Raising QMC AA max subdivs increases noise?

[studioDIM]

Odd issue, still with the same scene as the other post.
It's composed of sun+sky+physcam, a plane, and the max 5 (i think) gargoyle model, merged into one mesh and meshsmoothed.
There is no GI to the scene, just the material for the gargoyle is black in the diffuse channel, and 0.25 glossy reflective.
Now, when i set the QMC AA parameters to 1-4 (0.55 adaptive amount, noise 0.003 threshold for both qmc sampler and aa), i get a very nicely cleaned image, particularly in the dark shades (which have multiple glossy reflections going on).
When i just raise the AA to 1-16, it takes 25 seconds more to render it, but th noise somehow increases.
I would like to understand what pitfall i dropped into.
I am thinking sampling it more "finds" more detail, hence the noise (as the new detail isn't sampled enough).
It just does not concord with the fact that thresholds and adaptiveness are the same, so the noise should be somewhat higher when max subdivs is at 4.
here are the two samples:


1-4


1-16

Download and zoom in on the plant of the foot to see the evident differences.
Any idea?
Thanks for the time,

Lele

[Morbid Angel]

sounds like your adaptive amount prevents samples to be properly taken in the shaded areas. Because of image being really dark in shadows usually its hard to resolve noise there. Perhaps raising the adaptive amount will make it better.

[vlado]

I think the logic would be something like this.

When you have less AA subdivs, more of the adaptive sampling is done by the material; since you have the adaptive amount to 0.55 that means that the material will shoot first nearly half of the available samples for an evaluation, regardless of how the result compares against the noise threshold, and then take additional samples, if necessary.

When you have more AA subdivs, the sampling is moved more towards the image sampler, while less samples are taken for each material evaluation. The image sampler however can compare the results with nearby pixels to see if the result is below the noise threshold and may decide to terminate the sampling earlier.

In any case, the difference is relatively small; things were a lot worse in previous versions of V-Ray.

Best regards,
Vlado

[Paul Oblomov]

Vlado, can new sampler (LSS) help in this particular case ?

[studioDIM]

Aaaaah, ok.

So i basically overrode the material settings when raising the AA, introducing more noise coming from the AA pass, whereas when the max subdivs were lower, i would use the material higher sampling, and only "later" the AA would kick in.

Thanks for this one too, Vlado.

Lele

[vlado]

Vlado, can new sampler (LSS) help in this particular case ?

Not really, it does not change the amount and place of samples taken; the LSS sampler only changes which those samples are.

Best regards,
Vlado