Research map / Architectures and efficiency / RNNs and normalization
Real-Time Neural BRDF with Spherically Distributed Primitives
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 NeuBRDF, a material-agnostic neural BRDF that factorizes the 4D BRDF into two HEALPix hemisphere feature grids with a shared VQ codebook of neural primitives and a tiny MLP, enabling real-time evaluation.
Problem
Representing measured, simulated, and spatially varying BRDFs with compact memory while allowing real-time evaluation in rendering, since 4D BRDF inputs make feature-grid acceleration structures hard to build.
Approach
Factorize the BRDF into two hemispherical feature grids (incident and outgoing) discretized with HEALPix sphere pixelation, store per-grid integer indices into a codebook of neural reflectance primitives shared across materials, and decode with a tiny MLP.
Key results
- On MERL image-based evaluation, NeuBRDF reaches MAE 0.0017 and SSIM 0.996, versus 0.0028 / 0.995 for NBRDF and 0.0056 / 0.990 for NPF.
- Evaluating NeuBRDF takes about 4.5 ms in a 5-bounce path tracer at 1920x1080 (1 SPP) on an i9-12900K and RTX 3090, reaching 89 FPS overall in the Figure 1 setting.
- A BTF of about 500 MB is compressed to roughly 95 KB NeuBRDF plus a 131 KB neural texture on UBO2014, a compression ratio of thousands.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use NeuBRDF instead of NBRDF | MERL reflectance representation · MERL | MAE: 0.0028 -> 0.0017 | Helped |
| use NeuBRDF instead of NBRDF | MERL reflectance representation · MERL | RMSE: 0.0033 -> 0.0031 | Helped |
| use NeuBRDF instead of NBRDF | MERL reflectance representation · MERL | SSIM: 0.995 -> 0.996 | Helped |
| remove codebook quantization and store bulky grid features (ablation) | NeuBRDF · MERL raw data | MSE: 0.0167 -> 0.0160 | No clear effect |
| remove codebook quantization and store bulky grid features (ablation) | NeuBRDF · MERL raw data | Memory Cost: 0.07MB -> 1.60MB | Hurt |
| replace codebook quantization with low-bit (4b) quantization (ablation) | NeuBRDF · MERL raw data | MSE: 0.0167 -> 0.0189 | Hurt |
| replace codebook quantization with low-bit (4b) quantization (ablation) | NeuBRDF · MERL raw data | Memory Cost: 0.07MB -> 0.40MB | Hurt |
| use HEALPix sphere pixelation instead of the other tested pixelations (ablation) | NeuBRDF · MERL, RGL, and UBO | MSE | Helped |
| use Fibonacci sphere pixelation instead of HEALPix (ablation) | NeuBRDF · MERL, RGL, and UBO | indexing time: 5ms versus 0.5ms | Hurt |
Surprising findings
- Codebook quantization at 9-bit reaches MSE 0.0167 at 0.07 MB, which is close to the unquantized 0.0160 at 1.60 MB, while 4-bit low-bit quantization is worse (0.0189).
Limitations
- Quality on anisotropic and sparse measurements is deferred to supplementary material, so main-text evidence is limited to isotropic MERL-style data.
- Spatially varying extension relies on a simple neural texture plug-in, which the paper presents as naive.
Experimental setup
Datasets: MERL RGL LayeredBRDF Special Coatings UBO2014 UBOFAB19
Benchmarks: MERL RGL UBO2014 Special Coatings LayeredBRDF
Metrics: MAE RMSE SSIM MSE runtime (ms) FLOPs
Baselines: GGX Low et al. microfacet model Dupuy et al. factorization NPF NBRDF NLBRDF Neural Biplane
Builds on: Instant-NGP VQ-NeRF codebook (vector quantization feature grids) NBRDF HEALPix Neural Biplane
Closest papers by meaning
- Inverse Rendering of Glossy Objects via the Neural Plenoptic Function and Radiance Fields (CVPR 2024 · similarity 0.85)
- HyRF: Hybrid Radiance Fields for Memory-efficient and High-quality Novel View Synthesis (NeurIPS 2025 · similarity 0.85)
- NeRFPrior: Learning Neural Radiance Field as a Prior for Indoor Scene Reconstruction (CVPR 2025 · similarity 0.85)
- Differentiable Neural Surface Refinement for Modeling Transparent Objects (CVPR 2024 · similarity 0.85)
- NeISF: Neural Incident Stokes Field for Geometry and Material Estimation (CVPR 2024 · similarity 0.85)
- PBR-NeRF: Inverse Rendering with Physics-Based Neural Fields (CVPR 2025 · similarity 0.85)
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.