How-to

Image-to-Video Prompts That Work, With Real Examples

How to write an image-to-video prompt: what to describe, how to time it and what to leave out. Rules from our own dance tests, copyable prompts and three real clips from three different video models.

Published November 6, 2026

Describe only what moves. The start image already shows who is in the frame, what they wear and where they are, so a good image-to-video prompt gives the action and the camera in time order, then stops. In Studio you can even leave the box empty: a single start image then runs on a default of a locked camera with gentle, natural movement of everything in the frame.

Open Studio, add an image and try a prompt.

Open Studio

This guide comes from prompts we wrote and clips we rejected while building effects on reRAW AI, an all-in-one AI video and image studio. The timing and short-prompt rules come from a series of dance tests in July 2026. Where we saw a rule on more than one engine, we say so.

Start with the right image

Keep the aspect ratio between 0.4 and 2.5 (width divided by height) for the Seedance family and for Kling Motion Control. Outside that range the request is refused as too wide or too tall, so crop to something closer to normal first.
For body motion, show the whole body including the feet, with space around the person. Reflections in mirrors break once things move, so avoid mirrors. Fitted clothes are steadier than oversized ones, and a long skirt is a plus because it gives the motion something to do.
For Kling Motion Control, head, shoulders and torso need to be visible. Otherwise the engine finds no person, the run is refunded and you are asked for a clearer picture.

Write the prompt

Timed segments, one main action each

In our dance tests a flowing paragraph got only its first few actions performed. The same idea split into timed segments, with one main action per segment, fixed it. A prompt for a 10 s dance clip looks like this:

Timed prompt, 10 s

She dances a playful pop routine, full body in frame, camera locked static, vertical.
[0-2s] Shoulders bouncing on every beat, quick step-taps in place.
[2-5s] Both arms swing up, cross overhead and snap down to her hips in one move.
[5-8s] One fast half-turn and back to the lens, a small hair flick at the end.
[8-10s] She freezes: hand on hip, chin down, grinning at the camera.
Fast and precise, every move on the beat, smooth continuous motion.
AUDIO: upbeat pop with a clean drum loop.

If a segment is skipped, it held two main actions. Cut one or split them, and do not add adjectives. Engines also read a time stamp as a cut. When you want one unbroken camera move, write "continuous shot, no cut" inside that segment. We saw this on Grok and on Seedance.

Leave the scene out

Do not describe the room, the background or the outfit again. When a prompt names a new place, the model takes it as an instruction and can pull the subject out of the original frame. Camera moves and light changes inside the same scene are fine.

Re-describes the picture

A woman in a red dress sits in a sunlit Paris cafe with a latte on a marble table and a street full of pedestrians behind her. She turns and smiles.

Describes the motion

She turns toward the window and smiles, a light breeze moving her hair. Slow push-in.

Physical verbs, pushed further than you think

Snapping, bouncing, flicking and shimmying read well. Engines under-deliver on intensity, so overstate it: "hips snapping hard on every beat" rather than "swaying". Keep each clip to one kind of motion, because mixed dance styles break up into stutters.

End on a held frame

Our dance clips finish on a freeze with a look to camera, which gives the clip a clear ending.

Start and end frames

When you supply both frames, the prompt has to say how the first becomes the last. Studio will not run a start and end frame pair without text.

Start and end frame

From the first frame to the last frame: ribbons of light spiral up from her feet and wrap around her, the outfit rebuilds layer by layer, and the move ends on the pose in the last frame. The camera pushes in slowly.

Do not use the same picture as first and last frame when you want a full orbit. In our 360 spin tests the camera turned about halfway and came back, so that effect gives the engine a first frame only.

Three clips, three kinds of start

The inputs and the models differ, so these are not head-to-head comparisons. They show what each kind of prompt looks like in practice.

First frame, 720p, 10 s. The start frame is a drawn character placed in a real convention hall. The prompt is a short script of beats: staff lead the character in, a wave, a pose to camera, a turn to follow, a kiss over the shoulder. We write it separately for 5 s and 10 s clips.

720p, 5 s, first frame only. One instruction: the camera circles the subject once while the moment stays frozen.

1080p, first and last frame. The first frame is a photo of a cat and the last is a costume still, so the prompt only describes the change between them.

When the clip goes wrong

The action never happens. A segment holds two actions, so split it or cut one.
The subject jumps to a new place. The prompt describes a scene, so remove it.
The face melts during a spin. Ask for a half turn and back to the lens instead of a full spin.
The request is refused as too wide or too tall. Crop the image to an aspect ratio between 0.4 and 2.5.

Can I leave the prompt empty?

In Studio, yes, for a single start image. Anything that adds a second input needs text: an end frame, several references, a reference video or audio.

How long should an image-to-video prompt be?

Short when the image carries the look. Our dance prompts run to about seven lines, as in the example above. A text-to-video prompt with no start image has to build the whole scene, so it is longer.

Which model is best for image-to-video?

We have not run one photo through several engines with the same prompt and clip length, so this page does not rank them.

Add an image in Studio and write the first line of motion.

Open Studio