How to make a Tiny Desk AI video

Five seats, one portrait each, no prompt. Here is the whole process, including what to do when a face does not land.

Updated 2026-10-29

There is no timeline and no prompt box. The performance is fixed, so the only decision you make is who stands in each of the five seats. That is why the whole flow fits on one screen, including on a phone.

Step 1 — pick the seats you want to replace

The studio shows five seat cards, each with the face of the performer currently in that seat. Tap a card to add a photo. If you only want one person on stage, use the lead singer seat — it owns the opening shot and the close-up that carries the joke.

• 

  1.Lead singer

  Shots 1 and 3

• 

  2.Keyboard

  Shots 2, 4 and 5

• 

  3.Backing vocal · left

  Shots 2 and 5

• 

  4.Backing vocal · right

  Shots 2 and 5

• 

  5.Saxophone

  Shot 4

Not sure which seat is which? Who you can replace explains every seat with the shot it appears in.

Step 2 — choose photos that will survive the render

The model is doing identity replacement, not face-swapping two frames. It has to hold a face across head turns, a laugh and three camera angles. A sharp, front-facing portrait with even light is the single biggest quality lever you control.

• Front-facing, eyes open, whole head in frame with a little space above the hair.
• Even light. A window or a ring light beats a dim room or a hard side light.
• One person per photo. Group photos and screenshots of other videos perform badly.
• No sunglasses, no heavy filters, no hats pulled low over the eyes.
• At least 512 pixels on the short side. A blurry 100-pixel avatar will not hold up.

The full checklist lives in Tiny Desk AI photo requirements.

Step 3 — pick a resolution

Price moves in three cast tiers: 1 person, 2–3 people and 4–5 people. A five-person band costs more than a solo lead because every shot has to be rendered.

Step 4 — render, download, post

Rendering happens per shot. A solo lead only needs two shots, so it usually finishes faster than a full band that needs five. You get a 720×1280 vertical MP4 at 24fps with no watermark and the original audio bed.

If the platform mutes licensed music, use the silent copy and pick a track from the platform library. Label the post as AI-generated.

What to do when a face does not land

1. 1Check the photo first. If the face is small, side-on or shadowed, replace it.
2. 2Move important people into seats that get close-ups: the lead singer seat is the only one with a real close-up.
3. 3People at the edge of the wide shots are the hardest to replace. That is a limitation of the source footage, not something a retry will fix.
4. 4If the render is visibly broken, it counts as a failed render and the credits come back.

Costs, refunds and the exact credit rules are on is Tiny Desk AI free.