one take, four cameras

the problem

A talking video shot on one camera is one angle for ten minutes. Cutting between angles is what makes it feel like television instead of a webcam, and I did not have a second camera on me. I wanted to feed the machine my one angle and get dimension back.

what comes back

One ten-second take goes in. What comes out is the same take with real camera moves inside it: a close-up, a three-quarter angle, a slow push, a cut back to wide. My voice is my voice, my lips match it, and the outfit, the room and the colour carry through from the original. When the first ones came back I said, on the record, that they looked real. They do.

the three things that make it work

1. The voice is never regenerated. Most video models make up audio to match the picture. This method feeds my actual recording into the generation and locks it, so the picture is built to fit my voice instead of the other way round. Measured lip sync on the reference render: zero milliseconds off across the whole clip.
2. My words go in as dialogue, tied to a speaker. The model is told exactly what I say and who says it. Without that it invents speech and the mouth goes wrong.
3. Long, precise instructions with cuts on the clock. The prompt for one ten-second take runs to thousands of characters: what each shot frames, when it cuts, and where my eyes go, timed to the word I am saying. "Look at the floor" happens on the word floor. That is what makes it feel directed instead of generated.

what I stopped caring about

The cuts land a third of a second to half a second off the time you ask for. I approved two renders that missed every cut on the stopwatch, because shot quality and even pacing are what read as good, and nobody watches with a stopwatch. Optimise for the picture, not the timing.

one rule that saved a day

If lip sync matters, shoot against a plain background. My first test was in front of a wall of screens showing my own face, and every tool I used to measure sync locked onto the screens instead of me. On a plain room the same tool measured perfectly.

the part that changes everything

Instead of feeding the machine a still from my take, I fed it a character sheet: a different person, a different outfit, on a plain grey backdrop. The take still supplies the motion, the timing and my voice. The sheet supplies the person. Four sheets, four for four: a futa look with abs and freckles, a girl with glasses and devil horns, pink-and-black pigtails with a spiked choker, and a look with legible printed text on the waistband that stayed legible across every angle. It carried body type, not just clothes.

The one line that makes it hold: tell the model that nothing of the woman in the source video carries over except her movement, and that the person on screen is the one in the picture. Without that line you get a blend of the two people. Each swap takes about eight minutes on a rented GPU, and a sheet needs a plain backdrop, several angles including face close-ups, and its markers named in words, because what is named is what carries.

what it means

One shoot day of me talking to one camera is now raw material for a multi-angle edit and for any number of characters saying those words. That is the difference between a creator with a camera and a creator with a system.