Does it "look" into a single frame or does it take previous steps into account as well?