AI video prompts tested: what changed the result and what did not
Short answer: most of the prompt wording we expected to matter barely moved the result. What did matter was the words the person actually says, and the first picture the video starts from. We ran four small tests on Wan 3.0 for about $3.70 in total, and this is what we found, including the things that failed.
Why we ran the test
Some of our people videos looked flat. When we read back the prompts we were sending, the cause was ours: every clip carried instructions like "small relaxed expressions", "keep every action slow and continuous" and "the camera pushes in slowly". Three separate brakes on energy, and almost no direction pushing the other way. So we rewrote the prompt and tested it properly instead of trusting our own opinion.
Each comparison kept the actor, script and settings fixed at 480p, so the only thing that changed was the prompt. Each clip was scored out of 10 by an AI reviewer that did not know which prompt was which, for energy, expression, eye line, accent and lip sync.
Round 1: better wording, small gain
We compared the old prompt with a rewritten one on a talking-head clip.
- Energy stayed at 8 out of 10 in both.
- Expression range went from 7 to 8 with the new prompt.
- Lip sync went from 9 to 8, so no free lunch.
The new wording was no worse and a little more expressive. That is a modest gain for a full rewrite, which is itself a useful finding: prompt polish alone will not rescue a flat clip.
What did not work
Describing the accent
We asked for a Manchester accent in the prompt, right next to the spoken line. All three test clips came out London. Wan's own guide says accents are only loosely followed, and our result matched that.
Giving it a voice sample
We then made a short Manchester voice sample and passed it in as an audio reference. It was confirmed as Mancunian by an automated listener. The clip still came out with a London accent. Wan did not take the accent from the audio.
Energy caps in the prompt
We believe words like "relaxed" and "slow" act as caps, though in our first round the energy score did not change. If you want a livelier performance, describe the movement you want, such as a sequence of expression changes, instead of the calm you are trying to avoid.
Looking away from the camera
For a podcast host we wrote that his eyes stay on the guest, off to the right. He looked straight down the lens anyway. We then tried passing a picture of him turned to the guest as an extra reference. Still the lens. Our best explanation is that Wan treated that picture as "what he looks like", not as the opening frame, and every other reference image showed him facing the camera, so the pictures beat the words.
What did work
1. Dialect words in the script
When we wrote the line itself in Manchester phrasing ("Right ... proper unreal, our kid ... I'd do it again, me"), the accent was heard as Manchester, with northern vocabulary coming through. Energy was 8 and lip sync 8.
The lesson is that the accent lives in the words, not in the description of the accent. Keep it light and natural: "proper" and "our kid" for Manchester, never a caricature.
2. One generated opening picture as the only reference
For the podcast host, we generated a picture of him already looking at his guest and made that the only reference. The eyes stayed on the guest throughout the clip, the sound was clean, lip sync scored 10 out of 10 and the face matched. The picture cost about $0.15 to generate.
We also tried Wan's image-to-video start-frame mode. The eye line worked there too, but the audio came back about 12 dB too loud and clipped, so we did not keep it.
3. Naming references "Image n"
Wan 3.0's documented way to refer to uploaded files is "Image 1" or "Video 1", numbered in upload order. Some of our references had used an "@image1" style that does not appear in the documentation. Give each reference one job as well: "Image 1: identity only", "Image 3: location only", "Video 1: motion only".
The prompt habits we now follow
- Most important first: who, what they do, where, camera, light, voice, words, ending.
- Say each constraint once.
- Describe the state you want, not the thing you fear. Wan has no negative prompt field in the version we used, so "never look at the camera" tends to produce exactly that.
- One camera move and one main action per clip, 4 to 8 seconds for people, with short lines. Wan says lip sync to exact words is not reliable.
- Name the background sound you want. "No background music." removes music but keeps other sound, and if you name nothing Wan picks something.
Verdict
If your AI video looks flat or off, check three things before rewriting your prompt again: the words the person says, the picture the clip starts from, and whether any line in your prompt tells them to be calm. In our tests those moved the result far more than clever phrasing. All of this cost about $3.70, so it is cheap to check for yourself.
FAQ
Can I get a regional accent in AI video?
Not by asking for one. In our tests, a written accent description and a voice sample both failed. Writing the spoken line in light regional phrasing worked.
How do I make the person look at someone off-screen?
Start from a picture where they already do. As a reference image alongside others it was ignored. As the only reference, it held the eye line for the whole clip in our test, though that is not guaranteed.
Does prompt expansion help?
In our test, no clear gain or loss on an actor clip, and we could not see what the rewrite changed because the rewritten prompt was not returned. We left it off in our tests.