Short answer
The studio uses a fixed prompt, split into two 15-second halves. The prompt describes seven shots, names who sings in each, keeps one black handheld mic, forbids extra performers and franchise names, and never asks the model for the song. If you are building your own version, copy that structure rather than writing “make a Star Wars Gangsta’s Paradise video.”
Make my videoWhat changed
- — Launched the two-photo, 30-second 4:3 Gangsta’s Paradise AI generator.
Gangsta’s Paradise AI prompt guide
You do not write a prompt in the studio. If you are rebuilding the format yourself, this is the shot language that keeps it from turning into a generic music video.
Published 2026-11-09
A free-form prompt is the fastest way to lose the meme. The model will happily produce a stage, a crowd, a different camera move or a third performer. The format only reads as the trend when the set, the cuts, the mic and the two-hander are locked.
The seven shots
| Window | Camera | Performance |
|---|---|---|
| 0.0–5.8s | Medium-close to full body, Rapper centred | Rapper performs; Singer absent |
| 5.8–11.8s | Three-quarter wide, then push to face | Singer walks in/out; Rapper wipes brow |
| 11.8–15.0s | Wide, both walk toward camera | Rapper performs; Singer silent |
| 15.0–18.1s | Medium-close, Rapper centred | Rapper points at camera |
| 18.1–23.1s | Wide down the hangar, push to raised arm | Rapper walks, waves, points up |
| 23.1–24.2s | Close-up, Singer behind | Rapper performs; Singer silent |
| 24.2–30.0s | Wide choir, then foreground two-shot | Troopers raise mics; Singer enters ~27s |
Who owns the mouth
The single most important line in the prompt is not a camera move; it is the mouth ownership. The Rapper performs in shots 1–6. The Singer’s mouth stays closed while walking, standing and watching, and only opens for the final chorus. If you do not state that, both people mouth the entire song and the contrast collapses.
Do not write “lip sync.” The model does not receive the song, so there is nothing to sync to. Describe the performance in plain language: “Rapper performs the verse with clear mouth movement,” “Singer walks in silence with mouth closed.” The audio is added later, and the mouth follows the described action, not the recording.
The negative constraints
- One black handheld microphone, one hand, five fingers. No second mic until the trooper choir.
- No third performer, no crowd, no dancers, no band.
- Do not name the franchise, the characters or the actors. Say “white-armored troopers” and “white sci-fi hangar.”
- Do not ask for the song, the artist or a soundtrack. The model cannot clear it and may refuse.
- No text, captions, logos or watermarks.
- Do not keep the original performers’ faces, hair, helmets or robes.
Why two 15-second halves
Wan accepts at most 15 seconds of reference video per call and a reference-plus-output total of 30 seconds. A 30-second reference cannot be sent as one piece. The film is therefore generated as 0–15 and 15–30, each against its own reference half, then stitched on the 15.0-second cut. The prompt for each half only describes the shots inside that half; it does not repeat the whole film.
A prompt skeleton you can adapt
Start with the set and the frame: 4:3, 24 fps, photorealistic, white hangar, glossy floor, starfield windows, rows of white-armored troopers, cool key light, one black handheld mic. Then lock the identity: faces, hair, skin, body and clothing come from the two photos; replace the original performers; do not keep their robes. Then write the shots for this half with exact time windows and hard cuts. Then state who sings and who stays silent. Then list the negatives. Finish with “silent video; the song is added later.”
The same structure is used in the template breakdown, and the practical upload flow is in how to make the video.