Use a generator that separates camera-motion references from character-performance references. Assign your reference clip to the camera input, then describe your own subject's action independently. A dance or acting clip placed in an action-transfer field asks for a different kind of control, even when its camera movement happens to be attractive.
Identify what you actually want to borrow
Watch the reference once without concentrating on the actor. Does the viewpoint move sideways, approach the subject, turn around them, or change height? Note its starting view and final view. Then watch again for cuts. A sudden change of location might be an edit rather than a camera path that can be repeated inside one generated shot.
Write a small camera brief: “Begin beside the doorway, travel slowly backward, finish with both windows visible.” Leave the actor's gestures out. If the attraction is really the actor turning toward the lens, acknowledge that you want performance as well; a camera-only brief cannot describe the whole effect.
Choose the appropriate input contract
Adobe documents a camera-motion reference. Kling's Motion Control guide instead describes character and action references. These labels matter more than a tool's general “video reference” marketing.
| Your goal | Input to look for | What to verify |
|---|---|---|
| Repeat a viewpoint move | Camera-motion reference | Whether subject performance is separately defined |
| Repeat a gesture | Action or performance reference | Which character receives the action |
| Preserve an existing scene | Supported video-editing workflow | What can change outside the selected subject |
Read the current instructions for the chosen mode before preparing files. Duration limits, framing options and supported combinations differ. Do not assume that a reference-video slot accepts every kind of guidance.
Build a short, readable test
For a fictional planning example, a person stands outside a small cafe while the camera retreats to reveal the entrance. Your source photo establishes that person and the cafe. A separately recorded movement clip establishes the intended retreat. The person's only action is a slight head turn.
Test this simple combination before adding customers, rain or a passing vehicle. Keep the movement small enough that the starting scene can plausibly support it. A long retreat that reveals an entire street requires much more invented space than a short movement around the entrance.
In an Imagild film plan, record the camera and performance as separate requirements. The written brief is portable; the renderer still needs to accept the intended reference type. A reverse-prompt description does not itself create that input capability.
Inspect the returned viewpoint
Compare the start, midpoint and finish with your camera brief. Check whether the person shrinks as the viewpoint retreats, whether the doorway remains attached to the wall, and whether foreground objects move relative to distant ones. Also watch the complete shot: three pleasing still frames can hide a sudden jump between them.
Accept the result when the viewpoint change communicates the intended reveal and the person performs the approved action. If the actor copies an unwanted gesture, inspect the input assignment before adding more prompt text. If the scene bends or reveals implausible architecture, shorten the move or provide better scene evidence. Diagnose the failed layer before spending another generation.