How the score is calculated, line by line

A face-landmark model is downloaded to your browser and run against your photo. It returns 478 3D points - the face mesh plus iris points - each expressed as a fraction of image width and height. This is the only machine-learning step in the pipeline, and it is the only step whose output is not fully deterministic arithmetic.

The model needs WebGL, and there is no fallback without it: the landmark graph builds its image nodes on the GPU even when the CPU delegate is selected. The tool checks for WebGL before it starts and says so plainly when it is missing, rather than failing at you.

The image is never transmitted: no request carries it, and there is no server-side component of this site that could receive it.

Before anything is scored, yaw, pitch and roll are estimated from the cheekbone and midline points. A rotated head is the dominant cause of bad measurements, so the result panel grades pose and warns you when it is outside tolerance rather than quietly scoring a compromised photo.

Pitch tolerance is deliberately asymmetric: the typical selfie is shot from above, so up to 25° of chin-down is tolerated against 18° of chin-up.

Each paired feature is reflected across the facial midline and the mismatch is measured as a percentage of face width. Raw deviation is then mapped to a 0-100 score through a per-feature curve calibrated on real faces - not a linear penalty, which punishes ordinary sub-degree landmark noise far too harshly.

Feature pairs combine with these weights:

• eye - 25%
• eyebrow - 15%
• nose - 10%
• mouth - 15%
• jaw - 35%

The jaw carries the most weight because it is the pair a viewer notices first; the nose carries the least because it deviates least across the population.

Four sub-scores are combined into the 0-100 composite that the results panel shows. There is no division into a mark out of ten - that format reads as a rating of a person rather than a measurement of a photograph.

• Overall symmetry - 35%
• Facial thirds balance - 25%
• Facial fifths balance - 25%
• Golden ratio - 15%

Symmetry dominates because it has the strongest empirical link to human ratings and the least cultural baggage. The golden ratio carries the least weight for the opposite reason: it is the most inherited, and the least defensible as a universal target.

Scores are compared against a committed table of 7,997 previously measured scans, one value per percentile per metric. That table is what makes an "ahead of 72%" claim mean something specific rather than being a decoration.

Two deliberate constraints: claims cap at 99% - no face is ever told it beat everyone - and no claim is shown at all below the 55th percentile, because a rank is only useful information at the top and only discouraging at the bottom.

• 2D only. Depth, projection and profile geometry are largely invisible to a single front-facing photograph.
• Culturally loaded targets. The proportion ideals descend from one aesthetic tradition and are not universal.
• Self-selected calibration set. The percentile distribution comes from people who chose to scan their face, which is not a random sample of humanity.
• Photo-dependent. Angle, lighting and expression move the numbers, sometimes by more than the difference between two people.
• Weak outer claim. Geometry is a modest predictor of what people find attractive. That gap is a property of the world, not a bug in the code.

Is this how pretty am I test accurate?

It is accurate at what it measures and silent on what it does not. The geometry is exact and repeatable to the decimal. The inference from geometry to attractiveness is weak - published research finds symmetry and proportion explain only a modest share of how attractive people rate faces. So: a precise measurement of a narrow thing, not a precise verdict on a broad one.

Where do the percentile claims come from?

From a committed distribution table of 7,997 previously measured scans, one entry per percentile per metric. A score is looked up against that table to produce an "ahead of N%" figure. The claims are capped at 99% and suppressed below the 55th percentile, so a low sub-score never gets a rank attached to it.

Is the model biased by ethnicity?

Landmark detection performs well across skin tones and face shapes. The scoring is a different matter: the golden-ratio and thirds targets come from a European-descended aesthetic tradition, and there is no reason a face from a different population should match them. A lower proportion score often reflects that mismatch rather than any real asymmetry or imbalance. Symmetry scoring is free of this problem, because it compares your face only to itself.

Why does the test warn me about head angle?

Because head rotation is the single largest source of false low scores. Turning the head compresses one side of the face, which the symmetry pass reads as genuine asymmetry. The pose estimate grades yaw, pitch and roll, flags "fair" beyond 10° of yaw or 12° of roll, and "poor" beyond 14° and 15°.

Does the score use AI?

Only for landmark detection. A convolutional model locates the 478 points on your face - that part is machine learning. Everything after it is deterministic arithmetic: distances, ratios, weighted sums and a lookup table. No language model, no generative model and no black box produces the score itself.

• The pretty scale, and what each score band means
• What happens to your photo (nothing)