Reference mode is the part of Wan 3.0 that has no equivalent in most of the models it competes with, and it is also the part people use least well. This template exists to show the syntax in use, because neither of the two published examples in this library uses it at all.
The mechanism is simple and easy to get wrong. Attached materials are an ordered array, and the prompt refers to them positionally: `Image 1`, `Video 2`, `Audio 1`. Capitalised, with a space, numbered from one by position in the array. `image1` binds to nothing. `Image_1` binds to nothing. "The first reference picture" binds to nothing. And the important part: none of those produce an error. The request returns 200 and the model, having found no binding, invents a subject. You get a finished clip of the wrong person.
The second thing this prompt demonstrates is that assigning a reference and locking a reference are two different instructions. "Use Image 1 as the performer" tells the model whose face this is. It does not tell the model that the white hooded jacket, the ruffled skirt and the platform shoes are also part of what it is being given. In practice a single citation reliably preserves identity and unreliably preserves wardrobe, which for a fashion frame is the entire subject. The explicit second sentence — keep the face, the hair and the whole outfit exactly as they appear — is what closes that gap.
The third thing is the stillness, and it is the least obvious. Video models animate. That is what they are for, and given five seconds and a seated figure they will produce five seconds of small movement: a blink, a breath, a hand adjusting, hair settling. Every one of those is a moment where the reference can drift. So a reference portrait wants an explicit "does not move", and it wants exactly one named exception so that the clip is not a still with grain on it. Here that exception is a butterfly landing on an already-outstretched arm — soft, slow, and involving no change of pose.
It is worth naming what this shot gives up. There is no story, no reversal, no camera work to speak of. Five seconds and one butterfly. The trade is deliberate: reference mode is at its most reliable when the subject is asked to persist rather than to perform, and a great deal of commercial work — lookbooks, product-in-hand, character sheets, talking-head intros — is exactly that shape. Understanding that reference mode rewards restraint is more valuable than any individual clause in this prompt.
On the grade: "a heavily saturated grade that pushes the grass and the sky toward the same green" is the most transferable sentence here. It describes the look as an operation on two named elements rather than as a style word. Style words are ambiguous — "dreamy", "editorial", "cinematic" all mean six things — whereas an instruction that says which two things should end up the same colour has exactly one reading. Whenever a look is hard to name, try describing what it does to two objects in the frame instead.
The honest caveat, same as everywhere in this library that the label says reference render: the clip was published without a prompt and this text was written by reading it. What it demonstrates about citation syntax, wardrobe locking and stated stillness is true of Wan 3.0 regardless. What it will not do is reproduce this particular woman, because we do not have the image that was attached.
動画