arisris

muse-render turns a song and a sentence into a procedural music video you can keep editing. Most AI video tools return an opaque clip: timing is guessed, changes mean re-rolling the prompt, and creators cannot touch the result. muse-render instead gives the model a codebase. You upload a track and describe a world. The agent listens to the entire song, identifies its sections and energy arc, then writes a gapless timeline of three.js scenes that react to live audio features (beat, FFT, time). The browser preview is the output — edits appear instantly, with no offline render step. Creators can refine the result conversationally ("make the stars pulse on the kick", "calmer verse, brighter chorus") or attach reference images. The agent reads and patches its own scene files, validates the timeline, and reloads the preview. Because the output is plain ESM + three.js, every result stays inspectable, versionable, and editable by a human. Built with Bun, TypeScript, three.js, and Modal; the whole pipeline is tested (54 tests) and runs end-to-end today.