WeVisCut: A Single Agent for End-to-End Narration-to-Long-Form Video Editing
ORIGINAL / WeVisCut: The Whole Editing Room in One Agent From Narration to Long-Form Video
Existing agentic editors often rely on multi-stage workflows driven by general models, leaving the quality-cost trade-off of compact task-adapted policies unclear. WeVisCut connects tool execution with policy learning, using a single 9B policy for long-form video production and achieving performance comparable to Claude Opus 5 with far fewer generated tokens in a shared tool environment.
01 ABSTRACT
WeVisCut is a long-form script-to-video system whose agent harness maintains shared production state so tool feedback informs subsequent decisions and repairs. A four-stage data pipeline combining editing inversion with verified replay recovers editing intent from 3,750 finished videos, yielding over 130K training examples used to train a single 9B policy. The authors report that this policy substantially outperforms its base model on a 100-task benchmark and achieves performance comparable to Claude Opus 5 while using nearly 33x fewer generated tokens per successful output minute.
02 KEY FINDINGS
- Introduces an agent harness with shared production state that lets editing tool feedback drive subsequent decisions and failure repairs.
- A four-stage data pipeline inverts editing intent from finished videos and, via verified replay, produces over 130K executable supervision examples.
- A single 9B policy substantially outperforms its base model in a shared tool environment, approaching Claude Opus 5 with nearly 33x better token efficiency.
AI GENERATED SUMMARY / DISCOVERED BY HUGGING FACE DAILY PAPERS