Research map / 3D vision / NeRF and radiance fields
PBR-NeRF: Inverse Rendering with Physics-Based Neural 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.
Adds two physics-based losses (energy conservation and NDF-weighted specular separation) to a NeILF++ neural inverse-rendering pipeline to jointly estimate geometry, materials and illumination.
Problem
Inverse rendering from multi-view images is ill-posed due to material-lighting ambiguity; NeRF/3DGS methods ignore materials and illumination, and existing neural IR methods produce baked-in specular highlights in diffuse albedo.
Approach
Two physically motivated loss terms applied to the Disney BRDF: a ReLU-style energy conservation penalty on the hemispherical integral of the BRDF, and an NDF-weighted penalty (detached NDF, softmax temperature) on the diffuse lobe in specular regions, which pushes energy into the specular lobe.
Key results
- On the NeILF++ dataset, PBR-NeRF reaches mean albedo PSNR 20.08 vs 17.29 for NeILF++ (reproduced) and mean roughness PSNR 22.47 vs 21.83.
- On DTU, mean RGB PSNR improves to 28.63 from 27.85 for NeILF++, and mean Chamfer Distance is 0.828 mm vs 1.167 mm.
- Ablation: adding both losses improves albedo PSNR by 3.28 over the NeILF++ baseline without Lambertian loss (16.80 to 20.08), while RGB PSNR stays at about 31.27.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add Conservation of Energy Loss (ablation) | NeILF++ without physics-based losses · NeILF++ dataset (6-scene mean) | Albedo PSNR: albedo PSNR increasing by 2.34 | Helped |
| add NDF-weighted Specular Loss (ablation) | NeILF++ without physics-based losses · NeILF++ dataset (6-scene mean) | Albedo PSNR: albedo PSNR increasing by 2.35 | Helped |
| add Conservation of Energy Loss and NDF-weighted Specular Loss (ablation) | NeILF++ without physics-based losses · NeILF++ dataset (6-scene mean) | Albedo PSNR: 3.28 PSNR | Helped |
| add both physics-based losses (ablation) | NeILF++ without physics-based losses · NeILF++ dataset (6-scene mean) | RGB PSNR: maintaining on-par performance in RGB | No clear effect |
| add physics-based losses (IDs 2–4) (ablation) | NeILF++ without physics-based losses · NeILF++ dataset (6-scene mean) | Metallicness SSIM | Hurt |
Surprising findings
- The NDF-weighted specular loss and energy conservation loss each individually raise albedo PSNR by about 2.3 dB, while RGB PSNR is essentially unchanged.
Limitations
- Lighting estimation is not evaluated on DTU because LDR images and lack of ground-truth environment maps make it unreliable.
- Performance on Env scenes is slightly below SG-ENV for albedo, due to the implicit lighting MLP being under-constrained.
- Metallicness SSIM is slightly lower than NeILF and NeILF++.
Experimental setup
Datasets: NeILF++ synthetic dataset DTU
Benchmarks: NeILF++ dataset DTU
Metrics: PSNR SSIM Chamfer Distance
Baselines: NeILF NeILF++ PhySG SG-ENV NeRFactor VolSDF NeuS
Builds on: NeILF++ NeILF Disney BRDF NeuS VolSDF
Closest papers by meaning
- Inverse Rendering of Glossy Objects via the Neural Plenoptic Function and Radiance Fields (CVPR 2024 · similarity 0.91)
- UE4-NeRF:Neural Radiance Field for Real-Time Rendering of Large-Scale Scene (NeurIPS 2023 · similarity 0.89)
- Decoupling Scattering: Pseudo-Label Guided NeRF for Scenes with Scattering Media (AAAI 2025 · similarity 0.88)
- Rad-NeRF: Ray-decoupled Training of Neural Radiance Field (NeurIPS 2024 · similarity 0.88)
- NeRFPrior: Learning Neural Radiance Field as a Prior for Indoor Scene Reconstruction (CVPR 2025 · similarity 0.88)
- NerfBaselines: Consistent and Reproducible Evaluation of Novel View Synthesis Methods (NeurIPS 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.