Importance sampling is the key method for increasing the efficiency of
realistic image synthesis. The common approach is to importance sample
the local reflectance term only. This is not sufficient in complex
scenes. Rendering such scenes would greatly benefit from importance
sampling according to incident illumination, and indeed many such
algorithms were proposed, but none is currently widely used. We analyze
their deficiencies, and suggest an approach, that solves the previous
problems: it supports fast sampling, can be used with caching, is able
to capture high-frequency signals (such as caustics), and can be easily
plugged into many existing algorithms.