AI Video Camera Movement Prompts: The Half Everyone Omits

I spent an afternoon last week trying to get one camera movement to work.

Not a complicated shot. A woman at a kitchen table, and I wanted the frame to drift left and end up on the window behind her. I wrote "camera moves left." I got a static shot. I wrote "slow pan to the left, cinematic." Static shot with slightly different lighting. I wrote "the camera slowly pans left across the room" and got a shot that twitched about four degrees and stopped.

Then I opened the prompt guide that ships with the H3 weights, found the camera section, and rewrote the line as The camera pans left with large amplitude at slow speed, revealing the window behind her. First take.

Nothing about my intent had changed. What changed was that I stopped inventing the vocabulary and started using the published one — and I added the part I had been leaving off every single time.

A camera movement instruction has three parts, not one

Most camera movement prompts carry one: a direction. That is the reason most of them come back weak.

The format the model was trained on has three:

[motion type] + [amplitude and speed, when they matter] + [what the move lands on]

Here are the three examples MiniMax prints in its own guide, unedited:

The camera pushes in with small amplitude at slow speed toward the folded letter in her hands.
The camera pans right with large amplitude at fast speed, revealing the open doorway.
The camera holds a static shot as the runner exits the frame.

Read them again and notice what every one of them ends on. "toward the folded letter." "revealing the open doorway." "as the runner exits the frame." Not one of them stops at the direction.

That closing clause is what I had been omitting for an entire afternoon. A direction tells the model which way the camera movement goes. It does not say how far, when to stop, or what the shot is for. Left to guess, the model does the safe thing and barely moves — which reads, on your screen, exactly like the instruction being ignored.

The 20 camera movements the model actually reads

The published vocabulary is fixed. Twenty motion types, and anything outside them is a guess.

Lens pivots or zooms, camera body stays put

The whole camera moves

Defined by the subject

Roll and shake

I keep this open in a tab. There is also an interactive version of this table that assembles the full clause for you, if you would rather click than remember.

Pan is not Truck. Zoom is not Push.

These two pairs cause more wrong camera movements than the rest of the list combined, and the reason is that in casual speech we use one word for both.

Pan keeps the camera bolted in one place and swivels the lens, like turning your head. Truck picks the whole camera up and carries it sideways, like walking. On screen these look nothing alike: when you truck, near objects slide past faster than far ones and the scene gains depth. When you pan, everything sweeps together and the space stays flat.

So "camera moves left" is not just unpublished vocabulary — it is genuinely ambiguous. It could mean either. The model has to pick, and it will not necessarily pick yours.

Zoom changes focal length while the body sits still, which flattens faces and compresses the background. Push In physically carries the camera forward, which keeps perspective natural and feels like approaching someone. Directors choose between these deliberately. If you write "zoom" when you meant "push," you get a technically correct shot that feels wrong and you will not be able to say why.

Amplitude and speed, and when to say nothing

Two modifiers, each with two published values:

• with small amplitude / with large amplitude — how far the composition travels
• at slow speed / at fast speed — how fast it gets there

They are independent, which is the useful part. A large camera movement at slow speed is a reveal that builds. A small move at fast speed is a flinch. Same motion type, completely different shot.

The guide is explicit that medium amplitude and normal speed are normally left out. That instruction is easy to skim past, and it matters more than it looks: every extra word in a camera clause competes for attention with the words carrying real instruction. Writing "at normal speed" does not reinforce the default, it dilutes the sentence. Say nothing and you get the default.

Four of the twenty take no modifiers at all. Static Shot and POV describe a state rather than a movement, and Shake Slightly and Shake Strongly already carry their own degree — "shakes strongly with large amplitude" is vocabulary the guide never defines, so you are back to guessing.

Give the camera movement somewhere to land

This is the one that cost me the afternoon, so it gets its own section.

Compare:

The camera pans left with large amplitude at slow speed.
The camera pans left with large amplitude at slow speed, revealing the window behind her.

The first is a complete instruction by any reasonable reading. It still underperforms, because "pan left" describes a process with no end condition. The second gives the pan a destination, and the destination is what tells the model how far the move goes and when it is finished.

Three connectives cover almost everything:

• toward … — the move approaches something. pushes in toward the open watch case between his hands
• revealing … — the move brings something into frame that was not there. trucks right, revealing the queue that has formed behind him
• as … — the move is timed against an action. holds a static shot as the runner exits the frame

If you cannot finish the sentence with one of those three, that is worth pausing on. It usually means the shot has no reason to move, and a static shot would serve the scene better than a drift the model has to invent a purpose for.

Copy-paste template

[Shot 1] Live-action, cinematic, <opening framing and scene>.
The camera <motion type> <with … amplitude> <at … speed> <toward / revealing / as …>.
<What the subject does, if anything.>

Filled in:

[Shot 1] Live-action, cinematic, a medium-wide shot frames a watchmaker at a
cluttered bench under a single warm desk lamp. The camera pushes in with small
amplitude at slow speed toward the open watch case between his hands. He sets
down the tweezers and turns the movement a quarter turn to catch the light.

Note the order: the scene is established first, then the camera acts on it. Camera instructions stacked at the end of a prompt — , cinematic, pan right, 4k, dynamic camera — read as style tags rather than direction, and get treated with about that much weight.

Six mistakes that quietly cost you takes

1. Inventing the verb. "Drone shot," "slow zoom out effect," "camera glides" — all reasonable English, none of them in the published list. The model was trained on the published wording. A synonym is a coin flip.

2. Stopping at the direction. Half a camera movement instruction. Covered above. It is the single most common one, and the most expensive, because the output looks like the instruction was ignored rather than incomplete.

3. Asking one shot to change direction. Pan left for two seconds, then pan right — a single shot holds one camera movement. Instructions that fight each other inside one shot tend to cancel, and you get a camera that barely moves at all. If you need two moves, cut: [Shot 2] At 00:04.000, the camera cuts to … and give the second shot its own motion type.

4. Burying the camera in the scene description. Write the move as an action inside the shot, not as a tag at the end.

5. Modifying something that takes no modifiers. Static, POV, and the two Shake types. Adding amplitude to them produces a phrase the guide does not define.

6. Shaking for the whole clip. Shake Strongly across an entire take tends to destabilise everything else in frame — faces smear, text in shot becomes unreadable. If you want impact, cut into the shake rather than running it end to end.

A three-step routine

Write the shot static first. Get the framing, the subject and the action right with The camera holds a static shot. If the scene does not work standing still, camera movement will not rescue it.

Add one camera movement, with its landing. One motion type, one destination. Read the sentence out loud — if it ends on a direction rather than a thing, it is not finished.

Only then reach for modifiers. Amplitude and speed are for when the default is visibly wrong, not for every shot. If you cannot say what is wrong with the default, leave them out.

For shots that have to end on a specific frame rather than a specific move, that is a different prompt format with its own alignment line — the first and last frame builder writes that one. And the six-part prompt guide covers the structure the camera clause sits inside.

Bottom line

Camera movement in AI video is not a settings problem. There is no joystick and no keyframe track — the model reads the words you gave it, and it only recognises a fixed set of them.

So writing a camera movement prompt comes down to two habits. Use the published verb instead of a synonym. And finish the sentence: say what the movement lands on, not just which way it goes.

The second one is what fixed my afternoon.

Camera vocabulary in this article is from the Video Prompt Writing Guide published with the MiniMax H3 open weights, checked September 2026. This site is not affiliated with MiniMax.