Wan 3.0 Reference to Video

Wan 3.0 reference to video takes up to twenty reference assets — ten images, five video clips and five audio tracks — and treats each one as a constraint rather than a starting point. Those faces, garments, props, voices and spatial relationships stay consistent throughout a single take of up to thirty seconds, with audio generated in the same pass.

Images · clips · audio

10 + 5 + 5

Images · clips · audio

Assets in one request

20

Assets in one request

Any whole second

2–30 s

Any whole second

30 fps, sound included

480P / 720P / 1080P

30 fps, sound included

• Not an editor — it builds a new scene rather than changing an old one
• Not a face swap, and not a way to put someone in a video without their agreement
• No first frame alongside references — those two families exclude each other

• reference_image

• first_frame

⚠️ Wan 3.0 bills reference video seconds alongside your output seconds, and the two together cannot exceed thirty. Six seconds of reference plus a twenty-second output is twenty-six billed seconds and a valid request; six plus twenty-eight is not. Images, audio, documents and links add nothing.

Five video slots at fifteen seconds each would be seventy-five seconds, and the ceiling is fifteen — so on video the slot count is never the binding constraint, the running total is. The counter above the prompt box tracks both, in seconds and in credits, before you press Generate.

Twenty assets is the headline number; the number you should actually use is much smaller. One subject per image, one job per asset — the moment one file carries two conflicting jobs, or one image contains two people, the model has to guess, and guessing is the thing you uploaded references to prevent. Most good scenes are done with three or four. Would rather not write the rest of the prompt? The prompt generator writes it in Alibaba's order.

1. 01The character does not look like the reference

   Almost always because the image contains more than one person, or because that same image was also asked to define the location or the style.

   One subject per image, one job per asset.

2. 02Two characters blur where they touch

   Overlapping positions plus simultaneous action is the hardest case. Validate on a shot with clear separation first, then add contact.

   Two well-separated characters are dependable; a crowd that touches is not.

3. 03The motion ignores the reference clip

   That asset was also asked to define a character, so it is carrying two jobs and the model has to choose one.

   Split it into two assets with one job each.

4. 04The bill is larger than expected

   Reference video seconds are billed at your output rate, and they count against the same thirty-second ceiling. Images, audio, documents and links are free.

   6s of reference + 20s of output = 26 billed seconds.

5. 05The request is rejected before generating

   A first frame was sent alongside references. The two families exclude each other, and the refusal is the whole request rather than one input.

   Order: Image 1 · Image 2 · Video 1 — no frame inputs.

Start simple to validate reference quality, then add complexity. A scene that fails with five characters often works with two, and two clean clips cut together beat one muddled one.