Why the model understands film words
Because its training data contained the word and the footage that went with it. In video descriptions, scripts and trade material, "dolly in" or "crane shot" always sat next to the same kind of movement, and what the model learned is that pairing.
Two things follow. One is that the established, common terms work well — precisely the ones that turned up often. The other is that rare or insider jargon does not: one vendor explicitly recommends rewriting a less common term as "the term plus a plain description", because on its own it is not enough.
The six movements that land reliably
- Push in / pull out. The camera moves closer to the subject or away from it. The most reliable move there is; it almost always hits.
- Pan. The camera stays put and turns horizontally. Works well if you say which way.
- Tilt. The same thing vertically. A little weaker than a pan.
- Orbit. The camera circles the subject. Spectacular, and surprisingly reliable — but eight seconds is not much for a full turn.
- Tracking / follow. The camera follows a moving subject. Good when the subject's own movement is described too.
- Handheld. A slight hand-held wobble. It gives a documentary feel and is very easy to call up.
Not movement, but where the camera stands
Movement is only half of it. The other half is where you are looking from — and that is the half most people leave out.
- Height. "Eye level", "low angle", "high angle", "bird's eye". The single word with the biggest effect on the image.
- Shot size. "Wide shot", "medium shot", "close-up", "extreme close-up". The model handles these settled terms well.
- Lens character. "Wide angle", "long lens", "macro". What comes out is not an optical simulation but the look those images tend to have.
- Depth of field. "Shallow depth of field, blurred background" — one of the strongest quality upgrades available, because what separates a film image from a phone clip is the background.
The vendor caveat worth reading
One large provider's documentation carries the same sentence twice: certain advanced camera angles and lenses are not officially supported, and both the result and the reliability may vary. That is not an excuse, it is useful information.
Translated into practice: a camera word is a suggestion, not an order. With the settled, common terms the hit rate is high; with the unusual ones the model gives you whatever most resembles the description. Which is why it pays to ask for one camera move per clip. It will not execute two or three in sequence — what you get is an average that resembles neither.
The static camera that is not static by itself
This is the most common annoyance, and it has a simple explanation. In the training data the overwhelming majority of videos have a moving camera, so the default is movement. If you want a locked-off, still-looking shot, you have to say so separately — one vendor specifically suggests spending a whole sentence on it.
What works: "the camera stays completely still on a tripod, the framing never changes". What does not: "static shot" on its own, because that is too little for the model to override its default. The same holds in reverse: if you want a particular move, state its direction and speed as well.
Three routes to tighter camera control
Where words run out, there is a level above them.
- Camera tokens you pick from a menu. One provider offers the movements as menu items (pan, orbit, crane, push in, pull out). More predictable than free text, at the price of a shorter list.
- A shot-level script. At another one you lay the scene out shot by shot, in fields: duration, camera position, perspective, camera movement. This is the closest thing to what you would say to a cinematographer.
- A reference video. The strongest method: you do not describe the camera move, you perform it. One provider offers a grey, clay-like test render for exactly this: you move the camera on that, and the model takes the movement over.
A template you can copy
[CAMERA POSITION] + [SHOT SIZE] + [WHAT IS IN FRAME] +
[ONE camera move, with direction and speed] + [LIGHT] + [LENS CHARACTER]
Example:
Low angle, medium shot of a ceramic cup on a workshop bench.
The camera slowly pushes in over five seconds, ending close on the rim.
Soft window light from the left, shallow depth of field, the background
falls out of focus.
For a static shot, in a sentence of its own:
The camera stays completely still on a tripod, the framing never changes.
The template is in English for a reason. Video models are trained on English captions, and the vocabulary of camera work — the angle, the shot size, the name of the move — is most precise in English. If you work in another language, keep the prompt English and the shot plan in your own. How much that choice of language costs is what models and smaller languages actually measures.
Common misconceptions
- "More camera moves make it more dynamic." One clip, one move. Two of them average out.
- "Giving a focal length produces an optically accurate image." It does not. It gives the look, not the physics — the vendors say as much themselves.
- "Static is the default." It is the other way round: movement is, and static has to be asked for separately.
- "If it fails, I need a better prompt." Sometimes, but more often you need a different mechanism: camera tokens, shot fields or a reference video.
What to take away
One clip, one camera move, with direction and speed — that is half the job. The other half is the camera position and the lens character, because that is what gives the filmic feel. If the word is not enough, move up a level: tokens, shot fields or a reference video. For the full picture of what these models can and cannot do, How AI Video Works is the starting point, and on repairing a shot that came out wrong, Editing AI Video is the post to read next.