One bounce, two bounces

The diminishing returns of depth

A ray from the camera strikes a diffuse wall. To know the colour of that point, the renderer must estimate how much light arrives there — which means casting new rays out into the scene. Those rays strike other surfaces, which need their own incoming light estimated, and so the recursion begins. Bounce depth is simply how many times a path is allowed to continue before it is terminated.

At zero bounces, you see only surfaces that face a light directly; everything else is black. One bounce adds indirect light from the first reflection off any surface — the soft fill that spills into shadow, the colour bleed from a red wall onto a white floor. This is the single most important step. The difference between zero and one indirect bounce is usually enormous; the image goes from theatrical to credible.

The second bounce adds light that has bounced twice before reaching a surface: illumination that has turned a corner twice, the gentle glow deep inside a sofa cushion, the secondary colour cast in a corner. The gain is real but significantly smaller than the first step.

By the third bounce, returns have diminished further. Most scenes are dark enough in their deep recesses that paths reaching there carry little energy. In a bright exterior with large open skies, paths rarely need more than two or three bounces to converge. In a closed interior lit by a single window, energy must scatter many more times to fill the space, and truncating the depth too early leaves it unnaturally dark — the image cannot simply invent the missing light.

Physically, energy is lost at every non-specular interaction. A surface with 70 percent albedo passes on only 70 percent of the incoming light; after three such bounces, the surviving energy is around 34 percent of the original. After five, roughly 17 percent. Russian roulette termination formalises this intuition: paths with low surviving energy are terminated randomly, with unbiased compensation applied to the survivors, so the estimator remains correct on average while cutting wasted work.

The practical ceiling for most production diffuse paths sits somewhere between four and six bounces. Beyond that, variance from the extra paths is larger than the signal they contribute — you pay in noise what you can barely see you gained in light.