Rotation is not movement
This is the distinction the whole vocabulary rests on. When a camera rotates, its position in space does not change, so the geometric relationship between near and far objects stays fixed. The frame sweeps across the world, but the world does not reorganise itself. When a camera travels, near objects shift against far ones, and that parallax is the visual proof of motion.
The practical consequence: a pan describes a space, while a travelling move takes you into it. Panning across a room shows you the room's contents in sequence. Walking the camera into the room makes you an occupant. Neither is better, but confusing them is why some sequences feel oddly flat despite constant camera activity.
Working with a pan
A pan does one of two jobs. It follows, keeping a moving subject in frame, in which case the audience barely notices it. Or it reveals, moving off one subject to show something else, in which case the audience notices it intensely and expects a payoff at the end of the move.
Two rules make either version work. Give the move a reason, because a pan that arrives at nothing reads as a mistake. And give it a beginning and an end that are both compositions in their own right, since the frames the shot rests on are what will be cut into and out of. A pan that starts mid-composition and ends mid-composition is almost impossible to edit.
Working with a tilt
A tilt trades in scale and hierarchy. Tilting up a building, a statue or a person makes them large by making the frame work to contain them. Tilting down usually reveals consequence: what is at the bottom of the stairs, what is at the person's feet.
The common error is tilting when you needed a crane. A tilt keeps the camera's height fixed, so a tilt up from a low position keeps the low-angle distortion for the whole move. If you want the camera to actually rise alongside the subject, that is a crane, and no amount of tilting substitutes for it.
Where zoom fits
Zoom belongs to the same family in practice, because most camera interfaces and most prompt vocabularies group pan and tilt with zoom in zoom out as the basic controls. Optically it is different: focal length changes, magnification changes, perspective does not. A zoom flattens as it tightens and feels like scrutiny; a push-in feels like approach. Choose deliberately, because audiences read the difference even when they cannot name it.
Prompting pan and tilt in AI video
These are the two most reliable camera instructions you can give a video model, which is why they are worth reaching for before anything more ambitious. Even so, four problems recur.
- Rotation gets confused with travel. The single most useful clause is a negation:
camera stays in one place, rotation only, tripod. Without it a requested pan frequently arrives as a slow lateral drift, and a requested tilt arrives as a crane. - Direction is ambiguous.
pan leftcan be read as the camera turning left or the scene sliding left, and models are inconsistent about which. Treat the first generation as a coin flip, then either reverse the wording or reverse the clip in the edit. - Amplitude is not controllable. There is no reliable way to ask for exactly 30 degrees. Speed and scale words are all you have:
slow, slight panversussweeping 180 degree pan. Expect to re-roll for magnitude. - Combined moves collapse. Asking for a pan and a tilt at once usually gives a vague diagonal drift instead of two legible rotations. Write them as a sequence (
pans right, then tilts up) and accept that the model may only do the first.
A note on zoom: prompts for slow zoom in almost always produce a dolly-style push, because true focal-length compression is rare in captioned training data. If the flattened, non-parallax look matters, generate a high-resolution still and animate a crop over it in post. That gives you a genuine optical zoom, exact timing, and no drifting geometry, which is a better result than the fifth attempt at the prompt.