Hand tracking v11 — Hand cutout
The footage plays untouched except where hands move — their patch of video is cut out and stacked into offset, tinted echoes that trail the gesture.
The video-atlas idea, but the cutout is your HANDS instead of a full-body selfie segmentation. The full source video plays; MediaPipe tracks the hands in it; the footage inside the hand region is copied out and the recent cutouts are stacked as offset, depth-tinted, fading echo layers — so the hands repeat and trail away while the rest of the picture stays put.
Like video-atlas it keeps a RING of recent cutout frames and redraws them as a stack: newest on top (aligned with the live hand), older ones pushed further along the offset, scaled, twisted, colour-ramped and faded. The difference is the mask — a soft brush over the tracked finger/hand chains, not a body silhouette — and that the live video shows underneath.
Source + tracking come from the shared interaction layer, so the SAME Vision → Source options as splines-v1-interactive choose what inference (and the picture) run on: webcam by default, or a recorded video / image asset.
