The math behind your LooksLab score

The pipeline

A scan moves through four stages: capture, detection, measurement, and scoring. Landmark detection (MediaPipe via WebAssembly) runs in your browser. Scoring runs on our server, which receives only the landmark coordinates and your selected gender, not the photo. Photos are uploaded to a private Supabase Storage bucket so you can revisit your scan and delete it later from My Scans.

1. Landmark detection

Once you upload or capture a photo, we run Google’s MediaPipe Tasks Vision face mesh model on the image. The model returns 478 three-dimensional landmark points anchored to anatomical features such as the lash line, alar base, vermillion borders, jaw angles, and tragion. These points are the raw material for everything that follows.

2. Geometric measurements

From the landmark cloud we derive a structured set of measurements: distances, angles, ratios, and a few areal quantities. Examples include intercanthal width over alar width, lower facial third over total facial height, gonial angle, columella to upper lip angle, and the canthal tilt. All measurements are normalized to the interpupillary distance to keep them independent of image resolution.

3. Curve scoring

Each measurement is run through a calibrated scoring curve. A curve takes a single number, for example the ratio of upper lip to lower lip height, and returns a value on a fixed harmony scale. The shape of the curve encodes what we know about that measurement: where it peaks, where it falls off, and how steeply it does so.

4. Composite scores

Individual curve outputs are aggregated into four pillar scores: Harmony, Angularity, Dimorphism, and Health Indicators. We use weighted averages with weights tuned against a held-out calibration set, not equal weights, because some measurements carry far more signal than others. An overall score blends the four pillars.

The 66 calibrated curves

LooksLab ships with 66 distinct scoring curves. Each curve maps a measurement (a ratio or an angle) to a 0 to 10 score using cubic-bezier interpolation between calibrated control points. The control points were derived from population data so that values close to the modal range score highest and the tails fall off smoothly. We chose curves over hard thresholds for a simple reason: real measurements do not have a single right value with everything else being wrong. A nasolabial angle of 96 degrees is not a binary pass or fail, it is a point on a smooth response surface.

Each curve is defined by a small set of control points plus a mode that determines how those points are interpolated. Most curves use a custom cubic-bezier so a single measurement can be mapped through asymmetric, non-monotonic response shapes. A few simpler ratios use piecewise linear curves where the underlying literature does not support more nuance.

We ship per-gender curves today: male and female versions of the same metrics where the calibration differs. When we change a curve we ship the calibration data alongside it so the change is reviewable.

What the four pillar scores capture

Harmony

Harmony measures how close key proportions and ratios are to population-calibrated targets. It pulls in symmetry, midline alignment, the balance of the facial thirds, the horizontal fifths, the midface ratio, the lower-third proportion, eye separation, and the major intra-feature ratios that show up repeatedly in the cosmetic surgery literature.

Angularity

Angularity captures bone-driven structural sharpness: the gonial angle, the mandibular plane, the projection of the chin and zygoma, the sharpness of the jaw line in profile. Two faces can have similar harmony scores while one reads soft and rounded and the other reads chiseled. Angularity is what separates them.

Dimorphism

Dimorphism measures how strongly features sit on the typical masculine or feminine end of population distributions for your selected gender. Per-gender calibration means a measurement is scored against the modal range for that gender in our calibration set, not averaged across the full population.

Health Indicators

A small set of measurements that have been associated with general health and developmental signals in the literature. Think of these as correlates, not diagnostics. A high or low score here is not a clinical claim about anyone’s health.

What this score cannot tell you

A LooksLab report describes the geometry of a face in a single photograph. It says nothing about how that face moves, smiles, speaks, or carries itself. It cannot account for voice, posture, charisma, grooming, or the thousand small choices that shape how people experience you in person.

Lighting matters a lot. Harsh top light deepens shadows under the brow and chin, which can shift several angularity measurements by a meaningful margin. Camera focal length matters too. A photo taken with a phone held close to the face exaggerates the nose and flattens the ears relative to a portrait shot at the recommended one to two meter distance.

And, to be direct: a high or low score is not a verdict on a person. Attractiveness in the real world is shaped by familiarity, culture, fashion, social context, and individual taste in ways that no geometric pipeline can capture. The score is a snapshot of one photo through one lens. Take it as information, not as a ruling.

What we are working on next

We already ship per-gender curves (male and female versions of the same metrics where the calibration differs), so measurements with genuinely different expected ranges are scored against the right reference. The next step is widening the calibration set to improve demographic coverage of the underlying data.

Beyond that we are exploring confidence intervals on each measurement, so a report can tell you not just where you scored but how stable that score is across different photographs of you. If you have ideas for what should come next, we read every note that comes into hello@lookslab.app.