How to Make an Outstanding Tiny Desk AI Video

Updated 2026-10-28

This is a plain-English summary of a public technical write-up (Hugging Face blog) by another tool maker. Our own generator is launching soon.

1. Get the vertical clip. The viral version is a vertical, roughly 19-second cut (720x1280, 24 fps) with five shots. Working from that vertical version keeps the framing right; widescreen footage tends to get re-framed by the model.
2. Pick one clear photo per person. Use a sharp, front-facing photo of one person. Only use photos of yourself or people who agreed. A face close-up crop plus the full photo helps the model keep both the face and the clothes.
3. Use a video edit mode, not text-to-video. A video-edit model (for example ByteDance's Seedance 2.0 mini in edit mode) takes the clip as reference and keeps its motion, camera and timing while swapping the person. A short prompt that says to keep everything else the same works best; one public write-up found that mentioning the mouth or lip sync made results worse.
4. Split at the 13-second hard cut. These models cap a single generation at about 15 seconds, so split the clip at the hard cut at 13.0 seconds: one 13-second part and one ~6-second part. A seam on a hard cut is invisible.
5. One face per role. Assign each person to a single role. Mapping the same face to two roles caused the swap to fail in published tests.
6. Check the faces before sharing. A failed swap can still come back marked as successful. Look through the output (or run an automatic face check) to make sure the original people are gone.
7. Join the parts and add the song back. The model returns silent video. Join the two parts and lay the original audio underneath; keep a silent copy for apps that mute music. Label the post as AI.

Shot list of the viral clip

Timings from the same write-up.