Camera movements for AI video: what each one does to a frame

Camera movement is the single highest-leverage word in an AI video prompt. It decides whether a clip reads as a moving photograph or as a shot. Get it right and a mediocre image becomes usable; get it wrong and a beautiful image warps at the edges for six seconds.

This is a reference for the moves that behave predictably, what each one asks of the source frame, and the failure mode of each. It pairs with our image-to-video prompt guide, which covers the rest of the prompt.

The rule that governs all of them

One move per clip. A model given "slow push in while panning left and tilting up" will attempt all three and commit to none, and the result reads as a camera falling downstairs. If you want a compound move, generate two clips and cut between them on the timeline.

The second rule follows from the first: the move must be smaller than the depth available. Every camera movement reveals frame area that was not in the source image, and the model has to invent it. A move that reveals a little invents a little. A move that reveals a lot invents a lot, and invention is where warping lives.

Push in

The camera moves toward the subject. The most reliable move in the set, and the right default when you are unsure.

Phrase it: slow push in toward the subject, or gentle dolly in.

Needs: margin around the subject and a background with some depth. The push consumes frame area, so a subject already filling the frame has nowhere to go.

Fails when: the subject is cropped tight, or the push is described as fast. "Rapid push in" produces smearing at the edges almost every time. Keep it slow.

Best for: revealing detail, building tension, ending on a subject. It is the move that makes a still portrait feel like a moment.

Pull out

The reverse: the camera retreats, revealing context around the subject.

Phrase it: slow pull back revealing the surroundings.

Needs: more invention than a push, because everything newly visible has to be generated. It works best when the source image implies its own continuation — a room that obviously extends, a landscape that obviously continues.

Fails when: the source has a hard edge the model cannot extend plausibly, such as a subject photographed against a plain continuous studio backdrop. You get a backdrop that grows strange objects.

Best for: openings, establishing shots, the "it was bigger than you thought" beat.

Pan

The camera rotates horizontally from a fixed position, sweeping across a scene.

Phrase it: slow pan left across the scene. Name the direction; "pan" alone gets you a coin flip.

Needs: horizontal content worth sweeping across. A landscape, a interior, a row of objects.

Fails when: there is a strong single subject in the centre. The pan takes the subject out of frame and the model spends the rest of the clip inventing whatever is beside it.

Best for: landscapes, interiors, revealing a second subject.

Tilt

The vertical equivalent: the camera rotates up or down.

Phrase it: slow tilt up from the base to the top.

Needs: vertical content — architecture, trees, a figure seen from below.

Fails when: applied to a horizontal composition. A tilt on a wide landscape mostly reveals sky and ground, both of which the model invents flatly.

Best for: scale. A tilt up a building or a tree is the cheapest way to make something feel large.

Orbit

The camera arcs around the subject, keeping it roughly centred.

Phrase it: slow orbit around the subject.

Needs: the most depth of any move, because it reveals sides of the subject that do not exist in a flat image. The model has to hallucinate a three-dimensional object from one view.

Fails when: the subject is complex or asymmetric. Faces are the hardest case — an orbit around a generated portrait will usually distort the features. Products and simple geometric objects orbit far better.

Best for: product shots, single objects, anything you would put on a turntable.

Static with internal motion

No camera move at all. Everything that moves is in the scene.

Phrase it: static camera, [thing] moving. Say "static" explicitly — an unspecified camera tends to drift.

Needs: nothing but a subject with something plausibly in motion. Steam, water, cloth, hair, light.

Fails when: the scene has nothing to animate. You get an almost-still frame with subtle warping, which is worse than a still.

Best for: loops, backgrounds, ambient B-roll. It is the most underused move and the one with the highest success rate. Our looping video guide covers building clips that cut back to their own start.

Handheld

Small, irregular motion suggesting a person holding the camera.

Phrase it: subtle handheld camera shake.

Needs: nothing special, but it stacks with other moves badly. Use it alone or with a very small push.

Fails when: described strongly. "Heavy handheld shake" produces motion sickness and smearing.

Best for: documentary feel, making a generated scene read as captured rather than composed.

Quick reference

Move	Depth needed	Risk	Best for
Push in	Medium	Low	Subjects, tension, endings
Pull out	High	Medium	Establishing, reveals
Pan	Medium	Medium	Landscapes, interiors
Tilt	Medium	Medium	Scale, architecture
Orbit	High	High	Products, simple objects
Static	Low	Very low	Loops, B-roll, ambience
Handheld	Low	Low	Documentary feel

How to test a move cheaply

Draft at 360p, which costs 30 credits, and at 6 seconds — the default duration and the one where camera moves read most clearly. Compare two drafts of the same image with different moves before you spend 120 credits on a 1080p render. Two 360p drafts cost less than a fifth of one finished 1080p clip, and they answer the only question that matters: does this image support this move?

A free account starts with 150 one-time credits and no card, which is enough to run this comparison on several images.

For the craft behind a clip that holds up, read AI video prompts, Reference to video AI, AI video with sound, AI video quality settings, and What is a oner? The long take.

Frequently Asked Questions

Which camera movement works best for AI video?

A slow push in, by a wide margin. It needs less invented frame area than a pull-out or an orbit, it works on almost any composition with margin around the subject, and it reads as intentional. When a clip has to work on the first attempt, push in.

Why does my orbit shot distort the subject?

Because an orbit asks the model to render sides of the subject that do not exist in the source image. From one flat view it has to guess the geometry, and faces are where guessing shows most. Orbits work well on products and simple objects, poorly on people and complex scenes.

Can I combine two camera movements in one prompt?

You can write it, but the result is usually incoherent — the model attempts both and commits to neither. Generate the two moves as separate clips and cut between them in the editor instead. That is also how the move would be shot in reality.

Does naming a lens or focal length help?

Sometimes, and it is cheap to try. Phrases like shallow depth of field or wide angle shift the look more reliably than a specific millimetre figure. Treat them as style hints rather than camera settings, and do not stack more than one.

How fast should a camera move be?

Slow, almost always. "Slow" and "gentle" are the two most useful words in a motion prompt. Speed multiplies the amount of frame area the model must invent per second, and invention under time pressure is exactly what produces smearing at the edges. Common prompt errors are collected in our guide to AI video prompt mistakes.

Pick one image and run the table

Take a single source image with good depth and draft it three times at 360p — push in, static, and one other move. Watching the same frame under three different cameras is the fastest way to internalise what each does. The image-to-video tool page has the setup, and the Video Studio is where the drafts happen.