Thorsten Froehlich <438887d6$1@news.povray.org> Saturday 26 of November 2005
17:05

> would end up not taking samples, which would unbalance the distribution as
> a
> whole.  This in turn would create some strong but unpredictable (probably
> appearing randomly and visual as noise) bias in the samples.  Obviously
> this would make things a lot worse.

Concider checker plane y,0 with too small scale, it is very very aliased,
and if we are unlucky then in example first 4 rays could gave identical
result quickly fooling statistical function.

I did some testing with minimal samples setting, and It seems for checker it
is better(*) then any combination of variance/confidence/samples of regular
algorithm.

But I can be mistaking ofcourse, I will play a bit with that.

(*) better = in possible small render time eliminate the random very brights
pixels. I do not talk here about the "usual" quality of focal blur, but
about 100% green or 100% blue pixels showing when rendering very aliased
pigment { checker } plane.


-- 

Post a reply to this message