Facial harmony score: an explainable beauty score

A facial harmony score is a composite number that summarizes measured symmetry, proportion, and feature balance across a face. Unlike a black-box beauty rating, an explainable score can be decomposed: every contributing number traces back to specific landmark measurements.

What a beauty score is

Search "beauty score AI" and you will find dozens of apps that take a selfie and return a number — 6.4, 78%, "top 12%". Underneath, most work the same way: a neural network trained on faces that human raters scored for attractiveness learns to predict what those raters would say about a new face. The output is a prediction of average human judgment, compressed to one number.

A harmony score is a different construction. Instead of predicting a rating, it measures geometry — symmetry, proportions, feature relationships — and summarizes how those measurements compare with population reference ranges. The distinction matters: a rating-predictor tells you what a crowd might say; a harmony score tells you what your face measures, and lets you see why.

The word "score" still deserves caution in both cases. Any single number describing a face involves choices — which measurements to include, how to weight them, what to compare them against — and those choices shape the result. The difference an explainable score makes is that the choices are visible and the arithmetic is checkable, so the number can be interrogated instead of merely believed.

The problem with black-box face rating

Most beauty-score apps share a set of structural problems worth understanding before you trust any number they produce:

• No explanation. A single number with no decomposition is unactionable and unfalsifiable. You cannot check it, learn from it, or meaningfully disagree with it.
• Inherited bias. A model trained on rater judgments learns the raters' preferences — including their demographic and cultural biases. Studies of attractiveness-prediction models have documented systematic score differences across ethnicity and skin tone that reflect training data, not faces.
• Instability. The same face photographed twice — different lighting, angle, or camera distance — routinely receives different scores. A number that changes with the lighting is measuring the photo as much as the face.
• A verdict framing. Presenting one unexplained number as "your beauty" invites users, often young ones, to treat it as a fixed judgment of worth. That framing does harm without offering understanding in return.

How Refrakt builds an explainable score

Refrakt's harmony score is assembled from geometry upward, so that any number in the summary can be traced back down to landmarks:

1. Landmarks first. A guided scan maps 468 facial landmarks with x, y, and z coordinates, under controlled distance and pose so photo conditions do not masquerade as facial features.
2. Three measured components. From those landmarks, three families of measurement are computed: symmetry (left–right differences of paired landmarks across the midline), proportion (facial thirds, fifths, and feature-to-feature ratios compared with population ranges), and balance (how features relate to one another — jaw width to cheekbone width, lip fullness to lower-face height, eye spacing to face width).
3. Region-by-region sub-scores. Measurements are grouped into regions — eyes, nose, lips, jaw, cheeks, profile — and each region gets its own sub-score. The summary never hides which region contributed what.
4. Traceability down to millimeters. Every sub-score expands into its inputs. If the mouth region scores lower on symmetry, you can see the specific measurement behind it — for example, one mouth corner sitting 2 mm higher than the other — rather than taking a number on faith.

A conversation starter, not a verdict

The honest use of any facial score is as a starting point for understanding, not an endpoint of judgment — for three reasons the measurements themselves make clear:

• Geometry is a partial input. Attractiveness research consistently finds that symmetry and proportion correlate with rated attractiveness, but so do skin quality, expression, grooming, animation, and context. A static geometric score captures a slice of the picture, not the picture.
• The reference is a choice. Every score compares measurements with some standard, and every standard is chosen by someone. An explainable score makes that choice visible; it cannot make it universal.
• The useful output is the decomposition. Knowing your composite is 74 tells you little. Knowing your symmetry is high, your facial thirds are typical, and your jaw region drives the deviation — that is information you can actually think about, discuss with a clinician, or simply file away.

Ethics: whose standard is it anyway?

Any tool that scores faces owes its users honesty about where standards come from. Beauty ideals are not constant: they vary across cultures — studies of facial preference find different preferred proportions in different populations — and they change within a culture over time, as any comparison of ideals across recent decades shows. Features prized in one era or place read as unremarkable in another.

This is why Refrakt's score is framed as harmony — internal consistency of symmetry, proportion, and balance — rather than conformity with any single ideal, and why the decomposition always outranks the composite. A number can describe your geometry. It cannot rank your worth, and a tool that implies otherwise is overclaiming. Analyses are informational; they are not verdicts, and they are not medical advice.

Frequently asked questions