[Bf-cycles] Volumetric patch update
kartochka22 at yandex.ru
Mon Aug 6 12:40:46 CEST 2012
Still too buggy/unfinished for wide testing.
Finished Markov chain sampling of most bi-dir contributions, direct eye,
indirect eye, background. That had lot of bumps, crash in
light_distribution() when rand_t is exactly 1.0, now i clamp result to
max array number as as dirty hack. Still uncertain about camera edge
random walk, is it need to keep at integer pixel centers, or apply float
point fraction-of pixel steps.
In short, now MCMC can sample any light sources in case of paths started
from eye, so background, sun must work (un addition to already worked
point with radius=0 and emissive triangles). I have quick test of
indoor-like scene lit with small window and sun, - at least it not worse
then classic ^^.
Beware, as full bi-dir samples utilised, we get exploding of render
Actual MCMC Samples ~= samples * mutation * (max light_bounce * max
eye_bounce), try to keep mutation number low for experiments, or you
will wait minutes for even first picture appear. Mutations 8-12 is good
for start, not that dumb but only 10 times slower.
Not all bi-dir contributions work, there is still long time
unimplemented connect_both(), it require a lot of code, i postpone that
as lazy man. Will boost noise cleaning in many complex scenes, as more
possibility to connect paths, and MIS magic will help to dim bad hot
pixels. Will see.
As well, volume sampling probability for bi-dir is plain wrong, i now it
can render cool things, but it must be fixed anyway, sometime it produce
very wrong result.
-------------- next part --------------
A non-text attachment was scrubbed...
Size: 50993 bytes
Desc: not available
Url : http://lists.blender.org/pipermail/bf-cycles/attachments/20120806/3b7b3158/attachment.bin
More information about the Bf-cycles