Two failure modes, one loop. A feedback signal where there was silence, and a regression net where there was none.
Today's video models are black boxes. Bad physics? Temporal jitter? Wrong subject? You get a video — not a reason. So you reroll, reword, and hope.
Change a system prompt, swap a template, upgrade a model — and a dozen workflows regress without warning. There's no git blame for video quality.
A single prompt becomes up to three takes, each informed by the score of the last.
Interactive mode gives you the steering wheel — see every score, add a remark, decide whether to heal. Autonomous mode runs the whole loop and streams progress live.
Define a suite of benchmark prompts and thresholds. Run them before every template change. Retake grades each one against a saved baseline and gives you a clean pass/fail.
Paste twenty prompts, walk away. Retake runs two concurrently, streams progress into a live grid, and hands you a CSV when done.
We didn't train a model — we orchestrated two of the best ones. Generation and evaluation in a choreographed handoff.
State-of-the-art text-to-video. 720p, 5-second clips at 24fps. Retake handles prompt templating, parameter tuning, and retries.
A vision-language model that watches your video and scores it across prompt adherence, temporal consistency, and physical logic.