WeVisCut trains a 9B agent for narration-to-video editing
Original title: WeVisCut: The Whole Editing Room in One Agent From Narration to Long-Form Video
Trained on more than 130,000 examples, the 9B policy matches Claude Opus 5 while using nearly 33 times fewer tokens.
WeVisCut is a long-form script-to-video system that links tool execution with a learned editing policy. A shared production state lets tool feedback guide later decisions and repairs, and a four-stage pipeline recovers editing intent from finished videos through inversion and verified replay. Applied to 3,750 videos, the pipeline produces over 130,000 training examples used to train one 9B policy. On a 100-task benchmark with five metrics, the authors say it beats its base model and task-matched native editors under a shared tool setup.
Unconfirmed: Performance and token-efficiency comparisons are author-reported on their own benchmark.
- The data pipeline was applied to 3,750 finished videos.
- Supervision exceeds 130,000 training examples.
- The benchmark has 100 tasks and five quality and reliability metrics.
- Token use is reported as nearly 33 times lower per successful output minute.
From the original article
1 min read at huggingface.co
Producing long-form videos from narration requires interdependent decisions that remain coherent across shots. Existing agentic editors often use multi-stage workflows driven by general models, leaving the quality-cost trade-off of compact, task-adapted policies unclear. We address this gap with WeVisCut, a long-form script-to-video system that connects tool execution with policy learning.
By Chenyu Mu, Wenming Weng, Qinrui Wang · 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.