How to write image-to-video prompts that work

You have a portrait you love. You upload it, type a long description of everything in it, hit generate, and get five seconds of a face that barely moves, or a body that turns to rubber halfway through. The photo usually isn't the problem. The prompt is doing the wrong job. This guide shows you how to write image-to-video prompts that describe the motion you want, so one still turns into a clip you can actually post.

What you need

• One sharp, well-lit still — your avatar, a portrait or a product shot — made at the aspect ratio you'll post, such as 9:16 for reels.
• An image-to-video model — for example Runway Gen-4.5, Kling 3.0, or the image-to-video step on ai-creator.academy.
• Short clips — Runway's Gen-4 generates 5- or 10-second clips. Plan one scene per clip.
• A notes file for prompts that worked, so you can reuse them next week.

Step by step

1. Fix the still before you animate it. The image becomes the first frame and sets the composition, lighting and style of the whole clip. Runway warns that blurry hands or faces get worse once animated, so regenerate or retouch them first.
2. Stop describing the picture. The model can already see your image, and re-describing it in detail can reduce the motion. Runway's own example swaps a full description of the man's hair, suit and tie for "The man extends his arm to shake hands, then nods politely."

3. Use one sentence shape. Start with Runway's structure: "The camera [motion] as the subject [action]." For example: "The camera slowly pushes in as the subject turns her head toward the window." Call people "the subject", "she" or "the man", not their full description.
4. Pick one motion that fits the pose. Ask for movement that grows out of the frame: a slight head turn, a blink, hair moving in a light breeze. A seated subject who suddenly runs usually breaks the shot.
5. Choose one camera move. Slow push-in, slow pan, slow orbit or a locked camera. Pick one; combining pan, zoom and orbit with lots of background action in one clip is a common reason clips fall apart.
6. Say what you want, never what you don't. Runway's Gen-4 guide says negative phrasing isn't supported and can give the opposite result. Instead of "No camera movement", write "Locked camera. The camera remains still."
7. Add one element at a time. Get the core motion working first, then add subject motion, camera motion, scene motion and style one by one. That way you can see which addition helped, and which one broke the clip.

The following images were generated using Nano Banana 2:

Worked example — the window portrait. This still is built for a slow push-in: the profile pose, the window light and the loose strands of hair already suggest motion, so the video prompt needs only one line: "The camera slowly pushes in as the subject turns her head slightly toward the camera. Her hair moves gently in the breeze."

Prompt (Nano Banana 2, 1:1):
Medium close-up portrait of a woman in her thirties with shoulder-length auburn hair, in profile next to a bright window, soft natural window light on her face, loose strands of hair lifting slightly, cream knit sweater, plain light background, shallow depth of field, realistic skin texture with fine pores, editorial photography, 85mm lens. No text, no logos.

Worked example — the cafe shot. Here the camera stays put and the motion is small and inside the frame. The video prompt: "Locked camera. The camera remains still. Steam rises from the cup as the man looks up and smiles." Nothing about his face, clothes or the cafe, because the image already holds all of that.

Prompt (Nano Banana 2, 1:1):
Medium close-up of an East Asian man in his twenties sitting at a sunlit cafe table, both hands around a ceramic coffee cup with light steam rising, looking down at the cup with a relaxed expression, warm morning light from the side, blurred cafe background with plants, realistic skin texture, editorial photography, 50mm lens. No text, no logos, no signs.

Common mistakes

• Pasting your image prompt into the video box. Delete everything the picture already shows and keep only what moves.
• Stacking three actions into five seconds. Keep one action per clip. If you need a sequence, write "X happens, then Y happens" or use rough timestamps like [00:01] and [00:03], and pick the longer duration.
• Writing "no zoom, no shaking". Describe the camera you want instead: "Locked camera." or "Slow push-in."
• Fighting the image. Motion blur or a mid-step pose tells the model something is moving. If you want stillness, start from a still that looks still.

The bottom line

Image-to-video rewards restraint: a clean first frame, one sentence about what moves, one camera move and positive wording. Lock the look in the still, and let the prompt handle only the motion. If you don't have that still yet, start with your avatar.

FAQs

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

One or two sentences is enough to start. Runway recommends beginning with the most important motion and adding detail only when you need it.

Should I describe my character's face and outfit in the prompt?

Usually not, because the image already sets them. Describe visuals only when something new enters the scene, something changes a lot from the starting image, or two elements interact.

Why does my clip have cuts I didn't ask for?

Runway suggests trying a longer duration first. If cuts continue, check your prompt for wording that sounds like a scene change and add "Continuous, seamless shot."

Can I make a clip longer than 10 seconds?

Yes. Take the last frame of a finished clip, use it as the input image for the next one, and join the two clips in an editor.

What's the safest motion for a portrait?

A slight head turn, a blink or a small change in expression, hair moving in a light breeze, and a gentle push-in. Exaggerated facial movement and fast body turns are where faces start to drift.

Sources: Runway — Image to Video Prompting Guide · Runway — Gen-4 Video Prompting Guide · Vofy — How to use Kling 3.0 image-to-video