Runs MediaPipe's hand landmark network on its own — raw TensorFlow.js, no detector wrapper — so you can watch its 45 internal layers light up, read live kinematic features off the hand, and compare a hand-crafted pose subspace against the network's own learned one.
The 256×256 wrist-aligned crop the network actually sees.
Mean activation per stage — energy flowing through the pyramid.
Each tile is one channel. Hover a tile for its statistics.
Kinematics read straight off the landmarks every frame. Distances are in palm-lengths (wrist→middle knuckle), so they hold steady as you move toward or away from the camera.
Capture a handful of distinct hand shapes, then compare the subspace you get from hand-crafted geometry against the one from the network's own 768-d penultimate layer.