Photo segmentation v1 — Mask
Click a photo and the subject lifts off it: each click drops a focus point, a segmentation model masks it, and the cut-out floats sharp above a blurred copy of its own scene.
A koala lifts clean off its own photograph and hangs pin-sharp over a blurred full-bleed copy of the same scene, an optional drop shadow selling the float. Clicks steer the cut: each click on the picture drops a focus point that masks another zone, the subject on top is the union of every zone, and clicking a marker's circle releases that zone again.
Each focus point runs through MediaPipe's interactive segmenter — point at a pixel, get back a mask of whatever object owns it — one inference at a time, cached per point. The binary union is then blurred and its alpha remapped through a smoothstep whose centre pushes the edge in or out and whose width sets the feather, so the rim can be grown, shrunk or softened without re-running the model; the backdrop swaps between the untouched photo, a blur, a dim veil or a flat colour.
