | SSIM: 0.150 -> 0.754 | Helped |
| Use Aleth-NeRF with concealing fields instead of vanilla NeRF | NeRF · LOM low-light mean | LPIPS: 0.582 -> 0.417 | Helped |
| Use Aleth-NeRF with concealing fields instead of vanilla NeRF | NeRF · LOM over-exposure mean | PSNR: 9.60 -> 18.09 | Helped |
| Use Aleth-NeRF with concealing fields instead of vanilla NeRF | NeRF · LOM over-exposure mean | SSIM: 0.660 -> 0.745 | Helped |
| Use Aleth-NeRF with concealing fields instead of vanilla NeRF | NeRF · LOM over-exposure mean | LPIPS: 0.457 -> 0.488 | Hurt |
Surprising findings
- Pipelines that apply enhancement methods after NeRF rendering ('NeRF + enhancer') generally underperform pipelines that enhance training views first ('enhancer + NeRF') in low-light settings, especially for LPIPS.
- Aleth-NeRF achieves best PSNR and SSIM on over-exposure but not best LPIPS, which the authors attribute to training directly on degraded over-exposed images.
Limitations
- Aleth-NeRF must be trained separately for each scene, as with vanilla NeRF.
- May fail in scenes with non-uniform lighting or shadows.
- LPIPS on over-exposed scenes is weaker than some baselines, and the method relies on unsupervised training directly on degraded images.
Experimental setup
Datasets: LOM (self-collected low-light, normal-light, over-exposure multi-view dataset)
Benchmarks: LOM
Metrics: PSNR SSIM LPIPS
Baselines: NeRF RetiNexNet Zero-DCE SCI LLVE MBLLEN RAW-NeRF HE IAT MSEC
Builds on: NeRF Instant-NGP inverse tone curve
Closest papers by meaning
- Physical-aware Neural Radiance Fields for Efficient Exposure Correction (AAAI 2025 · similarity 0.91)
- Rad-NeRF: Ray-decoupled Training of Neural Radiance Field (NeurIPS 2024 · similarity 0.90)
- A Pre-convolved Representation for Plug-and-Play Neural Illumination Fields (AAAI 2024 · similarity 0.90)
- A General Implicit Framework for Fast NeRF Composition and Rendering (AAAI 2024 · similarity 0.90)
- NeRF On-the-go: Exploiting Uncertainty for Distractor-free NeRFs in the Wild (CVPR 2024 · similarity 0.90)
- Bright-NeRF: Brightening Neural Radiance Field with Color Restoration from Low-Light RAW Images (AAAI 2025 · 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.