Hi all,
I decided to post a few demo clips of tests I been doing.
[attachment=1]Rise.mov[/attachment]
Particles: 760.647 -> 275.782.976 via Repopulation
Rendertime: 1 second -> 5.30 min per frame average
[attachment=0]Type_Repop.mov[/attachment]
Particles: 300.000 -> 11.000.000 via Repopulation
Rendertime: 1 second -> 28 seconds per frame average
nice!
Here are some more details about the second test. I am sure Daniel will share his own experience with the problems and solutions he found…
Originally, the particle count from RealFlow was, as already mentioned, only 300,000.
Rendering the animation with just these particles produced a relatively solid result, although the typical SPH pattern in the particle distribution was obvious at closer inspection.
thinkboxsoftware.box.com/s/wd1k09g109nxp7whdr6e
(Click the Down Arrow icon in the upper right corner to download the full resolution file)
Unfortunately, once some extreme Depth Of Field was added to the mix, the results were quite disturbing.
Krakatoa DOF works by taking the particle’s position and jittering it within the “circle of confusion”, drawing many samples with lower transparency, producing an out of focus fuzzy circle. But around the focal distance, esp. given the slow pan of the camera, there were tons of noise. Each particle’s position would be jittered slightly, and when you add the filtering over 2x2 pixels against the final rendered image, it became pretty much unusable:
thinkboxsoftware.box.com/s/oph1id34yp2qagctbfq5
Increasing the DOF Sample Rate normally helps with improving the quality of the out-of-focus areas, but in this case the trouble was in the in-focus parts. Going from 0.1 Sample Rate to 1.0 improved the close up large circles, but did not get rid of the noise in the rest of the image, and caused longer render times.
RULE 1 OF KRAKATOA, once again, with passion: “If you see individual particles, your density is probably too high, your particle count is too low, or both!” 
So Daniel decided to give Repopulation a try, increasing the number of particles from 300K to 11 million.
The result was the file he posted in the beginning of this thread - the gaps between the RealFlow particles are now gone, and the DOF looks solid. The flickering is gone, too:
thinkboxsoftware.box.com/s/xm9ksm0aj22pirojgrdi
Render time went up from about 1 second (with no DOF) or 3 seconds (with DOF 0.1 Sample Rate) to 28 seconds per frame. Given that we increased the particle count 36 times, this is fairly reasonable.
We hope this gives you some additional ideas about how the render-time particle Repopulation can help you make low-count simulations look better!
Great breakdown! Thanks a lot