Noise level needs to be calculated from rendered content only

[pokoy]

I have some images that I'm rendering against photographed backplates, used in Direct Visibility Override slot. The rendered model makes only a small portion of the whole image which leads to a super low noise level from the start, even if the rendered model is noisy as hell.
It seems the noise level is calculated over the entire image area which may lead to very inaccurate values. Another negative side effect is is that if I have a noisy image background this may skew the whole noise level calculation and lead to a situation where a desired noise level is never met.

I'd like to request noise level to be calculated on actually rendered content only. Otherwise it's pretty much impossible to make sure image quality meets a certain level, I imagine this to be true for animation especially where the percentage of rendered content against a background may change a lot.

[Ludvik Koutny]

While this definitely does need fixing, adaptivity will always need to be evaluated for background as well. BG texture needs to be antialiased, otherwise it will come out jagged, jittery, or very noisy when in DoF or Motion-Blurred.

As a temporary solution, I would suggest to set noise limit to 0, and use pass limit instead. This way, image will still be evaluated adaptively, but won't stop prematurely. You will just need to do some test to see what pass limit is about right...

[pokoy]

As a middle ground solution we could have a simple checkbox in CoronaOutputMap that lets us exclude that map from noise level calculations.
I agree that DOF and mblur will need sampling, but we definitely need more control over this in case the background is really clean and doesn't need any sampling other than a few passes of AA.

[Ludvik Koutny]

As a middle ground solution we could have a simple checkbox in CoronaOutputMap that lets us exclude that map from noise level calculations.
I agree that DOF and mblur will need sampling, but we definitely need more control over this in case the background is really clean and doesn't need any sampling other than a few passes of AA.

Nope, that's not really a solution. If there was just regular environment (not fed through output map), and your model did not take large part of the image, exactly same thing will happen. This needs complex final solution, not implementing any middle ground workarounds ;)