Active Exploration for Neural Global Illumination of Variable Scenes

Neural rendering algorithms introduce a fundamentally new approach for photorealistic rendering, typically by learning neural representation of illumination on large numbers ground truth images. When training given variable scene, such as changing objects, materials, lights, and viewpoint, the space \( \mathcal {D} \) possible data instances quickly becomes unmanageable dimensions parameters increase. We novel Active Exploration method using Markov Chain Monte Carlo, which explores , generating samples (i.e., renderings) that best help interleaves on-the-fly sample generation. self-tuning reuse strategy to minimize expensive step samples. apply our generator learns render scene an explicit parameterization configuration. Our results show trains network much more efficiently than uniformly sampling and, together with resolution enhancement approach, achieves better quality uniform at convergence. allows interactive hard light transport paths (e.g., complex caustics), require very high counts be captured, provides dynamic navigation manipulation, after 5 18 hours depending required variations.