Creative AI

Madison

Open source

An AI agent assembles the edit. Madison shows the result in a browser: the rendered movie, every lane of the timeline, and the details of each clip.

Pick a moment or mark a range, write what should change, and copy the note back to the agent. It can also pull in the latest saved CapCut timeline.

Blender previs

14 cameras from one script

A script builds a gray box ranch set in Blender from a written spec, with 14 named cameras. Each camera renders one frame with labeled mannequins and one empty frame used as the background plate. A full run takes about 20 seconds.

Changing one value in the spec changes only that thing. A gray animatic decides where people stand, where the camera goes, and the timing. In one test the video model followed it shot for shot: every cut landed within a fifth of a second of the plan, with five people on their marks.

Character builds

Face cards to 3D heads

Each character starts as a set of locked face cards. Each face then becomes a 3D head mesh of about 90,000 points that can be turned and lit in Blender.

For instrument shots, a Blender mannequin plays along to a note map drawn from the song, and the video model paints the character over it.

3D sets in the browser

Explore in the browser

One approved picture of a bar, with the people removed, became an explorable 3D set using World Labs Marble. Misspelled signs were repainted and the set rebuilt.

In the browser the set assembles itself: every point starts scattered as a tiny square of color, then flies home in a wave. After that you fly through it with the keyboard and mouse.

Motion from sound

11 of 11 bow changes matched

It reads every note from a fiddle recording, then poses a 3D mannequin with the bow on the string and a finger on each note.

That motion study becomes the guide video for the AI render, so the bow changes direction when the music does. In the test clip the bow hand matched all 11 direction changes within about two frames. The earlier method matched 3.

Voice and render checks

Every render is downloaded, transcribed, and turned into a contact strip before anyone decides to keep it. A pitch measurement checks that each character’s voice stays in the same register from shot to shot.

This chart compares how wide the mouth opens with how loud the singing is, to find how far the lip sync is off. In this take the mouth ran about one and a half seconds early.

CapCut timeline tools

Scripts read a CapCut project and report the exact source and timeline range of every clip, plus gaps and tracks. New edits are built from a known good template, never by experimenting on the working project.

This chart is a speed curve for a final chorus, with each sung word marked, so the picture keeps time with the vocal.

Character continuity

Each character has a locked set of reference cards: close angles, full body, and on stage. No shot is made without them attached.

Continuity is settled in still pictures first, where tests are cheap, before paying for video. These are Alder’s cards.