The interesting part of this video is not the AI face. It is that I can now take any viral reel and get its exact skeleton in 1.6 seconds.
Nine shots: 2.00s, 0.50s, 0.87s, 1.53s, 1.50s, 1.00s, 1.50s, 4.50s, 2.67s.
That rhythm is the whole reason the original worked. A two second hook, seven fast cuts that never let you settle, then one long shot near the end where it finally shows proof. The content of a viral is replaceable. The architecture is not.
So I wrote the tool that pulls the skeleton out, and filled the slots with my own footage. 1.6 seconds to decode the reel, 9.4 seconds to render the finished twenty second video. I never needed anything from the original except its timings.
The face was the slow part. Video models are bad at holding a face. LivePortrait kept my identity but rippled the monitor behind me. Kling Video Edit gave me my moustache on someone else's face. A face swap painted my moustache onto my own hand, because the hand was covering the mouth and it found a face there anyway.
Seven models, five of them wrong, one afternoon.
What worked was inverting it. Image models hold a face, so I generated the face first with Nano Banana Pro from three photos of me, picked one of four variants, and let Kling v3 animate it. Two generations. About 460 credits a clip once the settings are known.
The first two seconds of this video are synthetic. The screen recording and the last shot are really me.
If you can spot the seam, say so. That was the test.