Measurements & methodology

01 / From photo to proportions

The shutter freezes a photo immediately. MediaPipe estimates facial mesh landmarks, head pose, and expression signals. The measurement pipeline accounts for image aspect ratio and aligns the eye-center axis to remove in-plane roll; ratios remove image scale. It does not correct head turn, nod, lens perspective, or distance distortion. A landmark is a model estimate, not a manually verified anatomical point.

The main measurements have preset ranges. More measurements contains additional shape, eye, and regional mirror descriptors. These compare with a front projection of MediaPipe’s canonical face model, labeled Model ref. They do not affect the score.

• Visible face height / width uses mesh forehead point 10 to chin 152, divided by cheek span 234–454. Point 10 is not the hairline. Jaw span / cheek span describes the projected outline, not jawbone structure.
• Nose-base position locates point 2 between the eye-center line and chin. Mouth position locates the mouth-corner midpoint between that nose-base point and chin. These separate vertical positions can change independently.
• Each eye has its own opening-to-width ratio and corner-tilt estimate. Eye A and B name mesh groups, not guaranteed anatomical left and right. One image cannot establish an eye-style classification or repeatability.
• Regional mirror differences compare eyes, nose, mouth, cheeks, jaw, and chin against an eye-centered vertical axis. The two eye-corner pairs share one region. The chin value is twice its lateral offset. These are 2D projection differences, not intrinsic facial asymmetry.

The fixed nasal span uses mesh points 98 and 327. It is an inner nasal mesh span, not a validated measure of the outer alar width; the mouth-to-nasal-span ratio inherits that limitation. The confirmed forehead ratio equals twice hairline-to-inner-brow height divided by inner-brow-to-chin height. It does not measure the three facial thirds independently.

Regional symmetry combines with a root mean square, so larger differences have more influence. The detailed regional values remain visible separately instead of being hidden inside that one number.

You confirm the visible central hairline or mark it unavailable. Your photo stays in browser memory; derived measurements and capture context are sent for report storage. Use a neutral expression, even frontal light, and the same camera and distance for comparisons.

Measurement uncertainty

The analysis stores estimated turn, nod, and tilt, plus expression and lighting signals. An estimated turn or nod greater than 3° marks the remaining geometry as capture-sensitive. Lighting or sharpness signals can also flag a report. Reports use a compact Provisional label for the affected reference, while its number or missing-feature range remains visible. These are unvalidated software checks that can misread a well-lit, still photo; they are not proof of a problem with the photo.

If the model cannot confirm a neutral mouth and jaw, mouth- and jaw-dependent descriptors are omitted and the reference is marked provisional. Brow-expression flags qualify usable brow measurements instead of discarding them: the photo review lets you confirm that the brows are visible and relaxed. Only missing or invalid landmarks, or marking the brows as obscured, omit brow measurements. The forehead measurement requires a confirmed hairline and usable inner-brow, nose-base, and chin landmarks. Expression flags can qualify it but do not erase a valid, confirmed measurement. Uncertain open eyes exclude eye-opening and corner-tilt descriptors.

The 3° threshold is a product heuristic, not a validated error tolerance or a pose correction. Expression signals are model blendshape estimates, not calibrated per-feature confidence. “Measured” means a usable calculation was produced under these checks; it does not guarantee that the model located every landmark correctly.

02 / The experimental reference

Each rule has one center and a scale. The center is the midpoint of its reference span, except symmetry errors, whose center is zero. The scale is half the span width, or the upper endpoint for a zero-centered rule. A measurement one scale away from its center scores 5.5. This scale is a product choice, not a standard deviation or population norm.

Reading the ranges

Preset range below a measurement is the span defined by the selected male or female preset, saved with that report. It is a product reference, not an expected population range, a medical normal range, or a boundary between acceptable and unacceptable faces. Historical reports retain their saved ranges.

In range means the displayed measurement falls between the displayed endpoints, inclusive. Above and Below show the distance to the nearest endpoint in the same units as the measurement. Percentage measurements use percentage points (pp): 2.0% against an upper endpoint of 1.5% is 0.5 pp above. Comparison and distance use the displayed precision so a rounded value at an endpoint does not appear outside it.

Target 8–10 below an overall or category score is the app’s goal band, based on the threshold for a close preset match. It is not the expected score of an average person. A missing-feature score range describes possible results for the unavailable categories; it is separate from this target band. Being inside a measurement’s preset range does not guarantee a score in the goal band: the calculation still measures distance from its center.

Bars highlight the preset span; the dark marker shows your value. Each bar has a fixed scale derived from its reference, so different results do not move the reference band. Extra mirror-difference bars use a 0–8% display scale; extra eye-tilt bars use −10° to 15°. These axis limits are display choices, not acceptable ranges. Values beyond a scale stay at the corresponding edge with an arrow and an Above scale or Below scale label. The number remains visible. Missing measurements have no marker.

