One model coordinates long-form video production
WeVisCut: The Whole Editing Room in One Agent From Narration to Long-Form Video · arXiv:2610.38352
In plain words
Making a long video from a script means choosing shots, visuals, narration, and timing that fit together. A change to one-shot can affect the rest, and the finished video does not show the decisions or tool failures behind it.
WeVisCut is a system that uses one model to plan, create, and check a video, while keeping track of production steps. On 100 fixed-script tasks, its reported quality summary was higher than the same model’s base version in the same tool setup.
The research
When one-shot changes, timing and other editing decisions may need revision too. Can a video-making system learn to manage those linked decisions when a finished video does not reveal the actions, feedback, and failures behind it?
WeVisCut combines a stateful production runtime, replay-derived training examples, and staged training to coordinate long-form video editing with one model.
The paper asks how a system can keep a long-video coherent when shot choices, visuals, narration, and timing affect one another, and when finished videos hide the production decisions behind them.
WeVisCut connects a runtime that records tool feedback with training examples made by replaying recovered shot structure, then trains one model across production roles.
On 100 fixed-script tasks, its five-part quality summary averages 90.5, versus 74.5 for its base model in the same tool setup; all five component averages are higher. This supports an improvement from the combined training procedure under that evaluation.
A separate comparison finds a similar overall average to Claude Opus 5, while native-workflow results extend the comparison to two other systems.
The results cover 100 fixed-script tasks and the tested tools and production setup. The Overall Score summarizes five measures but is not a separately validated measure of perceived video quality; the study also does not isolate the effect of each training stage.
What the evidence shows
- Interdependent editing decisions
WeVisCut is a long-form script-to-video system that uses one 9B model across planning, content generation, and quality inspection. The problem it addresses is that shot selection, visuals, narration, and timing depend on one another: a change to one-shot can require revisions elsewhere.
A finished video shows the output but not the tool calls, feedback, or failures that could teach a system how to manage those decisions.
- A runtime tracks production
The system’s runtime keeps a shared record of plans and production artifacts. It checks tool inputs and prerequisites, records validated results, and reports blockers when an action cannot proceed.
The model can use that feedback to inspect outputs, request bounded repairs, and replan while retaining unaffected artifacts. This makes proposed actions and completed production work distinguishable.
- Replaying intended edits
To create training material from finished videos, the authors recover intended shot structure and replay it through production tools. Inspections and bounded repairs produce recorded examples that include tool feedback and failures.
The resulting storyboards capture intended video structure, not the original editors’ tool actions. The pipeline turns 3,750 projects into 134,771 examples for training the shared model.
- A five-part quality test
The evaluation uses 100 fixed-script tasks from projects excluded from training, with five independent runs per system. Five measures cover script fidelity, visual support for narration, timeline realization, late-stage shot health and asset variety, and successful rendering.
The equal-weight average of the five component score averages is called the Overall Score; it is a descriptive summary, not a separate measure of perceived video quality.
- Results across two comparisons
In the shared tool setup, WeVisCut’s Overall Score is 90.5, compared with 74.5 for its Qwen3.5-9B base, and its reported averages rise on all five measures. Its 90.5 average is also close to Claude Opus 5’s 89.9, though that comparison changes both model identity and task-specific training.
In a separate comparison of complete native workflows, WeVisCut scores higher on average than Crayotter and VideoAgent on the same tasks.
Deep
- Original paper authors
- Chenyu Mu, Wenming Weng, Qinrui Wang, Xingsong Ye, Hao Yu
- Submitted to arXiv
- License
- CC BY 4.0