calculate the instance cutouts first, batch them and then process in one single inference call. Should make the instance-phase way quicker