How Do We Judge a Look-Alike? — The Science Behind Face Similarity

"Looking alike" feels different to everyone, and it's hard to put into words. So how does a computer turn that into a number like "your eyes are 82% alike"? Here's how Dalmum does it, in plain language.

Step 1: Find the "dots" on each face

First, we locate points at fixed positions on the face in the photo: the corners of the eyes, the top and bottom of the eyelids, the line of the eyebrows, the bridge and sides of the nose, the outline of the lips, the curve of the jaw — 68 points in total. These are called landmarks. Essentially, a face's shape can be summarized by where these points fall.

Step 2: Align both faces to the same size and angle

Every photo has a different face size, and heads are often tilted slightly. Comparing photos as-is would compare "how the photo was taken" rather than "how alike the faces are." So we scale both faces using the distance between the eyes as a reference, and rotate them so the line between the eyes is level — aligning both faces to the same conditions.

Step 3: Compare each feature

The aligned points are grouped by feature, and each feature's shape is turned into numbers. For the eyes, that might be width, height, and whether the outer corners tilt up or down. Comparing the same measure between two faces gives a high score for a small difference and a low score for a big one. This is how the six categories — eyes, eyebrows, nose, mouth, face shape, and feature layout — are calculated.

Step 4: Calculate the overall similarity

The overall score blends two things: the distance between two "fingerprint-like" summaries of the whole face (closer means more alike), and the average of the per-feature scores you just saw. These are combined into a single percentage.

So, should I trust this number?

Treat it as fun only. Since this method only looks at a few dozen points, it can't capture things like a shared "vibe" that people notice. Scores also shift with expression and angle, even for the same person. Places where faces are actually used to verify identity (security access, ID checks, etc.) use far more sophisticated and validated technology — an entirely different domain from a look-alike test.

For more consistent results, check out how to pick a good photo.