How Our AI Face Scoring Actually Works

WannaModel scores faces by mapping 50+ facial landmarks from a single photo and rating 14 categories — including facial symmetry, jawline definition, proportions, and photogenicity — against patterns learned from 200,000+ labeled faces. The result is an entertainment-grade estimate of how closely a face matches conventions common in professional modeling photography. It is not a measure of beauty, worth, or how attractive you are to other people.

The pipeline, step by step

When you upload a photo, the model first detects the face and maps 50+ landmark points: the corners and centers of the eyes, the bridge and base of the nose, the lip line, the jaw contour, and the outline of the face. From those points it computes geometric measurements — left–right symmetry deviations, vertical facial thirds, the spacing between features relative to face width, and proportion ratios in the family of the “golden ratio” (phi, ≈1.618) used in classical aesthetics analysis.

Those measurements, together with learned visual features (skin texture, eye-area contrast, bone-structure cues), feed 14 category scores. A separate model is used for male and female faces because the photographic conventions the fashion industry applies to each differ measurably. The overall “modeling potential” score is a weighted blend of the categories, and your percentile compares it against the distribution of previously analyzed faces.

The numbers behind the model

Two different figures appear on this site, and they mean different things. The scoring model was trained on a dataset of 200,000+ labeled face photographs. Separately, over 2.4 million faces have been analyzed with the tool since launch — that larger set is what your percentile rank is computed against, not what the model was trained on.

What your score does NOT mean

A face score is a statement about photographs, not people. It reflects how closely one photo, in one light, at one angle, matches patterns in modeling imagery. It does not measure beauty — which is culturally variable and personal — and it does not predict whether people find you attractive, which depends overwhelmingly on things no photo contains. A low score means the photo diverges from a narrow photographic convention; nothing more. Real modeling casting also weighs height, body proportions, movement, personality, and market demand — none of which a face scan can see.

Known limitations and bias

Every face-analysis model inherits the biases of its training data, and ours is no exception. Datasets of modeling photography over-represent certain ethnicities, ages, and beauty conventions, which means scores can systematically differ across groups in ways that reflect the industry's historical biases rather than any real difference between faces. Lighting, camera distance, lens distortion, and head angle all shift scores — the same face can score noticeably differently across two photos. We publish this openly because pretending otherwise would be dishonest; treat every result as entertainment with error bars, not a verdict.

Privacy

Photos are used to run your analysis and are stored in a private bucket visible only to you; they are never shown publicly, and shareable score cards display your scores — never your photo. You can delete your scans at any time from your dashboard. See the privacy policy for full details.

Questions or corrections

Researchers and journalists are welcome to cite this page. If you believe something here is wrong or misleading, contact us — we'd rather fix it than defend it.