Derivative Bug (B&W artifacts)

Vladimir says it's by design. You can apply a tiny offset to get rid of the artifacts.

It looks like a floating point error - the values of those B&W pixels will not change, even if you divide the result of the Derivative by some large number (the gray areas are affected, because here there the floating point values are still valid).

I just tested the offset trick - it seems to get around the problem at first sight.
It does so by luck though, since it avoid "stepping on the cracks".

Make it "step on the cracks" by enabling antialias and set the sample count to something high (I used 49 samples). Now the salt n pepper is back.

For the programmers:
Try uberzevs original filter and divide the immediate result of the Derivative with a constant (say 111). This should affect every sample going through the Divide right?

Now inspect the B&W samples before and after Divide: the value remains unaffected by the division, indicating that this is infact a bug related to invalid floating point result of the Derivative component.

Hmm.. I could be wrong (I'm on another computer now) and here it seems to "behave" as expected. Either I'm wrong or the error is CPU specific. I'll let you know when I get back to the other computer.

IMO this is a bug. Yes you can jack the anti-aliasing way up but that's not an elegant solution.

Hm, I thought maybe it was because it wasn't normalized, so I used the same technique I used for generating normal maps: square, add 1, square root, divide by that. The end result was the same. I had some similar artifacts without the normalize that normalizing seemed to fix. Maybe it has something to do with the lerp? I've not used that. Without the lerp, it's impossible to tell if the issue is there because the pixels are white or black around that point already.