How the facial harmony score is calculated

Follow the facial harmony score calculation: landmark distances, four weighted components, reference values and rounding. See exactly how the formula works.

Last updated: 2026-11-04

Landmark detection and scoring are separate steps

The facial harmony score calculation starts with estimated coordinates, not an AI opinion about beauty. Choose a clear photo containing one adult face and confirm you have permission to use it. The browser loads a face-landmark model from this website and estimates points around the visible face. The photo is processed on your device; the free tool does not send it to an AI provider or save it to a server.

The model estimates landmark locations. A separate, fixed formula calculates the score. The model is not predicting what a particular person will think of you.

Which measurements contribute to the total?

The 0–100 score combines symmetry (40%), facial thirds (25%), facial fifths (20%), and face length-to-width balance (15%). These weights and reference values are product choices, not a universal beauty formula or a population ranking.

Symmetry: compares selected paired points after correcting for image tilt.
Facial thirds: compares mesh-forehead-to-mean-brow, brow-to-nose-base and nose-base-to-chin fractions with equal thirds. The mesh boundary is not an anatomical hairline.
Facial fifths: compares all five projected horizontal parts with 20% each: cheek-to-outer-corner, first eye, inner-corner gap, second eye and outer-corner-to-cheek.
Length-to-width balance: compares the visible mesh length and width with the tool's reference.

The current engine (geometry-2.0.0) compares three visible vertical fractions with [1/3, 1/3, 1/3], five horizontal fractions with [0.20, 0.20, 0.20, 0.20, 0.20], and visible length-to-width with 1.35. Three and five sections each cover the whole visible dimension after eye-line roll correction. They are experimental product references, not anatomical norms or recommendations to change your face. If the boundaries do not form positive, ordered sections, the tool asks for a new photo instead of silently dropping a component.

Each component is clamped to 0–100 and rounded. The total is the rounded weighted sum of those four component scores. Symmetry measures the mean normalized mismatch of eight paired landmarks; it does not assess your entire face or skeletal symmetry.

Read the component descriptions alongside the overall score. A difference in one measurement can be hidden by the combined number.

Why a photo may be refused

No visible face, multiple detected faces, a small face, or a strong head angle can prevent a useful reading. The tool asks for a clearer photo instead of inventing a score. Passing these checks does not guarantee that every obstruction, filter, or lighting problem has been detected.

When the same formula produces a different result

A fixed formula gives the same output for identical coordinates. A new photograph usually produces new coordinates. Use a similar camera, distance, lighting and expression when comparing photos. Browser, model and preprocessing changes can affect the estimated points. Small differences are not evidence that your appearance changed.

The scoring formula in detail

For engine geometry-2.0.0, the component formulas are:

Symmetry: 100 − 500 × mean mirrored-pair error / face width. The error uses eight landmark pairs after correcting for head roll.
Facial thirds: 100 − 140 × sum(|third fraction − 1/3|). The fractions run from the visible forehead mesh boundary to mean brow height, then nose base, then chin.
Facial fifths: 100 − 140 × sum(|fifth fraction − 0.20|). The boundaries are cheek 234, outer eye 33, inner eye 133, inner eye 362, outer eye 263 and cheek 454 in MediaPipe mesh indexing. This uses all five parts, not just eye-centre spacing.
Length-to-width: 100 − 100 × |visible face length / face width − 1.35|.

Each component is clamped to 0–100 and rounded before weighting. The final score is round(0.40 × symmetry + 0.25 × vertical + 0.20 × spacing + 0.15 × ratio).

These constants describe our implementation, not proven beauty optima. Horizontal coordinates are scaled by image width and vertical coordinates by image height before measuring; treating a non-square photo as a square would distort the calculation.

Worked example: follow the contribution, not just the total

The sample report uses illustrative components of 88, 76, 80 and 79. Their contributions are 35.20, 19.00, 16.00 and 11.85 points. The sum is 82.05, displayed as 82/100. These are example inputs for explaining the formula, not a real person's result or an average score.

Facial thirds have the largest weighted gap in this example: (100 − 76) × 0.25 = 6.00. That makes it a useful place to start reading. A large gap is distance from a product reference; it is not a recommendation to alter a facial feature.

What AI adds to a paid report

The free analysis opens on a separate /result page with the total, four component scores and visible weights. It also includes a brow-based visible-thirds map, five horizontal parts, a mathematical golden-ratio comparison, an approximate outline description and general skin-care basics. The score card also shows the approximate face outline and a local skin-photo check. Two small cheek patches are sampled for mean RGB brightness (0–255), absolute left–right mean difference and the percentage of pixels with any channel at least 250 or all channels at most 5. These are image statistics, not skin-type, acne, hydration or health detection. Lighting, makeup, filters and compression can dominate the readings. Illustrative examples do not claim a skin measurement. The thirds and fifths panels use the same partitions as their scored components. The golden-ratio, outline and skin-photo panels do not change the score. The mesh forehead boundary is not the hairline, and the golden ratio is not a validated beauty target. A $9.90 in-depth analysis report contains the full 16 descriptive readings with units and definitions, a deeper written interpretation of the eye, nose, mouth and lower-contour measurements, measurement-based styling experiments for face framing, brows, jawline presentation and hair, reading priorities and a saved PDF. The full reading table and regional details are reserved for this report rather than displayed on the free result page. The server calculates the measurements before asking AI to explain them. AI does not choose a new score or see your photo.

Each styling idea explains the readings behind it and suggests a reversible comparison, such as changing a parting, brow styling, visual framing or photo setup. It does not prescribe anatomical changes or promise a score increase. General skin-care guidance is included separately with an American Academy of Dermatology source; facial geometry cannot establish skin type, skin health or suitable treatments.

Only derived measurements and component values go to a third-party AI processing service after payment verification. The generated explanation may make mistakes; inspect it against the fixed measurement table. The saved report includes its calculation and report versions. Compare free and paid features and read the separate report data flow before consenting.

Implementation evidence and reproducible checks

Checked on November 4, 2026, for engine geometry-2.0.0: the downloadable verification record lists actual synthetic test inputs, expected relationships and observed outputs. The synthetic landmark fixture contains no person's face or photo. These checks cover identical-coordinate repeatability, resolution/aspect-ratio handling, small translation, scale and head roll, plus rejection of invalid, undersized and excessive-pose inputs.

These checks exercise the formula, not real-photo detection accuracy, demographic fairness or human attraction. The maximum accepted roll is 18 degrees, the nose displacement proxy limit is 0.16 face widths, and the visible face must satisfy minimum pixel and image-fraction sizes. Passing these thresholds does not prove a photo is ideal.

The free path loads /models/face_landmarker.bin and /wasm files on this website; it runs landmark detection in the browser. You can inspect Fetch/XHR while taking the free test. An absence of photo upload in that browser session is narrower evidence than an independent infrastructure audit. Read the verification scope for what has and has not been tested.

Older saved reports keep their original engine version and component definitions. The former geometry-1.0.0 used 30/30/40 eye-bridge-based vertical segments and an eye-centre spacing reference of 0.46. Its archived verification record is retained; old and new totals should not be compared as if the formula were unchanged.