How the generator works

You do not need any of this to use the generator. It is here because knowing what happens in the twenty minutes after you press the button makes it easier to write a direction that lands, to pick a reference that works, and to understand why a scene is one bar long in the chorus and fifteen seconds long in the intro. To see what comes out at the other end, the AI music video generator page has the showcase films with their inputs.

#1. It listens to the song

Before a single word of story is written, the audio is analysed: the tempo, the beats and bars, where the sections begin and end, where the energy peaks, which instruments enter where, and which bars carry a vocal. The genre is recognised from the sound too. If you allowed the lyrics, the vocals are transcribed. Everything downstream, from the length of a scene to the pace of the cut, is read from this, not from your description of the genre.

#2. A director writes the treatment and the shots

A director model gets your direction, your references, the lyrics if you allowed them, and the analysis. It writes a treatment for the whole song, then a shot list addressed in bars, not seconds: every scene knows which part of the song it belongs to. The energy of the track shapes the rhythm of the list, so the loudest passages get short scenes that can be cut quickly, and a quiet intro or a breakdown can be one long take. The director also follows a set of production rules learned from many films, which is why frames come out lit the way the story says, empty when they should be empty, and without the things motion models get wrong. With Lip sync on, it stages performance scenes only on the bars where the song is sung. If you chose Research the world, a research pass runs first and adds real facts and photos of your subject to the brief.

#3. Every scene becomes a still

Each scene is rendered as a keyframe by an image model conditioned on your actual references, so a palette, a location or a drawing style carries into the frame instead of being described and hoped for. Scenes where your main character is visible get extra care so the face holds from take to take, and characters the director invents (a bandmate, a stranger on the train) get one look each, so the same person turns up in every scene they are in. A retake asks for a genuinely new frame; an edit changes the one you have. With Auto QC on, every frame is reviewed against the references and the scene direction, and the ones that fail are retaken once and reviewed again.

#4. Motion is shot per scene

After you approve the storyboard, each keyframe is animated by a motion model with a prompt written for that scene: the camera move, what happens, where it arrives. Scenes play at their natural speed rather than being stretched to fill their place in the edit. On a handful of musical arrivals per film the director adds a designed camera move (a push-in, an orbit, a fly-through that carries the camera across several scenes), never over a singing face. Scenes where the singer performs are shot from your vocal so the mouth follows it. Many scenes are shot in parallel, which is why a hundred-scene film still comes back in about twenty minutes. The quality option you pick (480p, 768p or 1080p, Fast or Cinematic) is the engine and resolution used here, priced per generated second. Every take is checked before it is used; a scene that fails is reported on its card so you can retake it.

#5. The edit follows the beat

The cut is built from the analysis, not by eye. Every cut lands on the beat, and the pace follows the energy of the song: quick in the loudest sections, held in the quiet ones. The genre sets the grammar: an ambient track drifts, a ballad breathes, a groove cuts on the bar, trap and bass music hit harder, and only the most aggressive music gets the full wall of flashes. Hits get accents, long takes get something happening in the middle, and in the biggest passages the edit follows the instrument that carries them, so a drop feels played rather than timed. It does not drift either: the last cut of a five-minute song sits as tightly as the first.

#6. Finish and render

A restrained visual layer, shown on the strongest bars rather than all the time, and a grade chosen for this film from its own frames and the song finish the picture (this is where the Black & White quick style is applied). Your original audio is joined to it, and the film is packaged for streaming at several qualities, with the original available for download.

#7. It stays open

A delivered film is not a dead end. Every scene keeps its takes, the edit is kept as a set of decisions, not only as a finished video, and the chat can change one layer while the rest stays as it was: a faster cut in the chorus re-cuts the same clips, a new look re-grades the same edit, a retaken scene drops into the same timeline. That is why a change costs the new parts and a render, not a new film.

#What this means for you

• Write what the camera sees, not what the software should do. The director already knows the tools.
• References beat adjectives. A real photo of the place, a scan of the artwork, a cover: the image stage works from them directly.
• Fix the still, not the motion. A wrong frame is 4 credits to retake; a wrong clip costs motion credits. Retake and edit before you press Make the film.
• Expect short scenes in the loud parts. That is the edit following your song. If you want a long, slow shot in the chorus, say so in the direction, or ask the chat for calmer cutting afterwards.