Photo segmentation v2 — Noise reveal
A blurred photograph in which regions snap into focus one after another: a seeded noise walk clicks around the frame and a segmentation model cuts out whatever it lands on.
Everything is soft except where the machine last looked: one region of the photograph — a band of building, a patch of water — stands in full focus over a heavily blurred copy of itself, a small circle marking the current point of interest. The sharp patch keeps moving; several times per loop the focus jumps somewhere new and a different object snaps into clarity.
Points of interest are hard to define algorithmically, so the sketch embraces the opposite: a seeded noise walk clicks around the picture on its own, and each pick goes to MediaPipe's interactive segmenter, which returns a mask of whatever sits under that point. The walk samples its noise field on a circle so it closes over the animation loop — same seed, same picks, deterministic capture — and the previous cut-out stays up until the next inference lands, so one revelation always replaces another with no blank frame in between. A manual click overrides the walk until its next pick fires.
