This prompt does something the other five do not: it starts with a pronoun. Not "a young man with a camera" — just "he". That is only possible because a first frame is attached, and it is the clearest demonstration in the library of what image-to-video actually changes about how you write.
The saving is bigger than it looks. A text-to-video version of this shot would need a paragraph before anything could happen: age, build, clothing, the camera he is holding, the square, the time of day, the light. All of it exists in the reference frame already, in more detail and with no risk of disagreeing with itself. Deleting it does not just shorten the prompt — it removes every place where the words and the picture could conflict, and conflict is where image-to-video runs go soft.
The second technique is the double. The prompt needs two instances of one face, which is normally the hardest thing to ask a generative model for, and it gets there by never describing the second one: "someone standing there who appears to be himself". The figure is defined by its relationship to a face the model can already see. There is nothing for it to average, because there is only one description in play. Contrast this with what happens if you write out the double's appearance — now there are two descriptions of what is meant to be one person, and the model treats them as two people who happen to be similar.
The pacing is worth stealing wholesale. Four verbs go by at a calm speed — lowers, looks, raises, slowly turns — and then one adverb changes register: "suddenly noticing". A five-second clip has room for exactly one beat, and this is how you spend it. The calm verbs are not filler; they are what makes the sudden one legible as sudden. Remove them and the realisation has nothing to be a change from.
Then there is the line this template exists to argue with. "The camera alternates between his viewpoint and a slow push toward his stunned face" is asking for two distinct coverage patterns in five seconds. Each needs a moment to establish before it means anything, and there are not two such moments available. The honest advice is to choose, and to choose the push — the point of view shot shows a tower, the push shows a reaction, and the reaction is the story. Keep the alternation only when you raise the duration.
This ran on Prime, and it is worth saying again what that did and did not buy. Prime is the high-speed model. Alibaba publishes the same output specification for both tiers — the same resolutions, the same two-to-thirty-second range, thirty frames per second — and prices Prime at about half again per second. It returned faster. It did not return a better picture, and nothing about this prompt needed it.
What is missing is the sound, and here it is more of a loss than on the other templates. The beat is a realisation, and the standard film grammar for a realisation is that the world goes quiet. Wan 3.0 generates audio in the same pass as the picture, so a sentence like "square ambience that drops away as he turns, no music" would have scheduled the picture beat as well as the sound one. Leaving it unwritten means the model picked, and a busy square that stays busy is the version that reads as nothing happening.
视频