Video Studio is the same idea as Procedural Sound Studio in a different medium: the score is the program. The model does not produce pixels or samples. It writes an editable score and the code that renders it, and the studio proves the output came from the score.
How it works
- You describe a video. A coding agent (Codex, Claude Code or a model on OpenRouter) writes
score.json, a timeline where every visible element is an event, andscene.html, which draws any frame from the score and the time alone. - The studio renders it frame by frame in Chromium and encodes the video with FFmpeg, along with a contact sheet.
- An output check validates the manifest, the score contract, the video probe and motion metrics. Its blank test strips every event from the score and fails the delivery if anything is still drawn.
- You edit by talking. Select events or drag a time range on the timeline to scope the next prompt. Every follow-up creates a new version and keeps the old ones.
There are two renderer modes: free JS, where the model writes its own renderer, and kit, where it writes only the score and SVG assets and the studio generates the renderer from a fixed motion kit.
Why the checks matter
An LLM that is asked for a video can quietly paint the result instead of building it from parts you can edit. The blank test closes that door: if removing every event leaves anything on screen, the output is rejected. The same contract makes each request a gradable task, with a pass/fail check that needs no human.
Facts
| Item | Detail |
|---|---|
| Stack | Electron and Node 22, Chromium for frames, FFmpeg for encoding |
| Providers | Codex CLI, Claude Code, or OpenRouter, using the user's own login or key |
| History | 74 commits across 7 branches; never public |
| Tests | Unit, integration and UI smoke suites that make no model calls |
| Availability | Private codebase, being prepared for non-exclusive licensing |
What you can buy
- The codebase, licensed non-exclusively once it is listed.
- Environments in the same style: generation tasks where the model must produce an executable artifact, with automatic checks such as the blank test as graders.
FAQ
Does the LLM generate the video pixels?
No. It writes a score of timed events and a renderer that draws frames from that score. The studio renders and encodes the frames, and a blank test rejects any output that does not come from the score.
How is it related to Procedural Sound Studio?
They share one idea: the score is the program. Sound Studio has models write SVG scores and the synthesis that plays them; Video Studio has them write timeline scores and the renderer that draws them.
Can I license Video Studio?
It is being prepared for non-exclusive licensing. Email me if you want to be told when it is listed, or to see it running.