How MaxxLook works — including the parts we get wrong.

Your scores — what they are, and what they are not

For members aged 60 and over, measurements without an age-matched reference use the ages 40–59 reference group. These comparisons are not age-matched and may reflect age differences. Your member percentile still compares you with members of a similar age.

The app gives you an overall score from 0 to 100 and nine area tiles: skin, hair, eyes, brows, face shape, jaw and structure, smile, makeup style, and overall harmony. Every number comes from measurements of your own photos; none is typed in by a person.

Each measurement is placed in a reference range measured from real photographs (see below) and scored by where it falls in that range, from 0 to 100. A measurement exactly at the middle of the range scores 50. For symmetry, tilt and offset measurements — how closely your two sides match, how level a feature sits, how far a ratio is from its target — smaller is scored higher. For the other measurements — widths, heights, ratios, spacing — the score rises with the value, so a wider-than-typical feature scores above 50 and a narrower one below it. A score says where a measurement sits in the range, not whether it is better.

One of those target ratios is the golden ratio: the face’s height-to-width ratio is compared with 1.618, a smaller gap scores higher, and it feeds the face shape tile and overall harmony. It is one measurement among many, not a verdict on its own.

An area score is the average of the scored measurements in it; the smile tile is scored from the lips on your front photo, because the smile photo’s own readings are measured but not yet scored. Your overall score is a weighted average of underlying dimensions built from the same measurements: symmetry and proportion count most (25% each), then balance, meaning closeness to the reference average (15%), then the structure of the jaw, brows and cheeks, and of the eyes, lips and cheeks (10% each). Nose, neck and ear measurements count toward these dimensions even though they have no tile of their own. Overall harmony is the average of symmetry, proportion and balance. A dimension or area with nothing scored is left out and the rest are reweighted, never padded with a made-up middle value.

An area the instrument cannot score from your photos says "unscored" instead of showing a number. Today that always includes hair and makeup style, and it includes skin until a skin reference range exists (see below), so six of the nine tiles carry a number. Hydration is not measured at all, so it never appears as a reading.

The word beside a score is a fixed range on the 0 to 100 scale: Excellent from 85, Good from 75, Fair from 60, and Room to grow below 60. Because a measurement at the middle of its reference range scores 50, a typical face often reads Room to grow — the word points to where small changes show most; it does not say anything is wrong. When your percentile can be shown, it replaces the word on your overall score.

Your percentile is the only comparison with real people: where your overall score sits among other MaxxLook members of a similar age (within five years) and the same gender and ethnicity, counting each member’s latest analysis once. It is withheld entirely until at least 30 such members have a comparable score — a percentile built on three people would be inventing precision we do not have. No member ever sees another member’s scores, photos, or identity.

What a score is not: a judgment of your worth, or a comparison with anyone you know. We will never compare you with a celebrity, a friend, or a partner, and Ask Maxx will not give you a hot-or-not verdict. Behind every score sit the same 170+ individual measurements, each with its unit, and what we would try first.

How the measurements are made

You submit six photos: front, both profiles, both quarter views, and a smile. A face-detection model maps up to 478 landmark points per photo, plus estimated ear and neck points where the model does not have them (these are marked as estimates and weighted lower).

From those points the engine computes 170+ measurements across 12 sections — width and height ratios, symmetry deltas, tilt, spread, golden-ratio deviation, facial-thirds balance, and more. Skin measurements come from pixels: texture energy, pore visibility, redness variation, under-eye shadowing. Every skin metric measures variation within your own face. We never rate skin tone or absolute lightness — that would be rating skin colour.

Some measurements cannot be made from a consumer photo at all. Fine lines, elasticity, hydration, and rosacea assessments are left unmeasured rather than given a fabricated number.

Where reference-group readings come from — and where they don’t

A reference-group reading requires a reference distribution: what this measurement looks like across people of your age, sex, and population group. Ours are measured, not asserted: we ran the exact pipeline that reads your photos over a large, openly licensed face-image corpus (FairFace, CC BY 4.0), kept only near-frontal faces of adults aged 18 to 59, and computed each measurement’s distribution per demographic group. Where a reference exists, your measurement is described by where it sits in that range — the middle reads as typical, not as a passing grade.

Two sections stay measurement-only on purpose. Skin metrics depend on controlled lighting, so a web-photo corpus would give misleading baselines until we derive them from capture-controlled sessions. Smile metrics depend on expression, and the corpus is overwhelmingly neutral-faced. For those you get the exact measured value — not a made-up comparison.

