How Do You Make a Tiny Desk AI Video?

To make a Tiny Desk AI video, the fastest route is a photo generator: upload a photo of the lead singer, add up to four friends, and about 6 minutes later you get the last 15 seconds of the clip with your faces in it. To match all 19 seconds cut for cut yourself, split the clip into two Seedance jobs at the 6.7 s hard cut and send one chest-up photo per person plus one photo of the empty room.

The meme recasts the end of Charlie Wilson's 2023 NPR Tiny Desk performance of "Outstanding". Background on the clip is in the Outstanding Tiny Desk meme.

Which route should you pick?

Both routes start from the same photos. The difference is who handles the shots, the cuts and the sound.

The generator uses the last 15 seconds because that is the longest motion reference Seedance 2.0 takes in one job. You lose the lead's side-on opening and the start of the singers, and keep the other four: the backing singers, the lead's close-up with the big laugh, the sax solo and the singers again.

Route 1: make it from photos

Open Tiny Desk AI and fill the seats. Only the lead is required. Every seat you leave empty is played by one of the site's AI band members, at no extra cost.

Then pick the model, the quality and the sound. Original sound puts "Outstanding" under the video; No sound leaves it silent for the trend sound on TikTok. The video is 15 seconds, vertical 9:16, with no watermark.

Pay with credits you have or buy a one-time pack (all credit packs). Rendering takes about 6 minutes. You can close the tab: the link is emailed to you and the video waits in My videos. If a render fails, its credits come back on their own.

Route 2: match all 19 seconds with Seedance 2.0

The full clip is five shots. Seedance 2.0 accepts a motion reference of 2 to 15 seconds and at most 9 reference images, so 19 seconds has to be two jobs.

Split at 6.7 s. It is already a hard cut, so nobody can see where the two parts meet. Pad each motion clip to a whole second (7 s and 13 s) and trim the output back afterwards.

Skip the per-shot keyframes. When one still per shot went in with the second job, the sax still was framed front-on while the motion reference had him in profile. The model played that still as an extra shot of about 1.8 s, and the lead's laugh shrank from 6.3 s to 4.5 s. Plain-background photos of each person plus one photo of the empty room fixed it: in a 5 s test across the laugh and the cut to the sax, the cut landed 1 frame from the reference with nothing inserted.

On cost, Seedance 2.0 Fast at 480p bills the input video seconds and the output seconds. 7 s plus 13 s, in and out, is 40 billed seconds: 272 credits on one API reseller, a bit under $1.40. Redoing one bad part costs about a third of that, so test the riskiest shot alone at 5 s before running the full jobs.

Which photos keep the likeness?

The same rule holds for both routes: chest-up or waist-up, one person, face to the camera. None of the five shots shows anyone below the waist, and in a full-body photo the face takes up so little of the frame that the likeness drops.

With five people you can't send two photos each in the DIY route anyway. Five headshots, two full-body shots for the lead and the sax player, and the empty room already fill 8 of the 9 slots.

Avoid blurry, dark or heavily filtered photos, group photos and pets (the generator does people only for now). Photos of children or public figures aren't allowed. Before you post, label the video as AI-generated where the app asks. ViralShort is an independent tool, not affiliated with NPR or the artists.

For the template route, see the Outstanding Tiny Desk AI template; for free options, Tiny Desk AI free.

FAQ