Turn a reference video into an image-to-video prompt

1. Give the image and the text different jobs

A reference analysis may describe the wardrobe, light, camera angle, action, and ending in one detailed record. Keep that record as your brief. For the image-to-video step, decide which information should already be visible and which information describes a change.

Runway's current prompting guide explains that an image establishes the visual setup while the text guides movement. Its advice also allows visual detail when introducing something new or describing a transformation. You do not need to strip out every noun; you need to avoid restating the whole image.

If you are still mapping the reference, begin with the scene-by-scene prompt workflow. Here, work on one selected shot rather than pasting a full sequence into one generation.

2. Prepare a starting image that can support the action

1. Write the opening state. Note who or what is visible and where the action begins. For a drawer opening, start with the drawer closed; for a pickup, start with the object within reach.
2. Choose your own visual content. Carry over a useful composition or movement idea, then design your own subject and setting. Prepare an original photograph, illustration, or generated still for the new shot.
3. Inspect the actual image. Check object edges, hands, faces, and the space needed for the action. A prompt describing a perfect image does not repair a flawed file automatically.
4. Check the selected generation mode. Use an image-to-video workflow that accepts a starting image. Follow that tool's current input and duration requirements; a text description is not an image attachment.

Google's video generation best practices similarly recommend a clear source image and motion-focused text. They distinguish camera movement, subject animation, and environmental changes. Choose which of those carries your shot before adding secondary movement.

Save palette and material decisions in your image brief using the color and texture card. Keep a copy of the approved starting image so later attempts can use the same input.

3. Write one movement with a visible finish

Draft the action in plain language, then read it while looking at your image. Can you point to the subject that moves? Is there room for the camera view you want? What would count as a completed action? Use the camera movement guide if you need to separate a moving subject from a moving viewpoint.

Starting image: [approved image file; keep this in your notes]
Subject action: [who moves + direction + pace]
Camera: [one movement or a fixed position]
Environment: [optional supporting motion]
Ending: [the visible state to reach]
Review: [one action check + one continuity check]

This is a planning card, not syntax to paste into every model. Combine the relevant motion fields into a short prompt. Keep the filename and review notes outside the prompt. A camera can move while a subject stays still, so state those choices separately.

Prefer a concrete endpoint such as a packet held at chest height to a mood instruction such as an inspiring reveal. If the shot needs several distinct events, first try the one event that makes its purpose clear. Add the others only after reviewing that attempt.

4. Try three different image and motion pairings

These original examples are illustrative briefs, not tested generation results. Prepare the described image first, then adapt the motion text to the tool you use.

5. Review the handoff from still image to moving shot

Watch once at normal speed, then inspect the beginning, middle, and end. Record whether the action completed and whether the important object stayed recognizable. A pleasing first frame alone does not answer either question. Compare against your new brief rather than expecting an exact reconstruction of the reference video.

The clip barely moves

Identify one visible action and remove repeated appearance notes. Check that the action can happen from the starting pose.

The wrong object moves

Identify it by position or a short visible feature, such as the person beside the table. Give the other subject a clear resting state.

The scene changes unexpectedly

Check for a hidden request for a second location or shot. Keep one continuous action, then plan the next shot separately.

The output begins with the action already finished

Revise the starting image so it shows the setup: hand beside the packet, closed drawer, or unturned object.

The ending never arrives

Shorten the action or use a longer duration if the selected tool supports it. Remove intermediate events before adding more timing instructions.

Save the image, prompt, model or mode, and generation settings with each attempt. Change either the image or the motion wording first so you can track your revision. Reusing the same inputs does not guarantee identical outputs; the log helps you compare decisions, not prove that one wording always performs better.

Image-to-video prompt FAQ