Consistency Test

Is RealSmile actually accurate?

We scored the same face from three different photos each. If the AI was random, the scores would jump around wildly. They don't — they land within a few points of each other.

Timothée Chalamet

Actor, measured across 3 red-carpet photos

Variance

±3 pts

Front, soft light

82

¾ turn, natural light

84

Stage lighting

81

Average: 82/100 · all three photos within 3 points.

Zendaya

Actor, measured across 3 editorial photos

Variance

±3 pts

Studio front

89

Street / daylight

87

Vogue cover

90

Average: 89/100 · all three photos within 3 points.

Henry Cavill

Actor, measured across 3 photos

Variance

±3 pts

Clean-shaven

86

Stubble

87

Beard

84

Average: 86/100 · all three photos within 3 points.

Anonymous user A

Verified paid scan, 3 selfies same day

Variance

±3 pts

Morning, bathroom light

67

Midday, window light

70

Evening, overhead

68

Average: 68/100 · all three photos within 3 points.

Why this works

RealSmile measures 17 geometric ratios (canthal tilt, FWHR, golden ratio, lower-third proportion, etc.) — not vibes. Ratios are stable across photos because they come from the distance between facial landmarks, which don't change based on lighting or mood. Free tools like UMax use overall-vibe CNNs that swing ±15 points photo-to-photo. We don't.

Methodology

How a photo becomes a score

Four stages. No magic, no black box.

01 · Input

Photo upload

Front-facing image, neutral expression preferred.

→ JPG/PNG normalized to a working resolution before analysis.

02 · Detect

Landmark detection

468 facial points mapped to eyes, nose, jaw, mouth, brow.

→ Each point is an (x, y) coordinate on the face mesh.

03 · Measure

Geometric ratios

FWHR, canthal tilt, gonial angle, phi, palpebral fissure.

→ Distances between landmark pairs become unitless ratios.

04 · Output

Score + composite

Per-metric scores roll up into a single composite read.

→ Same ratios + same weights = same score across photos.

Every metric, in the open

How each metric is measured

All 17 metrics, how each one is computed from the facial landmarks, the range it's scored against, and the research behind it. Competitors charge for this breakdown — here it's free.

Facial Symmetry

Mirror-distance of paired landmarks (jaw, eyes, brows, mouth corners) against the nose-to-chin midline.

Ideal92–100% · population average ~85%

SourceGrammer & Thornhill, 1994 · Evolution & Human Behavior

Canthal Tilt

Angle of the line from each eye’s inner to outer corner, relative to horizontal.

Ideal4–7° positive (upward outer corners)

SourceBashour, 2006 · Annals of Plastic Surgery

Facial Width-to-Height Ratio (FWHR)

Cheekbone-to-cheekbone width divided by brow-line-to-upper-lip height.

IdealMen 1.85–2.1 · Women 1.65–1.85

SourceLefevre et al., 2014 · PLOS ONE

Facial Thirds Balance

Brow→nose-base and nose-base→chin segment heights, scored against three equal thirds.

IdealClosest to 33% / 33% / 33%

SourceJefferson, 2004 · British Journal of Orthodontics

Eye Spacing

Inter-eye (inner-corner) gap divided by average eye width.

Ideal0.9–1.1 (≈ one eye-width between eyes)

SourceFarkas et al., 1985 · Anthropometry of the Head and Face

Jawline Angle

Angle of the jaw edge from the chin point up toward each gonial corner.

IdealMen 44–54° · Women 50–62°

SourceRexbye et al., 2006 · Age and Ageing

Nose Proportion

Nostril-to-nostril nose width as a percentage of cheekbone face width.

Ideal22–26% of face width

SourceFarkas, 1994 · Anthropometry of the Head and Face

Lip Ratio

Upper-lip height divided by lower-lip height (slightly fuller lower lip preferred).

Ideal0.5–0.8 upper:lower

SourcePopenko et al., 2017 · JAMA Facial Plastic Surgery

Midface Ratio

Eye-center-to-mouth-center distance divided by total face height.

Ideal0.42–0.48 (compact midface)

SourceMarquardt, 2002 · Journal of Clinical Orthodontics

Eye Shape (Hunter Eye Index)

Eye width divided by eye-opening height — higher reads more narrow/“hunter,” lower reads rounder.

Ideal2.8–3.5 width:height

SourceWick et al., 2008 · Ophthalmic & Physiological Optics

Brow Arch

Peak-of-brow rise divided by brow span (curvature of each eyebrow).

IdealMen 0.15–0.28 (flatter) · Women 0.25–0.40 (arched)

SourceRosen et al., 2015 · JAMA Dermatology

Philtrum Ratio

Nose-base-to-upper-lip distance as a fraction of total lower-face height (an aging marker).

Ideal0.26–0.35 of lower face

SourceSarnoff & Gotkin, 2012 · Journal of Drugs in Dermatology

Chin Proportion

Lower-lip-to-chin-tip height as a fraction of lower-face height.

IdealMen 0.32–0.42 · Women 0.28–0.38 of lower face

SourceMendelson & Wong, 2012 · Aesthetic Surgery Journal

Brow-Eye Proximity

Brow-to-eye vertical gap divided by face height (closer = more prominent brow ridge).

IdealMen 0.03–0.07 · Women 0.05–0.09

SourceWhitfield & Jordan, 2009 · Perception

Jaw Taper (V-Shape)

Jaw width divided by cheekbone width — lower means a more V-tapered face.

IdealMen 0.72–0.85 · Women 0.62–0.75

SourceFink et al., 2006 · Perception

Golden Ratio Score

Proximity of face height-to-width, midface, and orbital ratios to phi (1.618), averaged.

Ideal80–100 / 100 (closer to phi = higher)

SourceMarquardt, 2002 · Journal of Clinical Orthodontics

Orbital Tilt Symmetry

Left-eye canthal tilt vs right-eye canthal tilt — how closely the two eyes match.

Ideal85–100% match (most faces vary 5–10°)

SourceLittle et al., 2011 · Proceedings of the Royal Society B

No thumb on the scale

Every score is derived from fixed geometric ratios and published research, scored against ranges calibrated to human-panel attractiveness data — not tuned to flatter you. Two faces with the same measurements get the same number. A RealSmile score is one signal — geometric structure from a single photo — not a verdict on you.

Limitations

What this AI cannot measure

Geometry from a 2D photo has hard boundaries. Here's where they are.

Minor

Skin texture nuance

Pore-level detail, micro-blemishes, and skin quality aren't encoded in landmark geometry.

Moderate

3D depth from a 2D photo

Projection from one angle estimates depth; it does not measure it directly. Profile differences may be flattened.

Large

Voice, charisma, energy

A still photo carries zero information about how you sound, how you hold attention, or how you make a room feel.

Large

Movement and expression dynamics

Smile asymmetry in motion, eye contact, micro-expressions — none of that survives a single frozen frame.

Moderate

Body proportion context

Height, frame, posture, and body composition all shape perceived attractiveness and aren't in a face-only crop.

Minor

Cultural beauty preference variance

Geometry is universal; preference isn't. Different regions and subcultures weight the same features differently.

Honest take

A RealSmile score is one signal — geometric symmetry from a single photo — not a verdict on you as a person. Treat it as a baseline you can improve with grooming, lighting, and angles, not as a number that defines your worth.

Scan your own face →

Free score. Full report $14.99.