What should an image-to-video prompt say?
Describe what changes, not what is already in the picture: the motion, the camera, the sound and the timing, plus a line telling the model to keep the face and outfit exactly as in the image. The image carries the look; the prompt carries everything that moves.
Should you describe what is in the picture?
Mostly no. Re-describing the image invites the model to reinterpret it, and a description that disagrees with the picture (a different hair colour, a different room) makes the face or set drift. Our rule since making Penny, a character explainer built from 31 separate clips: the style sentence lives in the image prompt, and the video prompt describes only motion.
Name the image's role and a few anchors instead:
@Image1: The opening frame: the dancer, her outfit, the tower top and the twilight view. Open on exactly this composition.
That line, from our video Look Up, tells the model the picture is frame one, not just inspiration. When there are several people, list who is where, so each face is tied to a position.
How to add motion
- One action per beat, with a time range. "0-3s: she steps back into place" beats "she moves around".
- Say how the camera behaves. If it should not move, say "no camera movement". If it should, give one move: "a steady dolly push toward the subject".
- Give the environment something to do: wind in hair, steam, candle flicker, confetti. Small secondary motion makes a still image feel alive.
- Add sound if the model generates audio: three to five specific sounds and any dialogue in quotes.
- Add "no slow motion" for real-time movement.
Here is the whole video prompt for one Penny clip. It is short because the start frame already holds the character, set and style:
She strains and drags the huge gold bar a few steps across the cobblestones, sweat drops popping off her head. Camera follows. Low-poly 3D game animation, stylized, smooth motion. Nobody speaks; mouths closed or smiling. No captions, no text overlays, no new characters appearing.
How to stop the face changing
- Say it in the avoid list: "No change to her face." Our character prompts carry this line.
- Start from a frame painted from your master reference, so the model begins on-model instead of guessing.
- Do not change the outfit or hair in the prompt. If the look must change, make a new start image first.
- Stop new people walking in: "no new characters appearing". Extra faces tend to borrow features from the main one.
- Keep the face original. A face that resembles a famous actor can get the whole render blocked by copyright filters, after the render, in our experience.
More on this in our character consistency guide.
Model-specific details
- Seedance 2.5: in our image-to-video renders, sending an aspect ratio failed the job because the output ratio follows the first frame. Give it a 9:16 frame and leave the ratio out.
- Midjourney video is image-to-video only: the image is the first frame and the text prompt is optional[1].
- Kling 4.0 supports start and end frames plus up to 10 keyframes in image-to-video[2], so you can pin the ending image too.
- Veo 3.1 accepts up to three reference images for style and content[3].
- Going longer: take the last frame of one clip and use it as the start image of the next, with the rest of the timeline. Feeding the first clip back as a video reference restarted the scene in our test.
The free prompt builder reserves @Image1 for your picture and writes the motion timeline around it.
Where Cast fits
- You have no start image: Cast draws your character from a sentence (four Midjourney v8.2 looks for 15 credits) and uses the one you pick as the reference for every video.
- Animating the free Malfoid & Friends characters with one line of action.
Where it does not:
- Animating your own photo, drawing or product shot: Cast has no image upload.
- Pinning an exact end frame or keyframes: Cast takes character references, not a start-and-end frame pair.
People also ask
- Should an image-to-video prompt describe the image?
- Only enough to anchor it: who is who and that it is the opening frame. Spend the rest of the prompt on motion, camera and sound.
- Why does my character's face change in image-to-video?
- Usually the prompt re-describes the person differently from the image, new people enter the frame, or the camera moves far from the start composition. Add "No change to her face" and keep the camera simple.
- What is a good image-to-video prompt for Hailuo?
- The same structure applies, but we have not tested Hailuo's current model, which its site promotes as MiniMax H3 with Omni Reference[4]. Start with motion-only prompts and compare.
- How long should an image-to-video prompt be?
- Shorter than text-to-video. Our Penny clips used one or two sentences of motion per clip because the image carried everything else.
Sources
Claims about other companies come from their own pages, read on the date shown. Products and prices change; check the link for the current version.
- Midjourney docs: Video (read )
- Kling AI: AI video generator feature page (read )
- Google: Veo on the Gemini API (read )
- Hailuo AI: home page (read )
Related answers
- How do I write a prompt for an AI video?
- Why does my AI character's face change between videos?
- How do I keep an AI character consistent across videos?
- Can AI animate a drawing of my character?
- What is the best image-to-video AI in 2026?
Part of the Cast answers library (Prompting).
Updated