Visualizer Stream began with a practical live-performance question: how can a visual system move beyond a shuffled folder of videos and choose material that belongs with the music playing now? I built it as both a media-intelligence pipeline and a real-time playback system.

The analysis layer samples video into windows and candidate segments, measuring palette, brightness, saturation, motion, intensity, cut behavior, and boundary quality. A visual-rhythm stage combines motion onsets, scene changes, and flashes to estimate tempo and confidence. The catalog can therefore describe not just what a file is, but how portions of it feel and move.

Live music state supplies color, energy, tempo, and optional mood guidance. The matcher ranks analyzed candidates against that intent while accounting for quality, duration, audience policy, and recent selections. Explicit queue items still take priority, but autoplay can make informed choices when the operator hands control back to the system.

Playback sends a continuous H.264 RTP stream to render nodes while recording enough lifecycle and timing information to explain transitions. Show mode suspends expensive background work without suspending the performance. Together, those pieces turn a media library into a responsive visual instrument.