Adaptivity clamp

[jackieteh]

Thanks for this. Turns out the bigger issue in that shot was some flickering caustics. But I did increase the Adaptivity Clamp value a bit to make sure it covers the entire range. Thanks.

nice discussion, would like how to check how much Clamp Value we need to increase?

[Joelaff]

I went by the brightest secular highlight in the scene, not necessarily light sources themselves. But would be interested in hearing Vlado’s recommend.

[^Lele^]

It's down to what you'll do with the image in Post.
For the sake of argument, let me simplify this: the value specified in the control will be the value threshold in the image beyond which noise will be high.
So, if you will use the render exactly as it is, the default value of 1.5 will show absolutely no visible difference in the image, as all one can see is the 0.0-1.0 interval.
However, if for some reason you ended up halving exposure in post, then the 1.5f adaptivity clamp range would be halved as well, and your noisy areas would now sit starting at the 0.75f range, and as such they'd be visible.

So, it's a form of prediction: set it as low as you can considering the entire image history after the render is completed.

[Joelaff]

Good way of putting it, Lele. Generally have not had to touch it, but we had some shots recently that were more heavily tone mapped in post, with values more in the 0-3.0 or so range. Bumping it up to 3.0 does seem to have made a very slight difference.

Like anything else in 3d test, test, test, or set overly high, eat the render time, and don't worry about it. I like to test when I can so I can avoid the big render hits when possible.

[^Lele^]

As an important corollary, the higher noise level should not be a problem at all with a denoiser: the noiseLevel RE will direct it properly.
And as the Adaptivity kernel is bigger than one pixel anyway, there should be very little edge spillage and detail loss while denoising around the areas where the adaptivity clamp kicks in.
By this i mean that the noise level will show perfect black in the HDR areas where adaptivity kicked in, so that there is no risk of splatting massively energetic, and massively undersampled, pixels all over the image.
It would of course need proper in-the-field testing to be absolutely sure, but so goes the theory modelling and the synthetic testing.

F.e.: