How AI Face Matching Works (In Plain English)

No math degree required. Here's how a computer goes from a photo to a 'these two faces are 87% alike' verdict.

"The AI compared your faces" sounds like magic, but the idea is surprisingly approachable. Here's the whole pipeline without the jargon.

1. Find the face

First the software scans the image to answer a simple question: is there a face here, and where? A detection model draws a box around each face it finds. If it spots zero faces or several, a good tool will ask you for a cleaner photo — which is exactly why single-face, front-on shots work best.

2. Map the landmarks

Next it places landmarks — dozens of dots on the corners of the eyes, the edges of the lips, the bridge and tip of the nose, the line of the jaw. Think of it as the face's connect-the-dots outline. These points let the software rotate, straighten, and crop the face to a standard pose so two photos can be compared fairly.

3. Turn the face into a fingerprint of numbers

Here's the clever part. The aligned face is fed through a neural network that outputs an embedding — a list of numbers (often 128 of them) that captures the face's geometry. You can picture it as a coordinate, a single point in a vast 'face space.' Faces that look alike land near each other; faces that look different land far apart.

4. Measure the distance

To compare two faces, the software measures the distance between their two points in face space. A small distance means the faces are very similar; a large distance means they're quite different. That raw distance is then converted into the friendly percentage and verdict you actually see.

5. Score the confidence

Finally, a well-built tool reports how much to trust the result based on photo quality, how clearly the faces were detected, and how decisive the distance was. A blurry, side-angled photo lowers confidence even if it produces a number.

That's it — detection, landmarks, embeddings, distance, confidence. Dad or Mad may run quick photo-readiness checks in your browser, then uses private AI processing for previews and reports (details in our privacy policy). And remember: this measures how faces *look*, not genetics — see can a face tell you who a child looks like?.