Pre-Estimated Spectral Rendering

Abstract

Spectral Monte Carlo renderers are capable of reproducing several advanced phenomena of light, such as chromatic dispersion and fluorescence. As spectral renderers must sample the spectral domain, they are typically hampered by a multitude of sampling issues leading to notably poor convergence rates, which are reinforced when realistic emission or reflectance spectra are involved in otherwise simple scenes. We propose pre-estimated spectral rendering, which is a simple method that iteratively builds estimates of spectral radiance distributions before rendering, and subsequently uses these to efficiently guide importance sampling of the spectral domain. Our method significantly lowers variance and reduces chromatic noise with little overhead in multiple difficult scenarios, which we demonstrate with an implementation in a conventional renderer.