A video model can support reference audio in one workflow and reject it in another. Before preparing a first-and-last-frame request, check that specific combination. Alibaba’s Wan 3.0 guide excludes reference audio from its first-and-last-frame mode. This does not describe every Wan model or every audio-driven video route.
Name the actual production contract
Write down the model identifier, endpoint or interface mode, and required inputs. “Wan video” is too broad for a compatibility decision. Identify whether the task uses one starting image, two endpoint images, an editing video, or another reference arrangement. Then verify which audio role is offered: generated soundtrack, reference influence, speech driving or an uploaded track for later editing.
Those roles are not interchangeable. A reference recording may guide sound without being copied exactly. An externally assembled soundtrack can remain exact but does not automatically drive visible articulation. An endpoint image defines a visual boundary, not a guarantee that a spoken sentence will fit between the endpoints.
Decide what is non-negotiable
In a hypothetical anniversary clip, the couple’s opening and closing poses may be essential, while the approved music can be added later. In another project, an exact spoken line and synchronized mouth movement may matter more than two prescribed poses. These priorities lead to different production routes; trying to combine unsupported controls does not satisfy either requirement.
| Required result | Question before choosing a route |
|---|---|
| Approved visual endpoints | Does this exact mode accept both frames? |
| Exact recording preserved | Is it a real input track or only a reference? |
| Speech-driven mouth movement | Does the chosen route document that function? |
| Final combined film | Where will picture and audio be synchronized? |
Use the table as a decision record, not as evidence that a vendor can perform every row at once. Record an unsupported combination explicitly.
Separate a valid request from a good result
If visual endpoints take priority, prepare the supported silent or generated-audio request and plan the approved soundtrack in an appropriate editor, if suitable. Verify the duration and available pauses before assuming the recording fits. If speech driving takes priority, choose a route that supports the necessary audio input and review what visual controls are actually available there.
Do not rename an audio field, split the inputs across messages or use a different model’s example as proof of compatibility. A request being accepted can still lead to an ignored or transformed input. Preserve the actual request manifest and inspect the delivered file, including unexpected generated speech beneath the intended soundtrack.
Review the assembled deliverable
Check both endpoint appearance and the complete soundtrack after assembly. Listen for doubled music, leftover speech and abrupt endings. Watch for mouth movement that contradicts the approved recording. If the conflict cannot be resolved, revise the shot design or obtain approval for a voice-over approach.
An Imagild brief can state these priorities and contain the reviewed assets, but external mode compatibility remains a separate question. This article recommends a planning method; it does not report a successful combined-input experiment or promise exact endpoint and sound reproduction.