You can see a video of their technique here:
http://www.youtube.com/watch?feature=pl … qMovlLlr_0
The main issues I see for use in emulation are: 1.) it requires analysis and comparison of previous and current frames, which necessarily introduces “lag” and 2.) distortions are still happening, but they’re not happening uniformly, which is good for subjective measures of quality but not for game logic.
Here’s an example of #2: let’s imagine you’re playing Super Mario Bros and there’s a big pit you have to jump over. The algo sees an area with nothing important (i.e., just blank sky, not even any ground) and starts carving it out as you approach it (meanwhile, it’s preserving the goombas you’re jumping on, etc., so they all look nice). This makes the pit seem pretty small, so you don’t bother building up speed because a walking jump can cover it. Then, when you go to jump over it, it stops carving out those segments because they are now considered important (the main character is over/in them and they’re the focus of motion), and it seems like your jump slows down and/or the pit expands, causing you to fall into it.
Thus, even though simple squishing/stretching of the AR looks subjectively bad, it makes more sense in this case because it’s consistent, which preserves the game logic.