Faces Are More Than Pixels
When you look at a digital photo, you see a face. When a computer looks at the same image, it sees a grid of millions of numbers representing colour intensities. The fundamental challenge for any facial recognition system is that the same face produces completely different pixel values depending on lighting, angle, or the camera used. A photo taken in bright sunlight has entirely different pixel data to one taken indoors, even if both show the same person clearly. Raw pixels are therefore useless as a basis for identity comparison.
A facial fingerprint solves this by discarding pixels and extracting something more fundamental: a compact mathematical representation of the structural identity of the face. This representation is invariant to superficial conditions, the same bone structure, the same geometric proportions, the same inter-feature relationships will produce a fingerprint that stays recognisably similar across many different photographs.
What the 512 Numbers Actually Represent
A neural network processes your photo through a series of convolutional layers, each one extracting progressively higher-level features. Early layers detect edges and gradients. Middle layers combine these into shapes, nose bridges, eye socket curves, jaw contours. The final layers produce a 512-dimensional embedding: a list of 512 numbers that encodes the structural identity of the face in a format optimised for comparison. This is the facial fingerprint.
The network does not explicitly measure any single feature. It discovers, through training on millions of face photos, which combinations of measurements best predict whether two photos show the same person. The result is a representation more discriminative than any hand-designed feature set, one that captures subtle relationships between features that no human would think to measure explicitly.
L2 Normalisation and Similarity Search
The 512-number vector is L2-normalised, meaning it is scaled so that all values together have a length of exactly 1. This places every facial fingerprint on the surface of a unit hypersphere in 512-dimensional space. The practical consequence is that comparing two fingerprints reduces to measuring the angle between two points on this sphere, a clean, efficient operation known as cosine similarity.
Ollie computes a fingerprint for every celebrity photo in its database ahead of time. When you upload a photo, your fingerprint is computed once and compared with every stored celebrity fingerprint, a single matrix calculation that takes milliseconds. The result is a ranked list of the closest matches in that 512-dimensional space: your celebrity lookalike results.
Why Your Fingerprint Is Unique
The uniqueness of a facial fingerprint comes from the geometry of trained embedding space. The network is optimised so that same-identity face pairs land in tight clusters, while different-identity pairs are pushed to distant regions of the space. Even identical twins tend to produce fingerprints that are close but measurably distinct, reflecting small differences in feature placement that develop even from an identical genetic starting point.
Unlike raw pixel comparison, a facial fingerprint is largely invariant to lighting variation, small angle changes, and minor expressions. The underlying bone structure and geometric proportions remain consistent, and those are what the fingerprint captures. This is why you can upload photos taken years apart and still receive consistent top matches, the geometry of your face has not changed.
Practical Tips for Better Matches
Understanding what a facial fingerprint captures helps explain how to get the best results. Because the fingerprint encodes bone structure and feature geometry, photos that show your face clearly and at natural proportions produce the most accurate results. Avoid extreme angles that distort apparent nose width or jaw shape, and avoid heavy makeup that changes apparent facial proportions.
Front-facing, well-lit photos without strong shadows across the face produce the most stable fingerprints, meaning you will receive more consistent results across multiple uploads. The fingerprint approach also means that a photo from ten years ago may produce the same top match as a recent one: your biometric geometry changes slowly if at all.