How it works
Four steps, all running inside your browser tab.
Detect faces
Each photo is passed through SCRFD, an efficient face detector that returns a bounding box and five keypoints (eyes, nose, mouth corners) for every face it finds.
Align
A similarity transform warps each face to a canonical 112×112 pixel crop. Face-recognition networks are sensitive to pose, so this step matters.
Embed
The aligned crop runs through MobileFaceNet (trained on WebFace600K). Output: a 512-dimensional identity vector.
Compare & contextualize
Cosine similarity between two vectors is calibrated against three real distributions — random strangers, sibling pairs, parent-child pairs — to produce an honest percentile.
What this isn't
A paternity test, a kinship verification system, a medical tool. Face resemblance correlates only loosely with genetic relatedness — the network was trained to re-identify people, not to detect family. Numbers here are for entertainment.