Skip to Main Content

Real One

Built at Meta Global AI Developer Hackathon · Oct 26, 2026 · Remote

Real One — Demo video

Real One makes vertical music videos for Reels, TikTok and Shorts from one phone, for anyone performing to a song. The artist is never generated. Only the world is. I've directed music videos for ten years. Most of the artists I shoot for are independent, and the math is rough for them. One indie musician on Reddit paid up to £800 for visuals that "amounted to about 500 views" in a couple of years; a band in the same thread paid £1k+ a couple of times for videos that got "maybe 1-3k plays" [1]. The Source says an indie music video "may vary between $1,500 and $10,000" [2]. AI made videos bring their own problems. On Hacker News one commenter called an AI music video "far too much a literal direct copy of the lyrics" [3]. When fans suspect a video is AI, the reaction is brutal: "now it all just feels lazy and cheap" [4]. In the demo I perform "Real One", a song made for the demo: Muse Wonder wrote the lyrics, and the music and the vocal were generated with ElevenLabs Music, licensed for commercial use. The voice is not mine; I perform to it. You film yourself on a tripod, once or a few times, with the master playing out loud. Real One finds where each take sits in the song. SAM 3.1 cuts you out of every frame, Muse Image builds a world for every section, and the edit cuts on the downbeat and hands back a vertical video with the master audio in sync. With several takes, Muse Wonder watches all of them and picks the strongest one for each section, with a line on why. The lyrics land on screen word by word, as they are sung. On the last hook the world freezes mid air for two bars and only you keep moving. The last hook also carries the Real Receipt: every generated pixel turns grey, you stay in colour, and a counter reads how much of each frame was filmed. Over the demo cut, 37.0% of the pixels are filmed. After the render you can keep editing in plain words. "Put fire coming off my hands when I say heat" becomes a tracked effect: Muse Wonder names the part to follow and the word it lands on, SAM 3.1 segments that part on every frame, and the fire moves with your hands, landing on the sung word. Real One tracks the floor you stood on, solves where the phone was on every frame and rebuilds each world in depth, so the world moves with the phone (on a moving test plate the feet slip 0.4 px a frame, 6.3 on a still world). The demo is a tripod take, read as a locked camera. Built during the event: all of it. The repo's first commit is Oct 28, 2026, and nothing in it was written before the event. 388 tests pass on the 0.4.0 release, and I broke the guards on purpose to check their tests go red (logged in the CHANGELOG). [1] https://reddit.com/comments/1qdjcpp [2] https://thesource.com/2026/03/18/what-it-costs-for-an-indie-rapper-to-shoot-a-music-video-in-2026/ [3] https://news.ycombinator.com/item?id=48940066 [4] https://reddit.com/comments/1rr2z9a

Team