What we measure, how we score, and where we can be wrong.

Two product paths with different jobs

ARF Local Geometry v3.0 powers all eighteen free tools, including the overall score. It detects 478 visible face landmarks in the browser and calculates documented two-dimensional distances, angles, ratios and photo-quality signals. The free photo, landmarks and result are not submitted to our server.

ARF Full Report v1.0 is an account product that costs 10 wallet credits. After Google sign-in and explicit consent, it sends one resized adult portrait for transient model processing and saves the resulting text and scores to the user’s account. It does not save the photo.

Neither path determines objective attractiveness or identifies health, personality, ethnicity, gender, age, identity or social value.

The free local geometry engine

The browser loads MediaPipe Tasks Vision 0.10.35 and the site-hosted Face Landmarker model. A suitable face produces 478 normalized landmarks. AI Rate Face converts the visible x/y coordinates into the processed image’s pixel space before running its own formulas.

Eye-line alignment reduces roll distortion in symmetry comparisons, but two-dimensional coordinates cannot fully correct head rotation, camera perspective, focal length, expression or obstruction.

How the free overall score is composed

The total is calculated locally with a fixed formula:

• Visible symmetry uses pose-aligned left/right landmark residuals.
• Proportion balance combines facial thirds, fifths, eye spacing and selected ratio references.
• Feature balance combines jaw, eye and central-feature relationships.
• Photo presentation reflects local quality and pose suitability.

The weights and reference ranges are product design choices, not scientific laws or universal beauty standards. They change only with a new published version.

Face geometry versus photo presentation

A single combined score cannot tell you whether a low number came from your face or from your photograph. Every free result therefore also reports two separate values:

• Face geometry excludes the photo-quality term entirely, so it is the value that should stay roughly constant across photos of the same face.
• Photo presentation is the photo-confidence score, decomposed into six ranked drivers: head angle (24%), sharpness (22%), lighting (20%), tonal contrast (16%), framing (10%) and resolution (8%).
• The gap between them is reported in plain language. A gap of ten points or more is described as the photo holding the score back.

Face geometry is not fully pose-independent — symmetry in particular still drifts with head rotation, and expression genuinely changes the measured landmarks. The Face Score Consistency Test measures the spread of both values across two or three of your own photos rather than asking you to trust the claim.

Because the formulas are deterministic, the same processed image always returns the same numbers. There is no sampling, temperature or randomness anywhere in the free path.

How the credit-powered Full Report works

1. The browser first detects one face and checks pose, light, clarity and framing locally.
2. The user must be signed in, hold at least 10 credits, and confirm adulthood, photo rights and cloud processing.
3. The resized photo is sent through the same-origin Worker to a privacy-filtered OpenRouter model route.
4. The model returns a strict schema containing four scores, shape context, 4–6 visible-feature observations, style directions and a repeatable portrait plan.
5. The Worker normalizes and saves the report JSON. It does not save the original or resized photo.

The paid report uses the same 30/25/25/20 total formula after the model returns its four visual estimates. A fixed schema, low temperature and fixed seed reduce unnecessary variation but do not make a generative visual model deterministic.

If the result is no face, multiple faces, uncertain adulthood, low quality or a processor failure, all 10 credits are returned automatically.

What confidence means

Confidence describes whether this photo is suitable for a repeatable visual estimate. It is not the probability that the score or shape label is objectively true.

Confidence falls with strong head turn or tilt, close wide-angle perspective, uneven light, blur, heavy filters, obstruction, tight cropping or uncertain face detection. A low-confidence result should lead to a retake, not a stronger conclusion.

Known limitations

Attractiveness is subjective

Preferences vary across people, cultures and contexts. A photograph cannot capture movement, voice, chemistry, warmth, style or personality.

Photos are not neutral

Distance, focal length, head position, lighting and expression change visible width, symmetry and feature prominence. Different photo conditions are not a reliable before-and-after comparison.

Bias remains possible

We do not claim equal performance across every demographic group. A fairness claim would require a documented evaluation dataset and published results that we do not currently have.

No medical or high-impact use

Do not use the service to diagnose a condition, recommend surgery, screen applicants, rank strangers or make decisions about another person.

Adult-use and content rules

The service is restricted to adults aged 18 or older. Users must own the selected photo or have the pictured adult’s clear permission. The paid model is instructed to reject age-ambiguous photos; its refusal is a safety state, not an age estimate.

Insults, sexualization, discriminatory labels, medical claims and cosmetic-surgery recommendations are outside the product scope. See the Terms of Use.

Version history