Image-to-video prompt guide: subject, camera, constraints

5 min read · Published 2026-10-04

When you start from an image, the picture already decides how things look. Your prompt only has to answer three questions: what moves, how the camera moves, and what must not change. Keeping it that small gives you more predictable results than a long, poetic description.

Why image-to-video prompts are different

In text-to-video, the prompt has to describe everything: the subject, the setting, the light and the style. In image-to-video, the image supplies all of that. If you re-describe it, you give the model more chances to contradict the picture.

Write about change, not appearance. Say what should move and how the camera should behave, then protect the parts that should stay the same.

The three-part structure

Subject action: one small, specific movement. Steam rising, a hand turning a bottle, hair moving in a breeze.
Camera move: one move only. Slow push-in, slow orbit, gentle pan, pull-back reveal, or a locked-off shot with subtle motion.
Constraints: what must stay unchanged. Product shape, label, logo, colors, composition. Add what you do not want, such as no new objects or no cuts.

[Subject action] Steam rises slowly from the cup. [Camera] The camera makes a slow push-in. [Constraints] Keep the cup, saucer and table exactly as in the image. No new objects, no cuts.

Camera moves that tend to work

Slow push-in: adds depth and focus with little risk of distortion.
Slow orbit: good for objects, but keep it partial (30 to 90 degrees) rather than a full circle.
Pan or tilt: reveals a scene gradually. Works well for landscapes and interiors.
Pull-back reveal: starts on a detail and widens. Works for hooks.
Static camera with subtle motion: the safest option when the subject must stay exact.

Examples by subject

Product: A soft light sweep moves across the bottle while the camera slowly pushes in. Keep the shape, label and colors unchanged. No spinning, no new text.

Food: Steam rises from the dish, the camera makes a slow push-in, warm light. Keep the plating exactly as shown.

Portrait: The person blinks naturally and looks slightly toward the camera, very slow push-in. Keep the same face, clothing and background.

Landscape: Clouds drift across the sky and the camera pulls back slowly. Keep the horizon and composition unchanged.

Common mistakes

Asking for several actions at once. One main movement per clip is easier for the model to get right.
Mixing contradictory camera moves, such as zoom in while pulling back.
Using vague words like cinematic or epic without saying what moves.
Leaving out constraints, then being surprised that a logo changed.
Expecting readable text to animate correctly. Add text in an editor after generation.

Iterate one variable at a time

If a result is close but not right, change one thing: the camera move, the speed word, or the constraint line. Changing everything at once makes it impossible to learn what helped.

Keep a simple log of prompt, result and notes. After a handful of tries you will have a personal template that works for your type of image.