How the measurement actually works
This page exists so that every claim on the rest of the site has somewhere to be checked. Where we do not have something — a norm table for your group, a confident measurement on your photographs — it says so here first.
The capture set
Six required views and two optional ones. Every measurement declares which views it needs, and is suppressed if one is missing.
- frontal-neutral
- frontal-smile
- profile-left
- profile-right
- quarter-left
- quarter-right
- up-tilt-submental (optional)
- top-down-hairline (optional)
The views are not a photo shoot. They exist because a facial measurement is only comparable to another one taken under the same camera geometry, and a phone held at arm's length is not the same camera as a phone held at two metres.
Landmarking and the camera
Measurements are taken on a reconstructed three-dimensional mesh, not on pixels.
We place 521 landmark points and fit a three-dimensional model to them, solving for the focal length rather than assuming it. This is the step most consumer tools skip, and skipping it is why the same face measures differently in two photographs taken minutes apart: perspective at close range lengthens a nose, and a lens change moves facial width ratios by more than most real interventions do.
Every measurement carries the error budget of that fit. A number without one is not a measurement, it is an opinion with decimal places.
Every measurement, by region
The count is the sum of this table. It is checked against the specification on every build.
181 biometric tests across 13 facial regions
| Region | Code | Tests |
|---|---|---|
| General / whole-face | GEN | 8 |
| Eyebrows | BRW | 14 |
| Eyes & periorbital | EYE | 26 |
| Nose | NOS | 17 |
| Lips & perioral | LIP | 16 |
| Cheeks & midface | CHK | 13 |
| Jaw & mandible | JAW | 11 |
| Chin | CHN | 8 |
| Smile & dentition | SML | 13 |
| Neck & submental | NCK | 11 |
| Ears | EAR | 12 |
| Skin | SKN | 20 |
| Hair & hairline | HAR | 12 |
Reference populations, and where our coverage is thin
A percentile is a statement about a group. We name the group, and we publish where we do not have one.
Percentiles are computed against norm tables keyed by declared sex, declared ancestry group and age band. Ancestry is declared by you at intake and is never inferred from your photographs — that is an architectural rule, not a policy, and a build fails if any model output grows an ancestry field.
Coverage is not even. It is thinnest for South Asian, Middle Eastern and North African, and Latin American reference data. Where we have no adequate table for your group we show the raw measurement and say there is no percentile, rather than compute one against a population you are not in.
The bands a percentile is described in
| Band | Percentile range |
|---|---|
| Low | 1–20 |
| Mid-low | 20–40 |
| Mid | 40–60 |
| Mid-high | 60–80 |
| High | 80–99 |
What gets suppressed, and why
Every rule below is implemented in the analysis engine. This table is generated from it.
| Rule | When it fires | What it removes |
|---|---|---|
| 1 | Confidence is below 0.7. | The percentile |
| 2 | The 95% interval is wider than 20 percentile points. | The percentile |
| 3 | A required view is missing, or a required view failed QC. | The measurement |
| 4 | No norm for the declared population, or the test is descriptive-only. | The percentile |
| 5 | The test is beta and the user is not in the internal cohort. | The measurement |
| 6 | Measurement error exceeds 0.75 of the population SD. | The percentile |
| 7 | A required landmark is not visible in any usable view. | The measurement |
| 8 | The measurement spans depth and only the 2D path is available. | The measurement |
| 9 | Fewer than 90% of the Monte-Carlo draws came back finite. | The measurement |
| 10 | An invalidating QC condition covers this test family (makeup, eyewear, beard). | The measurement |
A suppressed measurement is shown to you as suppressed, with the reason. It is not silently dropped, because the absence of a number is itself information about your photographs.
