Wan 3.0 Prime reference-to-video holds a face, wardrobe, location or voice across a clip using up to 10 images, 5 videos and 5 audio files. It returns 2 to 30 seconds at 480p to 1080p, with reference video and audio each capped at 15 seconds.
Workbench
These examples were written for this endpoint. Replace the prompt and sample media URLs with your own accessible inputs, then check the current quote and model availability before submitting.
- Output5s
motion-style-copy The reference clip is a bottle turning under a spotlight, a different subject — the point: `reference_video_urls` carries only the camera's handheld orbit. `reference_image_urls` supplies the woman's face and dress in an empty studio the prompt never describes. Over `duration_seconds` `5` the frame drifts flat-on to three-quarter, keeping the source clip's unsteady feel.
- 720p
- 5s
- Output2s
same-anchor-same-voice Three stills — a presenter at a studio desk and two close-ups — hold the face and the set across shots; this run sent images only, with sound off, so nothing here pins the voice. Reference audio is how you would: the endpoint takes up to ten stills, five clips and five audio tracks together, and the voice is the part that is easy to forget — a presenter who looks continuous but sounds different between takes is still discontinuous. The prompt then only has to describe the move, a push that ends on a specific framing, because identity is already carried by the references.
- 480p
- 3:4
- 2s
- Output5s
moon-dune-scarf One identity plate shot flat on a plain studio background, and everything else in the scene — the dune field, the moon, the wind — written from nothing. The hard part is the scarf: several metres of it, streaming and coiling above her, and still attached to the same arm at the end of the take.
- 720p
- 16:9
- 5s
This example starts at $0.061 per output second. The Playground updates the estimate when you change the settings.