Chapter 20. GPU-Based Importance Sampling — Colbert, Křivánek

• Source: https://developer.nvidia.com/gpugems/gpugems3/part-iii-rendering/chapter-20-gpu-based-importance-sampling
• Volume: GPU Gems 3
• Note: add-only excerpt; images omitted.

Summary of full chapter read

Real-time glossy IBL via Monte Carlo in one shader; minimal precomputation vs wavelet/SH prefilters.

Estimator

Illumination integral Lo(v)=∫ Li(u) f(u,v) cosθ du. Sample BRDF PDF (normalized Phong lobe); weight by 1/pdf. Hammersley low-discrepancy sequence precomputed as constants (deterministic → aliasing).

Filtered importance sampling

LOD from solid angle of sample: mip level ≈ ½ log2((w·h)/N) − ½ log2(pdf) − ½ log2(d(u)), plus bias +1. Low pdf → blur more of env map. Dual-paraboloid maps (scale b=1.2) because cube-map faces mip independently. Diffuse via SH; glossy via this sampling. ~24 samples complex geo, ~40 for smooth spheres. Materials/env fully dynamic.