Research map / 3D vision / NeRF and radiance fields
Learning Neural Exposure Fields for 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.
Introduces Neural Exposure Fields (NExF), a per-3D-point MLP predicting optimal exposure jointly trained with a ZipNeRF-based radiance field for 3D-consistent view synthesis under strong exposure variation.
Problem
Neural view synthesis degrades on real-world captures with strong per-image exposure variation, producing over/underexposed regions and 2D-tonemapping-induced 3D inconsistency.
Approach
Learn a neural field mapping each 3D point to an optimal exposure value, supervised by well-exposedness and saturation weights of rendered pixels, and jointly optimized with the radiance field via log-exposure conditioning on the MLP bottleneck.
Key results
- On HDRNeRF, ID exposure PSNR 42.54 vs. 39.07 for HDRNeRF and 41.10 for HDR-GS, while training in ~11 min vs. 542 min for HDRNeRF.
- On Eyeful Tower, PSNR 26.48 vs. 22.82 for HDRNeRF* (a 57% MSE improvement per the paper) and LPIPS 0.234 vs. 0.263 for affine GLO.
- Ablation on Eyeful Tower shows removing well-exposedness drops LPIPS to 0.247 and removing regularization drops PSNR to 24.84.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use NExF instead of HDR-GS | HDRNeRF, ID exposure (t1, t3, t5) · HDRNeRF, ID exposure (t1, t3, t5) | PSNR: 41.10 -> 42.54 | Helped |
| Use NExF instead of HDRNeRF | HDRNeRF, OOD exposure (t2, t4) · HDRNeRF, OOD exposure (t2, t4) | PSNR: 37.53 -> 38.36 | Helped |
| Use NExF instead of HDRNeRF* | Eyeful Tower · Eyeful Tower | PSNR: 22.82 -> 26.48 | Helped |
| Use NExF instead of HDRNeRF* | Eyeful Tower · Eyeful Tower | SSIM: 0.836 -> 0.876 | Helped |
| Use NExF instead of Affine GLO | Eyeful Tower · Eyeful Tower | LPIPS: 0.263 -> 0.234 | Helped |
| add the proposed view-dependent MLP architecture (ablation) | NExF · HDRNeRF, ID exposure (t1, t3, t5) | PSNR: 33.85 -> 42.54 | Helped |
| add the proposed view-dependent MLP architecture (ablation) | NExF · HDRNeRF, OOD exposure (t2, t4) | PSNR: 28.20 -> 38.36 | Helped |
| add latent exposure conditioning (ablation) | NExF · HDRNeRF, ID exposure (t1, t3, t5) | PSNR: 39.88 -> 42.54 | Helped |
| add latent exposure conditioning (ablation) | NExF · HDRNeRF, OOD exposure (t2, t4) | PSNR: 38.33 -> 38.36 | No clear effect |
| add the well-exposedness criterion (ablation) | NExF · Eyeful Tower | LPIPS: 0.247 -> 0.234 | Helped |
| add the saturation criterion (ablation) | NExF · Eyeful Tower | PSNR: 24.30 -> 26.48 | Helped |
| add exposure-field regularization (ablation) | NExF · Eyeful Tower | PSNR: 24.84 -> 26.48 | Helped |
Surprising findings
- Without affine GLO embeddings, PSNR on Eyeful Tower slightly increases (27.44 vs 26.48) even though SSIM and LPIPS worsen.
Limitations
- Results degrade under extreme low-light or extreme over-exposure.
- Complex lighting effects such as semi-transparency and strong reflections are not handled well.
Experimental setup
Datasets: HDRNeRF dataset Eyeful Tower dataset v2 ZipNeRF dataset (Alameda) in-the-wild phone captures (statue)
Benchmarks: HDRNeRF dataset Eyeful Tower dataset v2
Metrics: PSNR SSIM LPIPS
Baselines: NeRF ZipNeRF 3DGS NeRF-W HDRNeRF HDR-GS GLO Affine GLO HDRNeRF*
Builds on: Zip-NeRF NeRF-W HDRNeRF generative latent optimization Bilateral Guided Radiance Fields
Closest papers by meaning
- ExtraNeRF: Visibility-Aware View Extrapolation of Neural Radiance Fields with Diffusion Models (CVPR 2024 · similarity 0.88)
- Efficient-3Dim: Learning a Generalizable Single-image Novel-view Synthesizer in One Day (ICLR 2024 · similarity 0.88)
- 3D Reconstruction with Generalizable Neural Fields using Scene Priors (ICLR 2024 · similarity 0.88)
- Denoising Diffusion via Image-Based Rendering (ICLR 2024 · similarity 0.88)
- NC-SDF: Enhancing Indoor Scene Reconstruction Using Neural SDFs with View-Dependent Normal Compensation (CVPR 2024 · similarity 0.87)
- Decompositional Neural Scene Reconstruction with Generative Diffusion Prior (CVPR 2025 · similarity 0.87)
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.