WeVisCut: The Whole Editing Room in One Agent From Narration to Long-Form Video

WeVisCut. The Whole Editing Room in One Agent From Narration to Long-Form Video. By Chenyu Mu and ten others, from Xidian University and WeChat Vision at Tencent. Published on the twenty-eighth of October.

Turning a narration into a long-form video means making many decisions that depend on one another and have to stay coherent from shot to shot. Existing agentic editors often handle this with multi-stage workflows driven by general models. What that leaves unclear is the alternative: how a compact policy, adapted to this one task, would trade quality against cost.

WeVisCut is a long-form script-to-video system built to answer that, by connecting tool execution with policy learning. Rather than a chain of stages run by general models, it puts one trained policy inside an agent harness and lets that policy run the whole production.

The harness maintains a shared production state, so feedback from each tool informs the decisions and repairs that come after it.

Training that policy needs supervision, and finished videos do not expose the editing decisions that made them. So the authors built a four-stage data pipeline that combines editing inversion with verified replay. Together they turn the editing intent recovered from finished videos into supervision that can actually be executed. Applied to three thousand seven hundred and fifty finished videos, the pipeline yields over one hundred and thirty thousand training examples.

That supervision trains a single policy of nine billion parameters, which coordinates production and recovers from failures within the same harness.

To measure it, the authors introduce a benchmark of one hundred tasks with five metrics covering editing quality and delivery reliability.

Under a shared tool environment, WeVisCut substantially outperforms its base model. It achieves performance comparable to Claude Opus 5, while using nearly thirty-three times fewer generated tokens per successful output minute. It also outperforms task-matched native editors on the proposed metrics.

The question the paper set out with was the cost side of the trade. Its answer is one small policy, trained on decisions recovered from finished videos, reaching performance comparable to Claude Opus 5 while spending a small fraction of the tokens on each successful minute of output.