Borrowing camera movement from a reference clip

Text prompts are a clumsy way to describe camera movement. "Slow dolly in with a slight arc to the left" is understood differently by every model, and the result is often a static shot with a zoom. Reference-video conditioning solves this by showing instead of telling.

The idea is straightforward. You supply a short sample clip whose motion you like: a handheld walk-through, a crane rise, an orbit around an object. The model extracts how the camera moves through that clip and applies the same movement to a new subject described in the prompt or given as images. Subject and motion are separated, so you can keep the motion and change everything else.

Practical observations from testing:

• Short, clean references work best. A three-second orbit transfers more reliably than a fifteen-second sequence with cuts.
• Combine the reference with an image of the new subject. Text alone leaves too much open; an image anchors what is being filmed.
• Use first and last frames when the ending matters, so the borrowed motion lands on a defined composition.

I ran these tests in Wan 3.0 (free AI video generator), which exposes this capability on the Wan 3.0 model, together with up to three reference images, frames, audio and document inputs. Outputs reach 30 seconds and keep characters and products consistent between shots. The tool states the credit price before each render and refunds failures, useful when iterating on many small motion tests, and it can be started without a card.

For anyone building video features, reference motion is the most underrated control surface in current models.