Seems like the main improvement is in upscaling speed rather than outright quality (there are existing techniques that achieve similar levels of quality, but do so much more slowly).

I'm curious if this has any video applications. They seem to be mostly focusing on still images so far. Not sure if that's because the technique is more difficult to apply to video, doesn't work as well, is too computationally intensive. Or if just no one has gotten to it yet.

But given that 4K resolution screens are being produced much more quickly than 4K content (or than bandwidth is increasing for that matter), seems like many of those 4K screens will be used to watch HD content. So it would be nice to have better video upscaling techniques.