Wan 3.0's omni-reference system is powerful but has rules. This guide shows how to combine each input type correctly, what the limits are, and which combinations produce the best results.
Start with reference images when identity matters. Up to 10 images can ride along with a request: a subject portrait, product shots from several angles, and a style reference for palette and lighting. In the prompt, reference them naturally — "the person in the second image walks into the cafe" — and let the images handle appearance.
Reference videos transfer motion. Up to 5 clips teach Wan 3.0 a dance, a walk cycle, a fabric flow or a camera pattern. One rule to remember: total reference video duration plus output duration must stay within 30 seconds, so trim references to the essential motion.
Reference audio guides voice and mood. Attach up to 5 clips — a voice sample for speaking characters, ambient recordings for atmosphere. Audio should be paired with at least one image or video; audio alone is not a valid request.
Documents and links are the quiet superpower. Attach one file — a product spec, a script outline — or one webpage link, and Wan 3.0 reads it as context. Point it at a product page and generate a commercial that respects the actual feature list, without copying details into the prompt by hand.
The mutual-exclusion rule: first and last frame inputs cannot be combined with reference materials in the same request. Pick one workflow per generation — storyboard control with frames, or identity control with references.
Everything is available in the Wan 3.0 studio: upload materials, write the prompt and watch the live cost estimate update as you go.