What An AI Face Report Actually Measures (And What It Doesn't)

1. Symmetry — left-right correspondence after midline alignment

Symmetry is the single most-cited variable in the AI face report category and the one the most tools surface first. The measurement is mechanical — the model finds the facial midline (typically a vertical line through the nasion, the philtrum center, and the chin point), reflects the left half across that line, and computes the average pixel-distance between each landmark on the right side and its reflected left counterpart. The number is normalized — 1.0 means perfect correspondence, lower values mean increasing left-right divergence — and it correlates moderately with perceived attractiveness across the open behavioral literature. The peer-reviewed NIH summary at PMC2781897 (Little, Jones, & DeBruine, 2011) reviews the cross-cultural evidence that symmetry is one of three independent attractiveness predictors alongside averageness and sexual dimorphism, with effect sizes that vary by population and by face. What symmetry does not tell you is whether the asymmetry is fixable — some asymmetry is structural (bone), some is muscular (asymmetric expression habit), some is photographic (head tilt at capture, lighting from one side). A good AI face report flags the symmetry score; a great one separates the structural from the photographic component so users know which lever moves which dimension.

2. Harmony and proportion — golden ratio, midface ratio, eye spacing

Harmony is the proportion-balance layer of an AI face report and it is structurally independent from symmetry. Where symmetry asks "are the two halves matched," harmony asks "are the internal proportions balanced." The measured ratios that cluster around perceived attractiveness in the literature include the upper-to-lower face ratio (forehead height to chin height), the midface ratio (cheekbone width to nose-to-chin length), the eye-spacing ratio (intercanthal distance to eye width — the classical "one eye-width between the eyes" rule), the lip-to-chin distance, the philtrum length, and the golden-ratio composite (a rolled-up score that summarizes how close several adjacent ratios sit to phi, 1.618, which the literature has mixed evidence for as a beauty constant — useful as one signal among several, not a single determining variable). A face can score perfectly on symmetry and below-average on harmony, or vice versa, and the report is more useful when both numbers are surfaced separately rather than rolled into a single composite. The fix paths are different — symmetry is partly capture-corrigible (head angle, lighting), harmony is mostly structural (jaw, cheekbones, hairline framing).

3. Facial width-to-height ratio (FWHR) — the masculinity proxy

FWHR is the bizygomatic width (cheekbone-to-cheekbone distance at the widest point) divided by the upper-face height (typically brow line to upper lip). The ratio is cited heavily because it is one of the most-studied single-number facial metrics in the perception literature — Carre and McCormick (2008) and follow-up work in Proceedings of the Royal Society B linked higher FWHR with perceived dominance, perceived aggression, and behavioral correlates in athletic and competitive contexts. The interpretation is more nuanced than the headline. FWHR varies systematically by sex (men trend higher, women trend lower), by population, and by age. A face report that returns "FWHR: 2.05" without a population-aware comparison set is less useful than one that returns "FWHR: 2.05, 78th percentile for adult men in the model's training distribution." The same number means different things for a 22-year-old and a 55-year-old, and a tool that does not adjust for that is shipping a measurement without context. FWHR is also the dimension where photographic capture matters most — a slight head tilt forward shortens upper-face height and inflates the ratio, and a head tilted back does the opposite. Run two photos and compare to verify the underlying structural number.

4. Jawline angle and taper — the gonial angle and lower-face geometry

The jawline metric in an AI face report is typically two numbers. The first is the gonial angle — the angle at the back corner of the jaw (the gonion), measured between the ramus (the vertical jaw line going up to the ear) and the mandibular body (the horizontal jaw line going forward to the chin). Lower gonial angles (closer to 90°) read as a more square, defined jawline; higher gonial angles (above 130°) read as a softer, rounder lower face. The second is the jaw taper — the rate at which the lower face narrows from the gonial angle to the chin point, often expressed as the angle between the mandibular line and the midline, or as a ratio of bigonial width (jaw at the back) to chin width (jaw at the front). Together those two numbers describe the entire lower-face silhouette without requiring a 3D scan. The fix paths are partially independent. Body fat percentage moves the apparent gonial angle (lower fat reveals the underlying bone angle) without changing the actual bone. Posture and chin tuck shift the angle the camera sees by a measurable few degrees. Beard density changes the visual taper without changing the bone. A good AI face report flags the measurement; a great one separates the corrigible (fat, posture, beard, hair framing) from the structural (bone) so users know what is realistically movable.

5. Skin uniformity — texture variance, redness, blemish density

Skin is the dimension where AI face reports diverge most from each other, because the underlying signal is per-pixel rather than per-landmark and the model has to segment the face skin region first before it can score it. The three load-bearing sub-metrics are texture variance (the standard deviation of luminance in small patches across the cheek and forehead — high variance reads as rough or uneven, low variance reads as smooth), redness or erythema (the average a-channel value in CIELAB color space across the skin region — elevated values indicate inflammation, irritation, or active acne), and blemish density (count of localized features detected by a small object-detector head — pimples, scars, dark spots per square centimeter of skin region). The score is highly capture-dependent. Lighting direction, camera lens, white-balance setting, and even the time of day a photo is taken move the skin score by 5-15 percentile points without any underlying skin change. A serious AI face report flags this — the skin metric is the most useful for tracking month-over-month change in matched lighting conditions, and the least useful for cross-photo absolute comparisons. Single-photo skin scores are directionally informative; treat them as such.

