What actually happens to your prompt

1. 01

   You write a brief and approve a storyline

   One or two sentences. A model turns it into a hook, a handful of beats and a payoff, each with the line to be spoken and the shot to be generated. You edit any of it. This step is free, every time, for the first storyline of a video — it is how you find out whether there is a video here before spending anything on one.

2. 02

   The narration is recorded first, and it sets the timing

   This is backwards from how most pipelines run, and deliberately so. The model estimates how long each beat will take to say; the narrator then tells us how long it actually takes, and the two are never the same. So the voice is recorded first and every scene is re-timed to fit what was really said. Nothing is cut off mid-sentence, and the expensive visual stage never runs against a timeline that is about to change.

3. 03

   Images, then video, at the corrected lengths

   An image for every beat. On the standard and cinematic tiers each beat also becomes generated video — and because no model will generate more than about fifteen seconds at a time, longer beats are cut into shots at points the storyline model chose, never mid-sentence. Each shot is seeded from the final frame of the one before it, so the joins are continuous rather than a cut you can see.

4. 04

   A score cut to the finished length

   Composed last, at the exact millisecond length of the edit, with sections that change where the story changes. It is instrumental: a vocal line competes with your narrator.

5. 05

   The edit, applied from a preset

   Cuts, caption style, transitions and how far the music ducks under the voice all come from the preset you picked. Captions are built from the words the narrator actually said, so they land on the syllable. Switch preset and the whole video re-cuts — for free, because nothing is regenerated.

When something goes wrong

A failure costs you nothing

Credits are held when a run starts and only charged when there is a finished file. If a scene fails, the scenes that succeeded are kept — you retry, and only the missing ones run again. If the failure is something you can fix, such as a model declining a prompt, the app says so and says which scene, rather than showing you an error code. If you give up on a run, the held credits come back.