I believe the claim is that the NN is trained to reconstruct pixels, not images. As in so many areas, the diffraction limit is probabalistic so combining information from multiple overlapping samples and NNs trained on known diffracted -> accurate pairs may well recover information.

You’re right that it might fail on noise with resolution fine enough to break assumptions from the NN training set. But that’s not a super common application for cameras, and traditional cameras have their own limitations.

Not saying we shouldn’t be skeptical, just that there is a plausible mechanism here.

My concern would be that if it can't produce accurate results on a random noise test, then how do we trust that it actually produces accurate results (as opposed to merely plausible results) on normal images?

Multilevel fractal noise specifically would give an indication of how fine you can go.

we've had very good chromatic aberration correction since I got a degree in imaging technology and that was over 20 years ago so I'd imagine it's not particularly difficult for name your flavour of ML.