6. Expression metrics — eye-open ratio, lip-corner angle, brow tilt

The expression layer of an AI face report is what distinguishes a structural-only tool from one that can answer photo-decision questions. The three measured sub-metrics are eye-open ratio (vertical palpebral aperture divided by eye width — squinted eyes read as lower scores, wide-open eyes read as higher), lip-corner angle (the angle of the mouth corners relative to the mouth line — upward angles correlate with positive affect perception in the Princeton first-impression literature associated with Alex Todorov's research program), and brow tilt (medial brow height vs lateral brow height — neutral brows read as composed, raised medial brows read as concerned, raised lateral brows read as engaged). The reason these matter is that two photos of the same face with the same structural geometry can score dramatically differently on perception sub-scores like trustworthiness, warmth, or dominance because of expression alone. The honest framing in an AI face report is that expression is the most-controllable variable in the entire breakdown — you cannot change your gonial angle for a photo, but you can change your lip-corner angle in 200 milliseconds. A good AI face report makes this trade-off explicit so users know which lever is one shutter-click away.

What an AI face report cannot measure (the honest section)

This is the section that most face-rating tools quietly skip and that we think is the most important one to publish. An AI face report is a measurement instrument with a known sensor (a 2D image) and a known model class (a dense-mesh landmark net plus a few small classifier heads). Things that fall outside that sensor envelope do not show up in the report, and any tool that pretends otherwise is selling something past what the model can actually do. Four explicit non-measurements follow.

1. Anything off-frame. An AI face report measures the face. It does not measure your voice, your height, your gait, your body composition below the shoulders, your hand geometry, your scent, or any of the cues that real-world first impressions integrate. A 9.5 face report does not predict real-world outcomes if the rest of the off-frame signal is poorly managed, and a 7.0 face report does not preclude them if the rest is dialed in. The face is one channel. The report scores that channel honestly. It does not pretend to predict the rest.

2. Dynamic and motion perception. A face in conversation is processed differently from a face in a still photo — micro-expressions, vocal prosody, eye-contact rhythm, head gestures, and listening cues all carry signal that a single frame cannot encode. The AI face report scores the still. It cannot score how you read in motion, how warm you come across in a five-minute conversation, or whether your laugh registers as genuine. Those variables are real, they matter for real-world outcomes, and they are not in the report. The honest move is to use the report for the photo decisions it can make and to leave the dynamic questions to dynamic measurement (which is its own discipline — Photofeeler-style human ratings from short videos are the closest production proxy, and they are still not what an AI face report does).

3. Context and styling. An AI face report does not know what you are wearing, what your hair looks like out of frame, what the room looks like behind you, or how the photo will be cropped on the destination platform. These variables move first-impression outcomes by margins that often exceed the structural variation the report measures. A face that is a 7.5 in the report can be an 8.5 in a Hinge feed with the right lead photo, the right haircut, and the right styling, and the same face can be a 6.5 in a different feed with worse versions of the same controllable variables. The report is the structural floor; the rest is delta you control.

4. Identity and personality variables. An AI face report measures geometry. It does not measure kindness, humor, intelligence, status, or any of the variables that determine actual relationship outcomes once the photo decision is made. The trap of the category is treating the report as a verdict on the person rather than a measurement of one channel. The healthier framing is the one we use internally — a face report is a photo and grooming triage tool. It tells you which lead photo to use, which haircut to choose, whether to grow or trim a beard, where the asymmetry is photographic vs structural. Past that, the report is silent, because silence is what the model owes you when the question is past its sensor envelope.

Visual scoring vs structural scoring — what each method actually does

Two methodologies dominate the AI face report category. Visual scoring runs the photo through a single end-to-end model trained to predict a human-rated attractiveness score directly — input photo, output number, with the model's internal reasoning opaque. Structural scoring runs landmark detection first, computes named geometric metrics from the landmarks, and rolls those metrics up to a composite using a documented weighting. The two methods answer different questions and have different failure modes — picking the right one for your decision matters.

The takeaway is that a structural face report is more useful when you are about to act — pick a lead photo, choose a haircut, decide whether to grow a beard — because the per-metric breakdown points at the lever. A visual score is more useful when you only want a percentile and you do not need the underlying breakdown. The RealSmile face report ships the structural side because the buyers we serve are about to make a photo decision and need to know which dimension is moving the score.

How to read an AI face report without over-trusting it

Three rules we use internally and recommend to anyone running a face report for a real decision. First, run the same photo twice. A reproducible tool returns the same numbers; a non-reproducible tool has a stochasticity problem you should know about before you act on the output. Second, run two different photos of the same face on the same day. Numbers that move by 5-10 points across photos are normal — that is the photographic component talking. Numbers that move by 20+ points across photos either flag a high-leverage capture variable (lighting, angle, expression) or a tool that is over-fitting to single-photo cues. Third, treat the percentile as a comparison-class anchor, not as a verdict. A 78th-percentile FWHR for adult men is structurally informative; a 78 with no comparison class is decorative. Tools that do not surface what they are comparing you to are doing less work than tools that do.

The trust signals worth checking on any AI face report tool before you act on the output: 38,000+ photos analyzed. Photos auto-deleted within 30 days. 7-day refund. Tools that publish all three plus a methodology page with citations are doing the work; tools that publish a number with no methodology are not. The honest test is whether the tool can answer "why does this number mean what you say it means" with a public document. If it cannot, the number is a marketing widget — useful for entertainment, less useful for decisions.