Two interactive 3D views of one song, Slow Down, as YuE2, an open-weight AI music model, makes it: the sound resolving from noise, and the song in four forms from plan to sound, on one time axis. The pages show the saved states of one real generation; the plan-to-sound replay is an illustration. Audio previews are peak-normalized and compressed for the web. Drag to orbit, scroll to zoom, and every view plays the audio it shows.
The song drawn as a landscape of sound at each of the 32 steps the model takes from random noise to the finished take. Move through the steps while it plays, and hear a quick estimate of where it is heading.
One song in four forms, read top to bottom: the score it plans, a rough draft of the sound, the detailed sound, and the audio. Light one moment through all of them, hear the phrase, read its lyric, and watch a sped-up illustration of the stages.