(L) [2013/07/18] [Zelcious] [Re: Optimization strategies] Wayback!>> ingenious wrote:Are you saying that many shorter chains is much worse than fewer longer chains? On what scenes? I find it a bit hard to believe that having a larger number of chains is always worse.
A tiny lightsource in combination with SDS path will certainly do. Only small mutations will be efficient and will require longer chains to actually move somewhere in screen space. Eventually large number of chains will catch up but initially it could be quite a bit of a difference.
(L) [2013/07/19] [spectral] [Re: Optimization strategies] Wayback!>> ingenious wrote:Are you saying that many shorter chains is much worse than fewer longer chains? On what scenes? I find it a bit hard to believe that having a larger number of chains is always worse.
In fact If you do MLT on the GPU with less chains (say 65000) you will have the "impression" that it is better... (at least with PT & caustics), because the caustics appear more quickly... but it is not a good solution... because for other pixels/chains you also need a good convergence (by example for simple diffuse surfaces).
By example, add a simple prism on a plane and a small light, the caustics will take a lot of time to appear and to be noise free, so in this case it is more interesting to have fewer chains (and then more iterations)... because the plane (diffuse by example) will be noise-free before !
So, what you have to do is to find the right balance... because the goal is to have a good "global" convergence and not good convergence only for caustics and hards paths...
It really depends of the scene...
BTW, multi-try will only help to find "better chains", but will also require more "global processing", so even on the GPU I'm not sure it can help !
(L) [2013/07/19] [ingenious] [Re: Optimization strategies] Wayback!Fixing the number of light transport paths sampled, the trade-off between the number of chains and their length makes a continuum between crude Monte Carlo and Veach's 1-chain Metropolis. (ERPT is nothing more than just setting this trade-off parameter much closer to crude MC, which in turn essentially corresponds to chain length of 1). It'd be quite interesting to see some results from exploring this continuum under different types of illumination conditions.