Select one complete action with a visible beginning and end. Do not paste a whole montage analysis into a request intended to produce one short shot. A multi-shot reference can supply a visual vocabulary, but its scene changes, subject changes and pacing need editorial decisions before they become executable generation instructions.
Find the action that carries the idea
Imagine a reference with a rooftop view, a close-up of a watch, a person walking, a passing train and a final smile. The distinctive idea might be a commuter noticing the arriving train, not all five images. Choose the action that communicates the purpose of your new piece.
Write its beginning and ending in plain language. “A person stands beside a closed door; the door opens and they turn toward it” is a complete change. “Urban cinematic lifestyle” supplies no action boundary and gives you little basis for evaluating the output.
Sort the analysis into keep, omit and save
Keep the subject, environment and motion needed for that action. Omit unrelated locations and secondary events. Save useful ideas for separate future shots rather than stuffing them into the same prompt. This is an editorial reduction, not an assertion that the omitted parts were unimportant in the original video.
| Reference detail | Decision for the new shot | Reason |
|---|---|---|
| Warm station light | Keep | Defines the chosen environment |
| Watch close-up | Save as a separate insert | Requires another framing |
| Rooftop drone shot | Omit | Belongs to another location |
| Turn toward opening doors | Keep | Provides one complete action |
Write an action-only first draft
A traveler waits beside the platform line. As the train doors open, the traveler turns toward them and takes one measured step forward. Keep the camera at shoulder height in a steady medium-wide view. The train and platform remain the same throughout the shot.
Add the essential visual context only after this works as a sentence. If an input image already establishes clothing and layout, avoid filling the text with conflicting redesign instructions. Runway's Gen-4 prompting guide is a useful official example of motion-focused instruction; its specific model behavior should not be assumed for another service.
Check whether the destination can execute the plan
Use Imagild Reverse Prompt to help describe your authorized source, then review and adapt the result. Analysis does not expand a generation model's available duration or provide unsupported multi-scene control. If you need three distinct framings, plan three shots and an edit rather than relying on a single overloaded instruction.
Watch the resulting clip for a complete action, not merely a resemblance to one reference frame. If the doors are already open at the start or the traveler never reacts, the beginning-to-end change is missing. Fix that issue before adding a more elaborate camera move.
Should I delete all camera instructions?
No. Keep the camera behavior needed to make the action legible. The aim is one coherent shot, not an instruction with no framing at all.
Can I preserve the montage's rhythm?
Yes, at the sequence level. Record the pacing separately and use it when assembling multiple accepted clips. Rhythm is not the same as asking one generated shot to contain every source scene.