[Bf-cycles] Volumetric patch update

storm kartochka22 at yandex.ru
Tue Nov 5 14:26:42 CET 2013

Fixed uninitialized pixel x,y in MCMC/MLT code, no more huge brightness
jumps in MCMC/MLT mode (Mutations>0). New debug checkbox "No importance
sampling". I found myself always trying to turn it off for diffusion,
and let MCMC discover that cosine pdf itself, so there is option (was
hardcoded it every time using #ifdefs). Few experiments show it only get
things worse.

With that bug fixed, i have tried to render simple animation, like
moving cube in fog, 2 bounce (max bounce = 1) in MCMC mode. Result is
mixed, brightness looks good now, but noise converging speed not very
good. As i target MCMC bidirectional mode as primary, that need some
speedup, maybe 2x to be usable. Also, as no QMC used, converging become
much worse then Sobol QMC, as 1/ SQRT(N). Have ideas how to utilize QMC
for Markov Chain, and i read papers that it possible, at least for
independent MC, but no working code yet.

Short plan is test it in more hard light scenes, if no big bugs found
then cleanup debugging things. Maybe with other help it will be merged
after all GsoC project hit trunk, as that MCMC things is fun to test,
even if slower then unpatched Cycels and unimplemented things.

Have idea how to handle transparent shadows, as bidir now work badly
with it, so volume things in transparent cubes have doubled light. Need
to add third temporal chain, insert vertices when shadow_blocked() used,
and respect it in MIS calculations. But it will increase code a lot,
postpone it, as more important things must be solved.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: blender_cycles_volume_mlt_spectral_61119.patch.bz2
Type: application/x-bzip
Size: 82074 bytes
Desc: not available
Url : http://lists.blender.org/pipermail/bf-cycles/attachments/20131105/01d0a9af/attachment.bin 

More information about the Bf-cycles mailing list