References for more measurements

Model ref is a comparison point calculated from the MediaPipe canonical face mesh, using the same landmark indices and 2D measurement functions as your report. We project the model’s X/Y coordinates straight onto a plane, flip the vertical axis to image coordinates, and omit depth. The official model documentation describes this static mesh as a reference for scale and face alignment.

These are model comparisons, not human averages or attractiveness targets. A single model supplies a reference point, not an acceptable range: its bar uses a highlighted line instead of an invented band. Both presets use the same model points. The six mirror-difference references are zero because the model is symmetric. No new scoring weights or penalties are applied.

The reference set is pinned as canonical-orthographic-v1 for the descriptor formulas used by geometry-photo-v4 and geometry-photo-v5. Version 5 fixes brow and forehead availability; it does not change these formulas. The source mesh, license, generation script, and tests are included in the project so the values can be reproduced.

*The two brow rules contribute equally to one 10% brow category. Their scores are combined with the same root-mean-square mismatch calculation below, before combining the six categories.

The calculation

rule score = 1 + 9 × 2^(−((value − center) / scale)²)

Only the exact center has an unrounded score of 10. Scores decrease smoothly toward 1 with distance from the center. There is no perfect-score plateau across a reference span.

reference = 10 − √(Σ(weight × (10 − category score)²) / Σ(weight))

This combines six categories by their weighted distance from 10. Larger mismatches matter more than they would in a simple average. The result rounds once to one decimal. There is no extra severity multiplier or curve. The v3 cubic penalty has been removed because lowering everyone's result did not improve measurement accuracy.

Only these six reference categories affect the number. Face outline, independent vertical positions, separate eye measurements, and detailed regional differences are descriptive evidence. They are not additional attractiveness predictors.

Rounding can produce ties, and a displayed 10.0 can include values that round up. No value on this scale has an established relationship to population rank or independent human ratings.

When a feature is unavailable

Missing brows or hairline produce a possible full-reference range without a single overall score or tier. The lower endpoint assigns missing categories a score of 1; the upper assigns them 10. Both calculations retain all six category weights. The lower bound rounds down and the upper bound rounds up to one decimal.

This range shows the effect of missing information. It is not a confidence interval or a prediction of the missing features. Coverage is 80%, 90%, or 100% of category weight, not measurement confidence. You confirm a partial report before completing it; rejected captures use no scan credit.

03 / The tiers

Complete references that pass the pose, lighting, and mouth/jaw checks use these app-specific looksmaxxing labels after rounding. They describe the experimental preset match, not attractiveness rank. Missing-feature ranges and references marked provisional have no displayed tier.

04 / Strengths and focus points

Available reference categories scoring at least 8 can appear as closer matches, highest first. Categories below 6 can appear as further from the reference, lowest first. Reports show up to three closer matches and two focus points; published category order breaks ties. These thresholds are product choices, not evidence of anatomical strengths or defects. When the reference is flagged for uncertainty, these labels and category scores are hidden; the overall experimental number or range remains visible.

Current single-photo scans omit the “hunter-style eyes” label because one photo cannot establish the temporal stability that its earlier rules required. The separate eye-opening and corner-tilt measurements are descriptive only. Optional profile observations do not change the frontal reference.

Suggestions address capture conditions, grooming, and care. They do not diagnose a recessed maxilla or promise changes to facial bones.

05 / Evidence and validation

Google’s Face Mesh model card evaluates landmark accuracy for mobile AR. It does not validate attractiveness scores. The canonical mesh defines the model points; it does not make their spans clinical anthropometric measurements. Research on selfie perspective shows that short camera distances can alter apparent facial proportions.

Research on individual preferences also finds reliable differences in how people judge faces. These sources explain limitations; they do not validate our centers, scales, weights, or tiers.

Empirical validation is not complete. Before claiming calibrated attractiveness estimates, we need a consented, commercially licensed adult sample, independent rater groups, and separate training, calibration, and test sets with no person shared between them. Testing must measure repeat-capture agreement, prediction error, top-score calibration, and performance across relevant subgroups and devices.

The SCUT-FBP5500 benchmark is restricted by its owners to noncommercial research. It has not been imported to validate or power this product. Passing software tests checks implementation, not agreement with human attractiveness ratings.

06 / Comparing reports

Compare results with matching presets, coverage, and capture conditions. Historical reports keep their original data: scoring v1 used broad bands and an arithmetic average; v2 introduced continuous reference curves and RMS aggregation; v3 added a cubic severity curve; v4 removes that penalty. Measurement v4 adds separate feature analysis; measurement v5 retains those formulas and fixes the brow/forehead availability filters. Complete v4 and v5 measurements remain compatible for rankings. The app recalculates confirmed brows and forehead from the frozen photo before saving. Earlier omitted measurements cannot be recovered from saved reports without another photo.

Free, credit, subscription, and test scans use the same calculation. Completed reports display the experimental score or missing-feature range, with a Provisional label when flagged. The detailed measurement definitions and limitations live on this page. Reopening a saved report displays its existing calculation, without another scan. Payment status does not change the evidence behind a result.