Research map / 3D vision / NeRF and radiance fields
A Pre-convolved Representation for Plug-and-Play Neural Illumination Fields
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 Neural Illumination Fields (NeIF), which model per-point incident lighting as a NeRF-style volumetric radiance field, combined with integrated lobe encoding and a multiscale pre-convolved background supervision for plug-and-play relighting and material decomposition.
Problem
Decomposing lighting and materials from multi-view images under non-distant, spatially varying illumination, where environment maps and spherical Gaussians assume infinitely distant light and per-point illumination networks lack spatial constraints.
Approach
Treat each 3D sample of the environment as a light emitter by representing incident illumination as a volumetric radiance field integrated along a roughness-adaptive vMF specular lobe, with an integral lobe encoding and multiscale Gaussian pre-convolved background supervision to stabilize decomposition.
Key results
- On Glossy Blender, the full method reaches 37.985 PSNR / 0.981 SSIM / 0.031 LPIPS versus 25.851 PSNR for NeILF.
- On real-world scenes, the method reaches 25.022 PSNR versus 21.664 for NeILF (averaged over scenes).
- Ablations show removing the background (w/o BG) drops Glossy Blender PSNR from 37.985 to 36.404 and removing ILE drops it to 37.362.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use NeIF instead of NeILF* | Novel-view rendering · Glossy Blender | PSNR: 27.279 -> 37.985 | Helped |
| Use NeIF instead of NeILF | Novel-view rendering · Real-world | PSNR: 21.664 -> 25.022 | Helped |
| Add integrated lobe encoding (ILE) (ablation) | NeIF · Glossy Blender | PSNR: 37.362 -> 37.985 | Helped |
| Add background guidance (ablation) | NeIF · Glossy Blender | PSNR: 36.404 -> 37.985 | Helped |
| Add multiscale pre-convolved representation (ablation) | NeIF · Glossy Blender | PSNR: 37.748 -> 37.985 | Helped |
| Add bilateral smoothness regularization (ablation) | NeIF · Glossy Blender | PSNR: 38.442 -> 37.985 | Hurt |
Surprising findings
- Removing the bilateral smoothness regularization yields the highest PSNR (38.442) on Glossy Blender but produces noisy, inconsistent materials unsuitable for editing.
Limitations
- Relies on geometry reconstructed in stage one (NeuS-style SDF), so decomposition depends on geometry quality.
- Real-world results degrade due to varying exposures, rough geometry, and limited background coverage.
Experimental setup
Datasets: Glossy Blender (self-rendered) BlendedMVS Bag of Chips Mip-NeRF Garden scene
Benchmarks: Glossy Blender BlendedMVS Bag of Chips
Metrics: PSNR SSIM LPIPS
Baselines: NeILF NeILF* Ne-Env
Builds on: Mip-NeRF NeILF NeuS Ref-NeRF PhySG
Closest papers by meaning
- IllumiNeRF: 3D Relighting Without Inverse Rendering (NeurIPS 2024 · similarity 0.91)
- Few-Shot Neural Radiance Fields under Unconstrained Illumination (AAAI 2024 · similarity 0.90)
- Aleth-NeRF: Illumination Adaptive NeRF with Concealing Field Assumption (AAAI 2024 · similarity 0.90)
- A General Implicit Framework for Fast NeRF Composition and Rendering (AAAI 2024 · similarity 0.90)
- ROGR: Relightable 3D Objects using Generative Relighting (NeurIPS 2025 · similarity 0.90)
- Physical-aware Neural Radiance Fields for Efficient Exposure Correction (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.