Put the person from a still into a driving video. No pose estimation, no segmentation, no masks — the only preprocessing is repainting the clip's first frame. 4 sampling steps. Takes 4–10 s of driving video; anything longer is cut to 10.1 s. If the clip carries a soundtrack it is pinned into the render and the mouth tracks it; upload a silent clip and the mouth moves without saying anything.
Qwen edits the driving video's first frame using your prompt; the reference image is optional. If you provide an already-edited frame, Qwen is skipped. SeedVR2 runs as a local CLI in this Space; its model/runtime dependencies still need to be compatible.
examples — click one for its driving clip, its painted first frame and the render
driving video — motion, framing and background are kept; its soundtrack, if it has one, is what the mouth follows
What should Qwen change?
optional: already-edited first frame (skips Qwen)
step 2: render
What it produces
Every example above, rendered at 4 steps with seed 42. Each is a single paint and a single render; the first and the last three swap in a person. Only the first one has audio — it is the only one whose driving clip had any.