Pair a still of your subject with a clip of the motion you want: a dance, a walk, a gesture. The model re-performs that motion with your character. One frame in, a directed shot out.
Two inputs: a still and a driving clip
Two files from your library: the subject (any still, whether a photo, an AI render or an illustration) and a driving video whose movement you want. The output is ordinary footage of your subject doing what the reference did.
Any still works: photos, AI renders, illustrations
The driving clip can be anything: stock, phone footage, another generation
The output is a normal clip, ready for the timeline
Direct the performance, skip the shoot
Change what the character does by changing the driving clip. The same subject can dance, walk or turn to camera from a reference you filmed on your phone, so one character stays consistent across shots that were never staged.
One subject, as many performances as you have reference clips
Record the motion yourself: a phone clip works as a driver
Chain with image generation: design the character, then move it
Questions
What inputs does motion transfer need?
A still image of your subject and a short driving video with the motion you want. Both live in your media library, uploaded or generated.
Is motion transfer the same as image-to-video?
No. Image-to-video generates motion from a prompt, while motion transfer copies motion you picked. Use it when the specific performance matters.
Does the subject need to resemble the reference actor?
No. The movement is mapped onto your subject's own body and framing, and the reference only supplies the performance.
How is motion transfer priced?
By the driving clip's duration, with the token price quoted before you run it. Short, clear performance clips work best.