sugestion of possible improvement of performance

[ProkopHapala]

Hello,
when I was rendering something with indigo I find that the biggest reason why is rendering with indigo so slowly is that you must wait until a few strange pixels (with extrem values in relation to the background)


I thing this can be done more quickly (with smaler total amount of mautations) if there wil be simple 2d pixel analyser, which compute gradients (for example; numerical derivatives in ale directions, very computionaly cheap and simple) and find areas of big discontinuity and cast more rays in this areas

also it can be generalised because in areas where the image (as a function f(x,y)) is close to be constant there can by the image blured, or something like thist and it needn't to be computed so precisely in the contrast the areas with big variation (where the gradinet is large) needs to be computed in much bigger detail with mode rays.

maybe can be used some typical methods of numerical mathematics. Local interpolation and aproximation can by find in gowing number of control points.
and the quality - in some areas degree of convergence can be find by mesurement how much is the pixel-value changing in time or when the new ray in this pixel is casted.

I know, these are cheap tricks, not very sofisticated, but maybe it can help.
I'm sorry if I'm wrong, I really don't understand therory of unbiased renders (I must read something).

thank's for indigo, I'm realy happy that secound most fine render in the world is free. It looks like a big chalange to develope it.
//and sorry for my english

[Zom-B]

Improving Indigo by the option to interact depending on Pixel Information
of the rendered Image wold be wonderfull and could be also usefull for other interesting features...

Something to think about would be to use the GPU for the Noise Detection,
Saving some CPU Time...
Just check the links at the last post!

It could be Also interesting to detect Fireflies by comparing the renderoutput over some time
(aka Noise that stays for long time) > seek & destroy

[ProkopHapala]

Zom-B > Yes, I thought similarly.
seek and destroy...for refining almost rendered image find strange artifacts or inproper areas and render more only there.

however I don't thing that this task (2d analysis) could take any relevant cpu-time (not needed to use some GPU help)

but still, some more sofisticated analysis (such as Radiance posted) which can quantify something like numerical instability of resulting pixel-value based on the rendering algorithm itselfs would be better than simple 2d noise statistics, I agree. or maybe combination.