Description
WeVisCut functions as a comprehensive AI agent for long-form script-to-video production, integrating tool execution with advanced policy learning. It is trained on a vast dataset of 134,771 examples derived from 3,750 completed video projects, utilizing a single 9B policy. This robust training enables the system to learn script-to-video production through executable supervision, incorporating editing inversion, verified replay, and a 100-task benchmark.
Read More
The core functionality of WeVisCut revolves around an iterative observe-think-act cycle. After the initial video cut is generated, a director can interact with the agent via chat, requesting specific modifications such as swapping a shot's source, changing music, or restyling captions. The policy then intelligently re-plans locally and re-delivers the revised content, ensuring that already successful elements are preserved and not re-shot.
Key to its operation are 'Editing Inversion' and 'Verified Replay.' The system recovers editing intent from finished narration-driven videos and then replays this intent using production tools. This process involves a four-stage pipeline that includes observation, bounded repair, and trajectory materialization, generating executable supervision for the AI. WeVisCut maintains a shared production state across all tool calls, allowing runtime feedback to inform subsequent decisions, inspections, bounded repairs, and replanning, moving beyond a fixed open-loop workflow.
The WeVisCut framework encompasses data construction, training, and inference, all supported by closed-loop tool execution. It intelligently selects appropriate tools based on the current production state, coordinating planning, execution, inspection, and delivery. The agent is capable of creating scripts and shot plans, retrieving visual and audio assets, and composing narration, decorations, and timeline edits. It also inspects returned artifacts against quality dimensions, identifies failures, and routes targeted repairs. Validated tool returns update the canonical state and artifacts, ensuring that each subsequent action leverages the latest run status. Local fixes are prioritized to preserve successful work, with inspection evidence feeding back into the policy for improved decision-making.
Performance benchmarks demonstrate WeVisCut's effectiveness, showing superior results compared to general-purpose baselines and reference models like Qwen and Gemini. It even performs on par with Claude Opus 5 while using significantly fewer tokens and calls per output minute. Native system comparisons highlight WeVisCut's advantage in speed and cost-efficiency, running 4.1 to 5.0 times faster and at 6.4 to 10.8 times lower cost than competitors like Crayotter and VideoAgent. Human evaluations further confirm its quality, with WeVisCut achieving a higher overall mean opinion score across various dimensions like theme, richness, narration, smoothness, and visual impact.
- CategoryAI Video Editors
- DeploymentCloud
- PlatformsWeb
WeVisCut's Core Features
Long-form script-to-video production
Single 9B policy trained on 134,771 examples
Iterative observe-think-act editing cycle
Director-chat for post-cut modifications
Editing Inversion for intent recovery
Verified Replay with production tools
Four-stage pipeline for executable supervision
Stateful tool execution with shared production state
Runtime feedback for dynamic decision-making
Planning, retrieval, composition, inspection, and repair capabilities
Automated script and shot plan creation
Visual and audio asset retrieval
Narration, decoration, and timeline edit composition
Quality inspection and targeted repair
Local fixes to preserve successful work
How to use WeVisCut?
Provide Script: Input a long-form script to initiate the video production process.
Generate Initial Cut: The WeVisCut agent will produce a first draft of the video based on the script.
Review and Chat: Watch the agent's cut and provide feedback or requests for changes via a chat interface.
Request Revisions: Specify modifications such as shot sources, music, or caption styles.
Receive Re-plan: The policy locally re-plans and re-delivers the video with the requested changes.
Iterate and Refine: Continue the review and revision cycle until the desired final film is achieved.
WeVisCut's Use Cases
- Long-Form Video Creation
- Narration-Driven Content
- Educational Video Production
- Documentary Film Editing
- Content Iteration and Revision
- Script-to-Film Automation
