MovieRoll: A Systematic Long-Horizon Framework for Long Video Generation

Written by

in

MovieRoll combines streaming fine-tuning, long-horizon rollout distillation, and long-duration inference to generate streaming videos longer than the training duration. Its hierarchical KV-cache compression method, HieraKV, reduces cache memory by up to 78%. After training on videos up to 15 seconds long, MovieRoll-Base remains competitive at 30 seconds, while MovieRoll-Distilled achieves the highest overall scores in single-event and multi-event evaluations of up to 60 seconds.

Source:Hugging Face 社区论文精选 · Read Original

Comments

Leave a Reply

Your email address will not be published. Required fields are marked *