MIT & Toyota Research Institute · Text-to-Video

MilliVid

Long-video generation using a hierarchical autoencoder (coarse levels capture layout/semantics, fine levels add texture) and a coarse-to-fine rollout that generates many low-detail frames then progressively refines them, keeping a constant token count (~10x more context than naive autoregressive). On Minecraft it sustains geometric consistency and object permanence far longer than FramePack at equal compute.