MovieRoll framework extends streaming video beyond its training length
Original title: MovieRoll: A Systematic Framework for Rolling Out Long-Form Videos
On HeliosBench, the distilled model posts the highest compared scores for clips up to 60 seconds after training on videos of at most 15 seconds.
MovieRoll is a base-to-distilled framework for streaming video that combines window-wise teacher forcing, hierarchical KV-cache compression, and cumulative long tuning with distribution matching distillation. Hierarchical compression is reported to cut cache memory by up to 78% while improving quality. KV recomputation and guidance annealing are used to generate past the training horizon. The paper says MovieRoll-Base stays competitive at 30 seconds and MovieRoll-Distilled leads the compared methods through 60 seconds.
Unconfirmed: Quality and memory figures are the authors' HeliosBench results, not an independent evaluation.
- Training videos are at most 15 seconds; evaluation runs to 30 and 60 seconds.
- HieraKV keeps full history in shallow layers and sink-recent history deeper.
- CuLT-DMD averages local gradients while holding parameters fixed during each rollout.
- FBCA attention and PS-VAE decoding support training and streaming output.
- AI models
- MovieRoll-BaseMovieRoll-Distilled
From the original article
1 min read at huggingface.coStreaming video generation must sustain quality beyond the training horizon while handling self-generated history and growing cache costs. To address these challenges, we present MovieRoll, a base-to-distilled framework that connects streaming fine-tuning, long-rollout distillation, and long-horizon inference.
By Yunyang Ge, Weijie Kong, Ryan Xu · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.