Cut to the Music, Moving Like the Film You Admire

The AI video category has settled into a recognizable shape: describe what you want, get a video assembled from stock, templates, or generated frames. Tools like Motion and Mosaic sit in that space and are good at it. Ghostmotion is built on two different decisions, and both of them are about motion rather than about content.

1. The music writes the edit

In most video tools, the soundtrack is the last step. You assemble the cut, then lay audio underneath it, and the two are related only by total length. It is the reason a lot of AI-generated video feels subtly wrong even when every individual shot is fine: the picture and the sound are not agreeing about where the beats are.

We invert that order. You pick the track first, or let the agent pick one. The waveform is then analysed for tempo, beat positions, and phrasing, and that grid becomes the edit's skeleton. Cuts land on beats. Transitions ride the phrasing rather than crossing it. Camera pushes are timed to hit the drop. Even the animation inside a scene, when an element enters and how fast it settles, is derived from the same timing map.

The practical consequence is one you can test: change the song and the same footage re-cuts itself to the new rhythm. The cut is not decoration applied over the audio; it is a function of it. Cutting on the beat is the single most under-rated variable in product video, and it is why the same shots can read as directed or as amateur depending on nothing but timing.

2. Automotion: motion learned from film, not picked from a library

The second decision is about where motion comes from. Template tools ship a fixed library of animations, which is why films made with the same tool tend to move the same way. You can usually identify the tool from the motion alone.

Automotion is our motion-learning model, and it works from reference instead. Paste a link to any launch video you admire, from YouTube, from X, from a portfolio reel, and it measures the animation in the footage: durations, easing curves, spring constants, stagger, rotation. Every value traces back to something measured from pixels, and when the footage is too short or too noisy to be sure, it reports the move as unknown with the reason rather than inventing a number.

Measured moves are then checked against the studio's motion library. Known moves are matched, near-misses widen an existing preset's range, and moves the library has never seen are distilled into new named presets. Because a preset can cite the right numbers and still move wrong, every learned move is rendered and re-measured with the same pipeline that studied the source, until the two match frame for frame.

The result is that your film inherits the motion grammar of the work you pointed it at, as measured values rather than as adjectives. It moves like the reference; it does not look like it. Every frame is still your own product's screens in your own brand.

What both decisions have in common

Neither is about generating more content. They are both about the part of video that people notice without being able to name: whether the thing on screen moves with intention. Timing derived from the music, and motion derived from film you chose, are two ways of answering the same question.

There is a third thing worth stating plainly, because it constrains everything above. We film your real product. The agent explores the actual interface at your URL and rebuilds those screens as live animated UI. That is a narrower promise than prompt-to-anything: we will not make you a brand documentary, a talking-head piece, or a conceptual explainer, and if it does not happen on a screen we do not film it. What we will do is make the product you actually shipped look like it was directed.