[Bf-cycles] Sobol sequence quality and SOBOL_SKIP=64

storm kartochka22 at yandex.ru
Mon Jun 17 18:12:04 CEST 2013

I read paper about recently added jmc sampler algorithm, and found line
about Sobol sequence, that it have "hierarhy" property that make it
perfect for Cycles like case, progressive render, as (that is my own
understand how it work) first samples (that close to 0) very well
distributed from low frequency POV, and later samples better cover high
order dimension discrepance. In other words, first samples are very good
to get overall guess of typical image when we have low number of
samples, and later samples good in make it in high order case, like in
hard complex weird geometry scenes.

But in kernel_random.h we are skipping first 64 samples. I have tryed to
replace it by 0, and got surprisely good result, many volume scenes
looks more "even" noise moire pattern at low sample number then before.
It is far from scientific, pure my subjective quess. So, if you feel
interested, recompile Blender with SOBOL_SKIP=0 and do some tests. It
have zero impact to speed, just different noise pattern.

BTW, I always turn off russian roulette algorithm (using same min=max
bounces) in my tests, maybe it matter for that SOBOL_SKIP.

More information about the Bf-cycles mailing list