GPU-based Acceleration of Path Guiding for Many-Light Global Illumination

Abstract

Rendering with global illumination has been used mainly in the field of film production thanks to its highly photorealistic image synthesis. When using Monte Carlo raytracing, which is one of the major rendering methods, the resulting images may contain some points that deviate significantly from final convergence values. It is needed to generate output images with a small number of samples and low variance by improving the variance reduction per sample. Path guiding is a method for reducing such variance by using importance sampling of the estimated radiance distribution in advance to construct efficient paths. However, in many-light scenes, estimating radiance distribution may be slow due to the complexity of distribution obtained by the light sources and it could make variance rather increased by constructing paths from the trained distribution of low accuracy. In this paper, therefore, we propose a method for accelerating the computational efficiency of path guiding with many-light global illumination by sophisticating light source sampling to better estimate radiance distribution. The proposed method was implemented as a GPU program, and we empirically proved that it reduced up to 46.8% of variance compared to the normal path tracing.