How MIRAÉ measures

MIRAÉ combines computer vision with AI so your scores are objective and your guidance is personalized. This page documents exactly how each number is produced, how confident we are in it, and where the limits are.

Geometry (the numbers)

A 468-point facial mesh runs on your device. From it we compute real anthropometric measures and calibrate to millimeters using the iris (~11.7 mm). Head pose is checked so off-angle frames don't skew results, and the front capture averages several frames to reduce noise.

What each number means

Symmetry. We mirror nine paired landmarks — eye corners, brows, cheeks, nose wings, mouth corners and jaw — across the facial midline and measure how far each side deviates.
asymmetry = mean(|left−right horizontal| + |left−right vertical|) ÷ interocular distance · ideal 0 = perfectly mirrored — Farkas, Anthropometry of the Head and Face (1994)
Facial thirds. We divide face height into three sections — hairline to brow, brow to nose base, nose base to chin — and check how close each is to one third.
balance = how close upper / middle / lower each sit to 33% · ideal ≈ 33% each — Neoclassical canons — Farkas et al. (2005)
Canthal tilt. The angle of the line from the inner eye corner to the outer eye corner, relative to horizontal, averaged across both eyes.
arctan(vertical rise ÷ horizontal run between the eye corners) · ideal ≈ 4–7° upward — Rhee et al., periorbital aesthetic norms
Facial width-to-height. Cheekbone-to-cheekbone width divided by upper-face height (brow line to upper lip).
fWHR = bizygomatic width ÷ (brow → upper-lip height) · ideal ≈ 1.8–2.0 — Weston, Friday & Liò (2007)
Jaw angle. The jaw angle measured cheek → jaw corner → chin, using the mesh depth (z) for a true 3D angle when the capture provides it.
angle(cheekbone, jaw corner, chin) · ideal ≈ 120–130° — Naini, Facial Aesthetics: Concepts & Clinical Diagnosis (2011)
Nasal proportion. Nostril (ala) width relative to nasal length, plus ala width against the inter-eye distance.
nasal index = ala width ÷ nasal length · ideal ≈ 0.7 — Farkas nasal anthropometric indices

Lighting-agnostic skin

Skin tone, texture, and redness are white-balanced against the sclera (eye white) and measured in CIELAB color space, so they reflect your skin rather than the room light. Brightness is reported only as an exposure indicator, not a quality score.

Fairness across skin tones

Skin scores are computed on white-balanced reflectance (von Kries color constancy against the sclera), so redness, evenness, and texture are read independently of your baseline skin tone. Your tone is reported separately and never changes your scores.

Your baseline tone is estimated with the Individual Typology Angle (ITA°) from white-balanced CIELAB, mapped to an approximate Fitzpatrick band (Chardon / Del Bino convention). It is shown for context and is never an input to any score.

Confidence & honest change

Every score carries a ± band that widens as capture quality and frame-to-frame stability fall — the real sources of measurement noise. A week-to-week change is only reported as real when it clears the combined noise floor of the two scans; otherwise we say “no significant change” rather than presenting noise as progress.

Capture quality

Before a scan is trusted we score the capture on head pose, lighting, focus (variance of Laplacian), framing, centering, neutral expression, and stability. Severe problems (large head turn, very dark, face too small) prompt a retake; softer issues are surfaced as advice so a usable photo is never rejected for a minor reason.

What each block is

Every block of a result is one of four kinds, and is labelled as such. A Measurement is computed from the landmarks and the outline and repeats within the stated band. An Estimate is a score or a figure derived from measurements and reference ranges; it is not a fact about you. An Observation is what the analysis read in the measurements (and, in full mode, the frames); it is never a diagnosis. An Option is something you could do, graded by evidence, with a professional where marked.

Each scan records which scoring rules and which prompts produced it, the model used, and a hash of its inputs, so a result can be explained, reproduced or challenged later. Population claims such as “Top X%” are shown only against an empirical, versioned benchmark that names its population; none is loaded today.

AI analysis agents

When enabled, analysis agents interpret the measurements and write your plan, grounded in the computed numbers and general dermatology/aesthetics guidance. A safety reviewer flags anything that warrants a professional. The AI explains and advises — it does not change the objective measurements.

Population context

Percentiles compare your score to a general-population estimate using a normal model (mean 70, SD 13). It is a heuristic reference for context — not a measured cohort — and is not adjusted for age, sex, or ancestry.

Validation & limits

The geometry engine is covered by self-consistency tests: we feed it faces with known construction (equal facial thirds, a set canthal tilt, perfect mirror symmetry) and verify it recovers those values within tolerance, so the picture you see and the number always agree. What this does not yet include is a validation against instrument-measured ground truth across a diverse cohort — until that exists, scores are cosmetic guidance, not a medical assessment, and norms remain heuristic. Extreme lighting, occlusion, or unusual angles still reduce accuracy; the in-app coach and the capture-quality gate help you avoid them.

← Back to home