The trend
The Tiny Desk AI prompt, seat by seat
If you want to build this yourself with an AI video model, the hard part is not the scene description. It is telling the model which face belongs to which performer, and keeping that mapping across cuts.
Why seat mapping beats “put these people in”
A model that receives five portraits and a vague instruction will happily blend them, swap them between shots, or drop one entirely. Name each seat, name a stable visual cue, and repeat that the mapping holds across cuts.
The prompt we use, shortened
Photorealistic live band performance, vertical 9:16, one continuous take. Scene: a cramped office desk turned into a tiny concert stage — bookshelves and record sleeves behind the band, two warm desk lamps, a vintage microphone on a boom stand, headphones, a small combo amp. Cast: @Image1 replaces the lead vocalist at the centre microphone. @Image2 replaces the saxophone player to the left of the desk. @Image3 replaces the keys player, seated. @Image4 and @Image5 replace the two backing singers behind the desk. Keep each identity distinct across every shot: same face, hair, skin tone, glasses, clothing and build. Never blend, morph, duplicate or swap two people. Shots: 0:00–0:04 wide behind the desk; 0:04–0:10 push in on the lead singing and laughing; 0:10–0:13 the sax solo; 0:13–0:15 back to the wide shot, waving. Audio: an original mid-tempo 1980s funk groove with wordless vocals, backing harmonies and a short sax solo. Do not reproduce any existing song, lyric or recording. Rules: everyone is an adult; no on-screen text, logos or watermarks; no fade to black; fill the whole clip.
What a prompt cannot guarantee
No prompt makes lip sync exact, and no portrait supplies a voice. If you need a specific recording under the clip, that is an editing job with a licensed file, and it will not make the lips match every syllable. Be honest about that in whatever you publish.
The part people skip
Unfilled seats. If your reference scene has five performers and you only supply two portraits, say explicitly that the other three stay as the original band — or that they must be replaced by invented musicians. Otherwise the model guesses, and usually guesses wrong.
Read next: how this generator applies the same seat map · what the format is.