Cleaning up motion blurred reflections

[kosso_olli]

This stuff has been bugging me for years. So thanks Lele for finding out how it's done!
I noticed too that this effect is highly dependant of the wheel geometry. The wheel I uploaded is kind of an oldschool wheel, you don't find these that often these days. The recent two-tone rims with diamond cut faces are much harder to get right.

[Wilyman]

This stuff has been bugging me for years. So thanks Lele for finding out how it's done!
I noticed too that this effect is highly dependant of the wheel geometry. The wheel I uploaded is kind of an oldschool wheel, you don't find these that often these days. The recent two-tone rims with diamond cut faces are much harder to get right.

Yep. And judging by some of the other shots in the series, the ones in the reference pic were these...



Nice big chunky chamfers everywhere.

[kosso_olli]

Nice big chunky chamfers everywhere.

Yes, this would be easy with that rim!

[Karol.Osinski]

So, it seems that for the while, the only way to properly clean up these kind of situations, without changing the scene setup, is with brute force camera sampling.
Namely, a 4-1000 (one thousand!) will render Olli's highlight perfectly fine to a noise threshold of 0.01, while not really taking any longer in the rest of the image which will never reach 1 million cam rays per pixel to achieve the given N.T..
Because of the very high camera sampling needed across the whole screen (the spinning wheel covers it all, there are no non-moblurred pixels), you will be able to shave some decent time from the rendering by lowering MSR to, say, 2.
I heard from a bird specific improvements are being studied, but this should work just fine, with minor variations, across most cases (f.e. with a slightly higher min AA, with a default MSR, and so on.), for the time being.

It goes without saying that this can be helped by lowering the intensity of the highlight in question (f.e., to 2 or 3 float, from 15-20. By whichever means.), which will shave potentially a great amount of time from the rendering.

EDIT: i should hasten to add that the sampleRate RE is your friend: if it shows red, there will be noise left in the image because max sampling has been reached before the noise threshold.
EDIT #2: below, a comparison of the original scene settings (1-32, 0.01) and the new ones (4-1000, 0.01) (left column) and the respective samplerate render elements (right column).
It took a *very* long time, however, in the parts which needed it, and are green in the Samplerate (green at 1000 subdivs. geez. XD).
Bright side? it resolved them exactly, and didn't take much longer elsewhere.

Wow Lele thanks, that was really impressive, and the results are great!. I know you guys are trying your best to popularize the usage of default settings and I totally understand it, but I love the technical side of things and I like playing with numbers so much, even though sometimes it is not worth the effort - but sure look - we all know how is it Anyway, may I ask you Lele if you you would recommend increasing the min. AA samples when dealing with highly reflective scenes? Something like a marble finished kitchen ,full of golden finishing, some metals and plenty of glass.

[^Lele^]

Wow Lele thanks, that was really impressive, and the results are great!. I know you guys are trying your best to popularize the usage of default settings and I totally understand it, but I love the technical side of things and I like playing with numbers so much, even though sometimes it is not worth the effort - but sure look - we all know how is it

Do tell! ^^

Anyway, may I ask you Lele if you you would recommend increasing the min. AA samples when dealing with highly reflective scenes? Something like a marble finished kitchen ,full of golden finishing, some metals and plenty of glass.

Well, it's going to be down to a few factors.
It's true that fixed sampling can be quicker than adaptive, for a pixel which needs to reach a specific sampling level (say, 512 samples, and a noise threshold which would require 512 samples).
It's also true that adaptive can save a *lot* of unneeded tracing in an image with uneven detail across the screen.
I wouldn't go overboard with min AA, as a result, leaving enough to ensure all the detail is found, so to be fed to the adaptive algorithm.
Say, 4 to 8 min AA subdivs ought to be plenty for most archviz scenes (64 samples per pixel will find the thinnest of geometry, i'd wager, even when motion blurred or DoFed.).
I'd go above only after experimental proof of the benefit on the specific scene (f.e., one which is *very* dark, and the adaptive sampler struggles with.), if i was to make the choice.

[Karol.Osinski]

Thanks so much for your answer ^Lele^ . I am looking forward to do some testing in my "free time" - if I can even call it like this nowadays lol.