Nvidia with Chrome can already do it in real-time, if you trained models specifically for it - it would be even better.