The seven trait axes
Reported separately, never blended. Each one states how strong the evidence behind it is, including where that is weak.
| Axis | Evidence | What it is, and what it is not |
|---|---|---|
| Symmetry | Moderate | Fluctuating left-right deviation after pose correction. Real but small, and inflated by two-alternative forced-choice designs: significant in 2AFC, null in ratings across two independent labs (Lee et al. 2021; Jones & Jaeger 2019). Facial asymmetry does not track childhood health (Pound et al. 2014, N = 4,732). |
| Averageness | Strong | Distance from the ancestry- and sex-matched population mean shape. Strong and replicated, but a distance, not a destination: the preference is calibrated by each observer to their own diet of faces (Apicella et al. 2007), so the number is a statement about a named reference population rather than a universal property. |
| Dimorphism | Mixed | Position on the feminine-masculine shape axis, learned within the declared ancestry group and never across it (Kleisner et al. 2021 found a 2.5x range in sexual shape dimorphism across eight populations). Grade A for female femininity; grade D for male masculinity, where the sign itself is unstable across samples — so for male users the axis is reported as a position with no valence and no linked intervention. |
| Neoteny / visual age | Moderate | Estimated perceived age minus chronological age, predicted from an explicit feature vector rather than pixels so that the reasoning is inspectable. Grade A for the skin channel, grade C for the shape channel: most of the measurable effect runs through skin and facial contrast, not skull shape (Fink 2006; Matts 2007; Porcheron et al. 2013). |
| Adiposity | Strong | Facial fat signal from cheek, submental and jawline geometry. Reported as a position and never as a distance from a preferred point: the relationship with attractiveness is curvilinear, and the attractiveness turning point for women sits below the WHO healthy-weight midpoint (Coetzee et al. 2009). A linear treatment would make this a thinness score. |
| Homogeneity | Strong | Skin colour and texture evenness. The strongest and most actionable determinant in the evidence base, and measured as homogeneity rather than colour deliberately: colour effects do not survive multiple regression between people (Foo et al. 2017) while evenness does. Foundation makeup invalidates it entirely. |
| Proportionality | Weak — low evidence, labelled as such in the report | Aggregate deviation across ratio and angle tests from the matched reference distribution. Grade D as a beauty predictor, grade B as a description of morphology, and retained only for comparability with the category. It is not canon conformance: the neoclassical canons hold in 0-33% of healthy people in every population studied (Khoshab et al. 2022; Bozkir et al. 2004 measured 0% for the three-section canon in 500 Turkish adults). |
There is no eighth axis that adds these up. 7 trait axes reported separately is the whole design; see the refusals below.
Human review
Every analysis is checked by a trained reviewer before you see it, and their name is on it.
The reviewer can change or remove anything. They see watermarked working copies of your images, never your originals, through a viewer that requires them to state a reason, ties the grant to one task, and expires after sixty minutes. Every access is logged with the person, the reason and the time.
The turnaround is 48 hours because the person is real. That is a contractual commitment with an automatic refund behind it, not an estimate.
Visualizations
Your own pixels, moved by the size of a change we described.
A visualization is a deterministic warp of your own photograph, paired with the unmodified original. It is not a generated face. An identity check runs on the output and the render is discarded if it stops looking like you.
Rendered images are watermarked in the pixels, carry their provenance in the file, and are captioned as digitally rendered. We will not render an outcome we could not first describe as a measurement.
What we refuse to build
These are architectural, not editorial. Each one is asserted by a check that fails the build.
- No attractiveness score. No field anywhere in the system may be named one, and a lint over every schema, column and template fails if it is.
- No inferred ancestry. No model output may carry an ancestry field. Ancestry is declared at intake or the question is skipped.
- No third-party model APIs. No photograph, measurement or report content leaves our hardware. The machines that run inference have no outbound internet route, and a build fails if one appears.
- No minors, and no third parties. Adults, photographing themselves.
- No diagnosis. We are not a medical provider and we do not screen for disease.
The evidence register
Not published yet
Every empirical claim we make about the world will be listed here with its citation, its effect size and the over-reading it refuses. The register is authored as a reviewed file rather than as page copy, and it is not finished; publishing half of it would be publishing the half that happens to be ready.
Changes to norms and tests
Nothing has changed yet
Any change to a norm table, a test definition or the landmark set creates a version boundary for year-over-year tracking, and every one of them will be dated here. There have been none, because nothing has shipped to a customer.