Image-to-video motion prompts: words that move a picture, and words that warp it

When a still becomes a clip, the model keeps your picture and adds the one thing you cannot draw: time. The prompt is where you tell it how time should pass. Most disappointing image-to-video results are not model failures but vocabulary failures, a prompt that describes the picture again, or asks for four kinds of motion in a 5-second clip. This is a glossary rather than a tutorial: the words that reliably move a picture, the words that reliably melt it, and five prompts rewritten from one to the other. The step-by-step guide covers how to attach the frame and pick a length; this post is only about what to type.

The rule behind every word on this page

A motion prompt has three slots: what the camera does, what the subject does, and what the air does (wind, water, smoke, light). Fill each slot with one concrete, directional phrase and leave the rest empty. The model reads a short clip as a budget of change; every extra phrase spends some of it, and when the budget runs out the model starts changing things you did not ask about. That is warping.

Camera words that hold up

• Static camera, locked off. The safest phrase in the language. Use it whenever the subject is a face, a product or anything with straight edges.
• Slow push-in or slow pull-back. A gentle move toward or away from the subject. "Slow" is doing real work here; drop it and the model may rush and smear.
• Gentle pan left or gentle pan right. Rotates the view. Works best when the start frame has something interesting at the edge to pan toward.
• Slow tilt up. Good for buildings, trees and anything tall with sky above it.
• Small orbit around the subject. Flattering for objects on a table; risky for people, because the far side of a head has to be invented.
• Handheld drift. A faint documentary wobble. Reads as real footage; keep it "subtle" or it becomes an earthquake.

Subject words that hold up

Pick verbs with a direction and a speed built in. "Turns slowly toward the camera", "steam rises from the cup", "the curtain lifts in a breeze", "the fox lowers its head to drink". Each names one body part or one object, one movement, one tempo. If the subject is a person, the smallest motions are the most convincing: a blink, a breath, a glance, hair shifting. A full gesture across the frame asks the model to redraw the hand twelve times a second, and the hands and text guide explains why that rarely ends well.

Atmosphere words that hold up

• "Clouds drift slowly across the sky"
• "Light flickers from the candle"
• "Rain streaks down the glass"
• "Dust motes float in the beam"
• "Waves roll in and foam over the sand"

Atmosphere is the cheapest motion you can buy. It happens behind or around the subject, so the model can keep the subject exactly as drawn while still making the clip feel alive. If a still has no obvious atmospheric element, name one that plausibly could be there: "a thin haze in the air" is almost never wrong.

The words that warp

• "Dynamic", "epic", "cinematic", "dramatic". They carry no direction and no speed. The model fills the gap with random large movement.
• Two camera moves in one clip. "Pan right then push in" is two shots. Make two clips.
• Transformation verbs. "Turns into", "becomes", "morphs". You are asking for the picture to stop being your picture.
• Fast anything. "Quickly", "rapidly", "sprints". Speed means more change per frame, and more change per frame means more invented pixels.
• Walking toward the camera. The feet freeze, the head drifts, the legs scissor. If a walk is essential, keep it short and add "natural gait, feet moving".
• Re-describing the scene. "A woman in a red coat standing by a lake at sunset" is already in the image. Every word of it competes with the motion words for attention.

Five prompts, before and after

Pacing words and clip length

"Slow", "gentle", "gradual" and "subtle" are not decoration: they lower the amount of change the model attempts per frame, which is exactly what keeps a still looking like itself. Pair them with a short clip. The studio renders up to 30 seconds on some models, but a 5-second clip of one clean motion is almost always more usable than a long one, and it is also cheaper, because price scales with seconds and resolution and shows in the composer before you confirm. Test the words at the lowest resolution first; once the motion is right, re-render the final. If a render fails, the amount returns to your balance automatically.

Which model listens best

Every video model in the studio accepts a start frame through the Animate action or the Reference button, but they differ in how literally they read motion words. Motion Real is the one to reach for when physical plausibility matters, water and fabric in particular. Motion gives the most detail at the top resolutions and tolerates a slightly fuller prompt. Motion Turbo is for finding out quickly whether an idea moves at all. The model comparison sets them side by side, and the text-to-video guide has the longer camera glossary if you are generating without a start frame.

What no prompt can do

A motion prompt cannot add something that is not in the frame and have it look like it belongs; a new character walking in will be invented from nothing. It cannot keep a long line of text legible once the camera moves, so lock the camera for shots with packaging or signs. It cannot follow a subject out of frame and show what is behind it. And it cannot turn a blurry or badly composed still into a sharp clip: the start frame is the ceiling. Fix the picture first, then open the studio, attach it, and type three slots' worth of words.