Camera Movement Prompts for AI Video: 24 Phrases Models Actually Follow

Ask for a cinematic shot and you will get something vaguely moody and usually static. Ask for a slow dolly push in on a subject and you will get a camera that actually travels. The difference is not taste, it is vocabulary: video models were trained on footage described in shot-list language, so the closer your prompt reads like a line from a shot list, the more consistently the motion comes back.

This page is a working vocabulary rather than a theory essay. Every phrase below is written to be pasted as-is, each one is paired with a subject so the camera has something to move relative to, and each carries a note on why it holds up. Where a move is fragile, that is stated too.

Two rules run through all of them. Name one move per clip, and give the camera a subject to hold. Most failed camera prompts break one of those two rules before they break anything else.

The core idea

Four principles behind every prompt on this page.

Step-by-step

The order that keeps you from debugging five things at once.

1. Fix the shot size first

   Decide wide, medium or close-up before you write the move. A crane up reads completely differently over a wide landscape than over a face, and the model cannot infer which you meant.

2. Choose exactly one move

   Pick from push in, pull back, orbit, lateral track, vertical crane, handheld follow or static. Write it as a verb phrase rather than a description of the result you want.

3. Add direction and speed

   Slow, gentle, steady or fast are the words that actually change pacing. Slow dolly push in behaves very differently from quick dolly push in, and both are more reliable than medium-paced.

4. Give the camera a subject

   Name what the camera is looking at and, where it helps, what it passes. Tracking shot alongside a cyclist on a coastal road, camera at hip height gives the model both a target and a path.

5. Test at five seconds, then extend

   Render five seconds first and watch whether the move completes. If it stalls at three seconds, the move is too ambitious for that subject — shorten it or change the shot size rather than hoping a longer render fixes it.

Prompt recipes

Paste as-is, then swap the subject for yours.

Slow dolly push in on a potter's hands shaping wet clay on a spinning wheel, shallow depth of field, warm window light from the left, medium shot, 35mm

Slow dolly pull back from a single espresso cup on a marble counter to reveal a quiet empty cafe at dawn, soft ambient light, wide shot

Slow orbit left around a matte black sports car parked under concrete, late afternoon sun, low angle, reflections sliding across the bodywork, wide shot

Steady lateral tracking shot following a cyclist riding along a coastal road at golden hour, camera at hip height, background palms drifting past, medium wide

Slow crane up from a close-up of wet cobblestones to reveal a narrow European street with laundry lines above, overcast light, vertical reveal

Loose handheld shot following a chef plating a dish in a busy kitchen, subtle camera sway, practical warm light, close medium shot

Fast whip pan from a graffiti-covered wall to a skateboarder dropping into a concrete bowl, motion blur through the transition, wide shot

Locked-off static camera, a horse walking through shallow water at dusk, ripples spreading, distant hills, wide shot

Common mistakes

The four that account for most bad results.

Why one clear move beats a paragraph of adjectives

Camera language works because it is the language footage is described in.

Training captions for video tend to be written the way a shot list or a stock footage description is written: subject, action, camera behaviour, lens, light. A prompt that follows that shape is closer to the distribution the model learned, which is why a plain sentence like slow dolly push in outperforms a decorated one like a breathtaking cinematic journey into the soul of the city.

The second reason is ambiguity. An adjective can be satisfied in dozens of ways, and the model will take whichever is most common in its training data — usually a static, slightly graded shot. A named move has far fewer valid interpretations, so the output distribution tightens and the same prompt reproduces more consistently across runs.

That also explains the failure mode everyone hits: the move that starts and dies. When a prompt contains two competing motion instructions, the model allocates attention between them, and the resulting clip often begins with one move and drifts into a static frame. Removing the competing instruction fixes it more reliably than adding emphasis or repeating the phrase.

Practise it in OpenFilmo

Every recipe on this page runs in the same workspace.

A five-second clip costs 5 credits and one image costs 2. The Free plan gives you 30 credits every month with no credit card, no watermark and commercial use included — enough for six five-second tests before you pay anything. Paid plans start at $9 for 400 credits and add faster generation, parallel tasks and higher export resolution.

Related guides

Longer reads on the same workflow, from the OpenFilmo blog.

AI tools

Single-purpose tools that run in the same workspace, on the same credit balance.

Vidu models

The Vidu Q1 and Q2 models this site runs on, and how they compare to Kling, Veo and Runway.

Related workflows

Closely related tools and guides on the same problem.

Frequently asked questions