A reading on the high side of typical does not mean "better than most people". And reference-group comparisons against other members are withheld entirely until a demographic cohort has at least 30 completed analyses — publishing a comparison built on three people would be inventing precision we do not have.

Fully automated — no queue, no waiting on a person

Every analysis runs end to end without human intervention: photo quality checks, measurements, reference comparisons, and the final report are all produced by software. There is no review queue and no approval step; your report is released the moment the pipeline finishes.

This is a deliberate trade-off. A human reviewer can catch things software cannot — but a promised reviewer who does not exist would mean reports that never arrive. The pipeline is honest about its own limits instead: measurements it cannot make from your photos are marked as unmeasured in the report rather than guessed, and every automated release decision is recorded with its reason and is auditable.

How the system stays honest without a person in the loop

There is no human correction channel and we do not pretend otherwise. Reliability instead comes from what software can check about itself: every metric carries a measured noise floor, every analysis records which measurements succeeded and which could not be made, and every automated release decision is stored with its reason so it can be audited.

Where we can verify a claim about your photos, we make it. Where we cannot, the report says so — an honest "unmeasured" is worth more than a confident guess, and it is the reason you can trust the numbers we do show.

What we can and cannot detect over time

Between full analyses, weekly skin checks measure the same metrics from a single selfie. Each check is compared against your own previous readings — a change measured against yourself needs no population reference, which is the most defensible number this platform produces.

Every metric carries a minimum detectable change: a noise floor below which we tell you "within noise" rather than showing a trend. Be clear about what that floor is: it is derived from controlled perturbation tests of photographs (re-capture simulation under lighting, angle, and compression changes) plus conservative assumptions — a floor, not a clinical test–retest study. We treat it as provisional, publish how each figure was derived next to where it is shown, and would rather show "no measurable change" than celebrate noise.

Habit insights pair your adherence with skin trends — but they always say correlation is not proof. We would rather under-claim a real effect than teach you to celebrate random noise.

Ask Maxx — grounded in your results

Model-written replies are not switched on yet. Today every Ask Maxx reply is assembled from fixed templates built on your own measurements and plan, and nothing you ask is sent to an outside AI provider. The rest of this section describes the model-written replies that will follow.

Ask Maxx is an AI assistant. Its replies are written by Anthropic’s Claude model from your own results: your scores, area summaries, opportunities, plan, goals, and the styling answers in your profile. It is never sent your photos, age, gender, or ethnicity.

It may only quote numbers that are in your results, and replies that invent a score are thrown away. Questions we recognize as medical, or as a sign of distress, get a fixed, human-written response instead of a model reply. When the model is unavailable, Ask Maxx answers from fixed templates built on your measurements.

Generated images are always labeled

Try On uses a generative image model, through our provider Higgsfield, to render a hair, brow, makeup, or beard look on your own front photo. Every rendering is labeled as a generated image. It shows a look, not a prediction of how you will look, and it never changes your scores. Because your photo leaves our servers to make it, Try On asks for its own permission first.

Try On renders are not switched on yet. Until they are, Try On shows each look’s reference image and sends no photo anywhere.

Before/after pairs are produced by constrained, photometric transforms — smoothing skin texture within the skin mask, subtly lifting under-eye shadow, refining jawline contrast. Each is checked to preserve at least 95% structural similarity with your photo and is labeled "Illustrative concept — not a predicted result". Neither kind of image is a promise of results.

What happens to your photos and data

Analysis photos are encrypted in transit and at rest (AES-256 via Cloudflare R2). Skin-check photos and free-preview photos are measured in memory and discarded — never stored. Biometric consent is explicit, enforced server-side, and withdrawable at any time.

We do not sell biometric data. We do not track you across other companies’ apps. Ask Maxx’s model-written replies and Try On renders will use outside AI providers once they are switched on; today neither is, and nothing is sent to one. Our privacy policy names each provider and exactly what it receives. Account deletion permanently removes your data within 30 days.

To get a copy of your data at any time, email privacy@maxxlook.app from the address on your account.

What this is not

MaxxLook is not a medical or dermatological assessment. It does not diagnose conditions, and it is not a substitute for a physician or dermatologist. It is an appearance-focused coaching tool for adults 18 and older.

If you are struggling with how you feel about your appearance, the resources below are here, always:

• BDD Foundation →
• International OCD Foundation →
• 988 Suicide & Crisis Lifeline (US) →