Prompting Motion for Photo Animation: Describing Movement a Model Can Follow

A video prompt is built differently from an image prompt. For an image you describe a state: what is visible in the frame. For video you describe a process: what changes from start to finish. Mixing the two modes causes most failed clips.

Two independent axes: subject and camera

Any motion in frame combines two things: the subject moves and the camera moves. The model distinguishes them, and they should be described separately.

• Subject motion: "slowly turns her head to the right", "blinks", "smiles faintly", "hair moves in the wind".
• Camera motion: "slow push in", "gentle pull back", "slight lateral drift", "camera is static".

Saying "camera is static" explicitly is not a wasted word. Without it the model often adds camera movement of its own that conflicts with the subject's motion, and the frame drifts in two directions at once.

The structure of a request

1. What is in frame — briefly, so the model does not redefine the content: "a woman in a coat against a street".
2. Subject motion — one action, two at most.
3. Camera motion — one, or explicit stillness.
4. Tempo — "slowly", "smoothly", "barely perceptibly".
5. What stays unchanged — "background static", "pose unchanged".

Why one action beats three

Every additional action shares the same number of frames. A five-second clip physically cannot hold a head turn, a smile, a wave and a camera push — the model will try to fit them all and each will come out jerky.

The rule is simple: one clear action per clip. If you need more, make two clips and cut them together.

Phrasings that work

The model reliably understands descriptions familiar from filming practice: they appeared many times in the data it learned from.

• "slow camera push in" — predictable and almost always clean;
• "a slight head turn to the right", with the side specified;
• "blinks, then smiles faintly" — two micro-phases fit a short clip;
• "wind moves the hair, background static";
• "slow pan left, subject motionless".

Phrasings that do not

• Abstractions: "bring the photo to life", "make it beautiful", "add dynamics". The model does not know what to turn these into and chooses at random.
• Negations: "no sudden movements". Negation works poorly in motion descriptions — say what should happen instead: "slowly and smoothly".
• Precise quantities: "turn by 23 degrees". The model does not measure angles; it follows general meaning.
• Plot: "the person walks to the window and opens it". That is a scene, not the motion of one frame.

Tempo as its own parameter

Models tend towards sharper movement than the author expects. The words "slowly", "smoothly", "barely" almost always improve the result: slow motion is easier to keep consistent between frames, so there are fewer artefacts.

The reverse holds too: trying to get fast, energetic movement nearly guarantees flicker and deformation.

Iterative refinement

The first clip is almost never the final one. A practical cycle: get a result, identify the single biggest problem (too fast? camera interfering? wrong motion?), change only that in the prompt, and rerun.

Changing everything at once is pointless: you will not know which edit worked and will spend twice as many runs reaching the same conclusion.

When the prompt cannot help

If the source is unsuitable for animation — hands in frame, a head cropped flush to the edge, a busy background — no phrasing will fix it. The prompt governs what the model does, not what it has to work with.