Wan 3.0 takes reference material with one request. The published caps are:
The thing worth noticing is the word and. These are three separate allowances, not three slices of one pool. You cannot trade unused image slots for a sixth audio clip, and you cannot spend a shorter video budget on a longer audio one.
Video and audio each carry two limits — a count and a duration — and they bind independently:
So the planning question is not "how much reference do I have" but "which of the two ceilings am I closer to". If your material is a handful of long takes, duration binds first and you should be trimming. If it is many short beats, count binds first and you should be concatenating before you upload.
That second case is the one people miss. Six one-second clips is 6 seconds of material — comfortably inside the duration budget — and it still fails. Joining them into two three-second clips fits, carries the same footage, and costs nothing.
The caps are the easy part. The harder constraint is that reference material and first/last frames cannot be combined. A request carrying both is rejected before it generates.
This is not a soft preference or a "results may vary" warning. It is a hard branch in your integration, and it means the UI decision is made for you: reference-driven and endpoint-driven generation are two modes, and your interface has to pick one per request rather than letting a user fill in both and discovering the conflict at submit time.
The good news is that the rejection happens before generation, so a mistake here costs a round trip rather than a clip.
The last piece: there is no identity or session state. Reference material is per request, and nothing is remembered between them. If you want the same face in shot 4 that you had in shot 1, you re-send the reference in shot 4's request — every time, at full budget.
Which means the three budgets are not a per-project allowance you draw down. They are a per-request envelope you refill on every call, and the cost of consistency across a sequence is re-uploading the same material for each shot.
The reference caps, the mutual exclusivity with first/last frames, and the absence of cross-request state are published parameters, not inferences from behaviour. They are listed with citations and check dates on the Wan 3.0 specification table.
Disclosure: this site is written by the people who run a third-party browser interface to this model. We are not affiliated with Alibaba or Alibaba Cloud.