Frame generation can only guess the next frame based on previous ones and motion vectors so when you do something it can’t predict (press a button) it still has to draw that normally. As I understand it because of this the input latency is going to be the same as if frame generation is off it will just look smoother.