Real One

How it's made, replaying my real run from Nov 4. Today it renders on a Mac from the command line.

Drop your takes.

Film yourself performing to the song, once or a few times, with the master playing out loud. Real One listens to each take and finds where it sits on the song's clock.

ALL THREE TAKES AT THE SAME MOMENT OF THE SONG, 9.5 S

Your song, in sections.

Say where each section starts and Real One snaps it to the bar. Every cut, world and lyric hangs off this clock.

THE BAR GRID. CUTS LAND ON THE DOWNBEAT; A NEW SECTION GETS A NEW WORLD.

The director picks a take.

Muse Wonder watches every take on the song's clock and picks the strongest one for each section, with a line on why. One pick here was overridden after a blind critic.

A world for every section.

Muse Image builds each world from the director's prompt, lit like your take. Then each world is rebuilt in depth on the floor you stood on, so you stand in it instead of on top of it.

Every word on your voice.

Demucs pulls the vocal out of the beat, and MMS forced alignment places each word of the lyric sheet on it. Every lyric card lands when the word is sung.

Edit it by saying it.

Type what to add. Muse Wonder names the part of you to track and the word it lands on. The time comes from your word timings, never from the model. SAM 3.1 then follows that part on every frame.

OPEN THE EDITOR
THE CUT AT "HEAT"

Your video.

A vertical music video for Reels, TikTok and Shorts, with the master audio in sync. On the watch page, press R for the Real Receipt: every generated pixel turns grey, you stay in colour, and a counter reads how much of each frame was filmed.

WATCH THE VIDEO
RUN IT YOURSELF

On a Mac with Python and a Meta Model API key, from the repo:


    
FINISHED
REAL RECEIPT · YOU IN COLOUR