LongTake: Learning to Sustain Dynamics in Long-Horizon Video Generation
LongTake targets the failure mode where long autoregressive video rollouts go static or degrade.
The paper introduces a two-stage training pipeline using Long-Horizon Teacher Forcing on curated real long videos. That trains the model to predict later frames from long ground-truth prefixes, extending supervision beyond short clips. The authors report stronger initialization for distribution matching distillation without the usual intermediate few-step distillation stage. In their tests, LongTake improves 30-second rollouts at comparable aesthetic quality, while Hybrid DMD reaches the highest dynamic degree at both 30 and 60 seconds. HF Daily Papers' note
The paper introduces a two-stage training pipeline using Long-Horizon Teacher Forcing on curated real long videos. That trains the model to predict later frames from long ground-truth prefixes, extending supervision beyond short clips. The authors report stronger initialization for distribution matching distillation without the usual intermediate few-step distillation stage. In their tests, LongTake improves 30-second rollouts at comparable aesthetic quality, while Hybrid DMD reaches the highest dynamic degree at both 30 and 60 seconds. HF Daily Papers' note
score 5