Seems like the key to consistency will be generating multiple images in a sequence simultaneously, rather than generating frames independently. Are models being developed along these lines?