MovieRoll: A Systematic Framework for Rolling Out Long-Form Videos
MovieRoll’s distilled model posts the highest HeliosBench total scores among the compared methods in evaluations up to 60 seconds.
The framework combines streaming fine-tuning, long-rollout distillation and inference techniques to sustain video quality over longer sequences. Its hierarchical cache compression cuts cache memory by up to 78%, while also improving generation quality, the authors report. Distillation combines supervision from multiple points in a generated sequence while retaining only one local computation graph. The base model remains competitive at 30 seconds despite training on videos no longer than 15 seconds.
HF Daily Papers' note
The framework combines streaming fine-tuning, long-rollout distillation and inference techniques to sustain video quality over longer sequences. Its hierarchical cache compression cuts cache memory by up to 78%, while also improving generation quality, the authors report. Distillation combines supervision from multiple points in a generated sequence while retaining only one local computation graph. The base model remains competitive at 30 seconds despite training on videos no longer than 15 seconds.
HF Daily Papers' note
score 5