LongTake: Learning to Sustain Dynamics in Long-Horizon Video Generation

LongTake. Learning to Sustain Dynamics in Long-Horizon Video Generation. By Byoungwoo Park and three others, from KAIST and Georgia Tech. Published on the twenty-ninth of September. Its code repository is a placeholder: a readme, and nothing to run yet.

World models, game simulators and long-take video creation all need scenes that evolve coherently and dynamics that are sustained over long durations. Autoregressive video diffusion is a natural framework for that kind of long-horizon generation, yet extended rollouts often turn near-static, or lose visual quality. The authors hypothesize a cause: short-video supervision gives limited guidance on how ongoing scene dynamics develop over longer durations. Their central finding follows from addressing it. A training stage that supervises the model on long real videos strengthens the direct initialization of distribution matching distillation, or DMD, under student self-rollout, and does so without the intermediate few-step distillation stage used in standard pipelines.

That stage is Long-Horizon Teacher Forcing, the core of LongTake, a two-stage training pipeline built on curated real long videos. It trains the autoregressive model to predict later frames conditioned on long ground-truth video prefixes, which extends direct supervision beyond the short training horizon. It is designed to help the model sustain dynamics and preserve visual quality. Under the same five-second DMD training setup, initializing this way yields substantially higher dynamic degree than initializing from short-horizon teacher forcing, measured on thirty-second rollouts at comparable aesthetic quality, and it surpasses the evaluated baselines on both measures.

A further step, Hybrid DMD, reuses this teacher to extend supervision to later frames of the self-rollout, while retaining bidirectional joint supervision over the initial window. On long-horizon self-rollouts, LongTake lies on the Pareto front of dynamic degree and aesthetic quality, and Hybrid DMD attains the highest dynamic degree among the evaluated methods, at both thirty and sixty seconds.

The finding has stated edges. The comparison between initializations is made under one fixed setup, five-second DMD training, and the rollouts measured are thirty and sixty seconds long. There are two measures, dynamic degree and aesthetic quality, and the claim is stated against both together: more motion at comparable quality, not motion bought by giving quality up. The cause it starts from remains a hypothesis, that short-video supervision is what leaves long rollouts without guidance. What the evidence supports directly is the training stage built on that hypothesis, and what it does for the distillation that follows.

Long videos go still because the model was never shown how long videos keep moving. Showing it, with real long footage and long prefixes, does two jobs at once. It sustains the dynamics, and it gives distillation a starting point strong enough that one stage of the standard pipeline can be skipped.