I've actually been wondering if the next step in all this ML upscaling is for all the successful trained models to be put in a database somewhere with notes about the aesthetic they're good at upscaling and all games it's been successfully used for, with users able to download pre-trained models ready to go. So if you wanted to upscale an arbitrary game, you look for the one with the closest aesthetic notes on the database, download it, and give it a go. If it succeeds, you make a note of it on the database.
Of course, other stuff would be needed too, like algorithms for auto-processing the masking and stuff.
This would hopefully eventually turn into a comprehensive collection of pre-trained upscaling setups, with attendant other software, so that if you wanted to upscale a game and no-one's done that one you can skip the month/s of training a neural network from scratch and just wait for the day or two of running the various processes. Add to that the appropriate injectors for various emulators and pc game engines, and you could upscale virtually any 3d game that functions on a pc, whether natively or by emulator, within two or three days on any decent pc!
Dunno if that will ever come to be, but it's a nice thought.