The commercial use cases for generative art is not to make images that experts cannot discern as fake. It would be very expensive to annotate training images to have physically correct reflection images and the value would be negligible. Realistically, if you want to produce something that is impossible to prove fake, you would have a vastly easier time doing such edits manually. We are very, very, very far from being able to push button churn out undiscernable fakes. Even making generically good outputs for art is a careful process with lots of iteration.

Indeed, but useful nonetheless. Solving it may be a challenge for a while, and deep fakes generated before a solution becomes broadly available will remain detectable with this technique.

Or just use conventional algorithm, since the fix is about formal physics. Although it will not be a true 100% fix, it could be good enough to make this test rather useless, because even now:

> "There are false positives and false negatives..."