Millivid

MilliVid is a research project on long-range consistency in video generation. It addresses the problem that even a few dozen frames require impractically long transformer sequences. The approach pre-trains an autoencoder that compresses each frame into a hierarchy of tokens, from the usual latent resolution down to a handful of tokens per frame. Coarse levels capture scene layout and semantics, while finer levels add texture and high-frequency detail.

A video diffusion model is then trained to generate these tokens by coarse-to-fine rollout. It first generates many coarse frames, then refines them level by level, conditioning on past frames and on coarse future frames. The number of tokens seen per step stays constant, and with three hierarchy levels the project page says this gives roughly 10x the context of naive rollout at the same sequence length budget. The authors contrast hierarchical latents with resolution cascades and report better reconstructions at small token budgets.

Evaluation uses a custom dataset of long Minecraft videos, with action-conditioned rollouts (forward/left/right) from up to 256 context frames. The authors report more consistent and higher-quality rollouts than FramePack and standard autoregressive rollout, and stable rollouts where the baselines degrade from exposure bias. It is a project page for a research method; the provided material shows no released code, weights or product, and the validation is limited to Minecraft.