Generates videos directly in pixel space without a pretrained autoencoder or cascaded refinement stages. It can jointly produce RGB frames and disparity maps, helping generated scenes follow target camera motion.