How FaceAuraAI Works
FaceAuraAI estimates face shape from facial landmarks and published geometric rules; results are styling estimates, not identity or medical facts.
Quick answer
Published rules, re-checked by script.
FaceAuraAI finds 468 landmarks on your face, measures widths, lengths and angles, and scores them against published rules for seven face shapes. Each result carries a confidence label based on your photo's quality, and every recommendation names its measurement. Your photo isn't stored. The rules are deterministic, and a script re-checks them on 1,003 fixed portraits.
Check your shape from a photoFrom photo to recommendation
- Your photo is resized and stripped of location and camera metadata in your browser, where supported.
- Landmarks are found in your browser, or, if it can't, on our server (currently by Amazon Rekognition).
- Measurements: widths, lengths and angles, normalised.
- Result and confidence: scored against seven shape rules, with a confidence label and reasons.
- Recommendations: hair, frames and beards, each citing its measurement.
Landmarks and the ratios we measure
MediaPipe Face Landmarker places 468 points on your face; the rules use twelve.
The landmark points behind each measurement
| Measurement | Landmarks | In plain words |
|---|---|---|
| Length ratio | 10–152 ÷ 234–454 | Mesh's forehead top (usually below the hairline) to chin, over width between the zygia (outer cheekbone points) |
| Jaw ratio | 58–288 ÷ cheekbone width | Width between the gonia (jaw corners) |
| Forehead ratio | 103–332 ÷ cheekbone width | Upper forehead width |
| Temple-to-jaw ratio | 127–356 ÷ 58–288 | Temple over jaw width |
| Jaw angle (∠jaw) | at 58 and 288, from 234/454 to 152 | Average jaw-corner angle |
| Chin angle (∠chin) | at 152, from 172 to 397 | How pointed the chin is |
Why photo size and distance don't matter
Coordinates are rescaled so width and height count equally, so the camera can't stretch your face, and ratios and angles don't change with your face's size in the frame.
The seven published face shape rules
These are the rules as the classifier runs them. Conditions add or remove points, and the highest score wins: each rule is a scoring condition, not a pass/fail line. Length, jaw, forehead and temple-to-jaw are the ratios above.
| Shape | Scores when | Adjustments |
|---|---|---|
| Oval | length 1.18–1.23, with diamond and triangle under 50 | + temple-to-jaw 1.13–1.18 · + ∠jaw ≥ 136° · + forehead < 0.70 with jaw ≤ 0.89 (also lowers square) · − ∠jaw ≤ 136.2° with jaw ≥ 0.89 and ∠chin ≥ 115° |
| Round | length ≤ 1.15, ∠jaw ≥ 138° | more if shorter · + ∠chin 105°–125° |
| Square | length ≤ 1.25, ∠jaw ≤ 136.5° | more the tighter ∠jaw · + jaw ≥ 0.885 · + ∠chin ≥ 113° · − length over 1.25 |
| Heart | ∠chin ≤ 100° | + temple-to-jaw ≥ 1.14 · + temple-to-jaw ≥ 1.15 with ∠chin ≥ 115° and forehead ≥ 0.70 |
| Diamond | forehead ≤ 0.71, ∠chin 115°–122°, ∠jaw 133°–137°, length ≤ 1.15; or ∠chin ≤ 110°, forehead ≤ 0.73, jaw ≤ 0.89, length ≥ 1.18 | — |
| Oblong | length ≥ 1.25, ∠chin ≥ 105° | more if longer · + jaw ≥ 0.87 · + ∠jaw 135°–138° |
| Triangle | forehead ≥ 0.74, ∠jaw ≥ 144°; or ∠jaw ≥ 137°, jaw 0.88–0.90, ∠chin 112°–117°, length 1.17–1.20, forehead ≤ 0.69 | — |
Diamond scoring 100 or more sets square to zero; triangle scoring 100 or more sets heart to zero. A close runner-up is named too (the blend). If no rule scores, the result is oval with Lower confidence.
Body shape: three ratios
The body shape detector measures shoulders (pose landmarks 11–12), hips (23–24, scaled ×1.52 to the outer hip) and waist, from a segmentation mask on the row 58% of the way from shoulders to hips. Ratios match the nearest of five silhouette prototypes, on your device.
How confidence works
Every face read gets High, Medium or Lower confidence from four signals: lighting and sharpness, head angle, how many landmarks were measured rather than estimated, and how close the next shape scored. Reasons come in plain words, and the band is never shown in red: Lower confidence is an honest result, not an error. A photo whose quality can't be measured is flagged and never labelled High confidence. The number beside the label shows how strongly the winning shape scored, not the chance it is right.
Re-run our rule check
What the check shows, and what it doesn't
We assembled a fixed set of 1,003 single-subject, front-facing portraits from the FairFace and CelebA-HQ validation sets, keeping photos with one face, a near-frontal pose and a complete, unobscured landmark mesh. On the cached landmarks, the rules returned the same label for all 1,003 (macro F1 1.000; 95% Wilson interval on the agreement rate 99.62%–100.00%). The rules assigned those labels in the first place, so this shows the method is deterministic and reproducible. It doesn't show how often a person would agree with a label, which is why we publish no accuracy figure (quality and evaluation explains).
How to reproduce it
Researchers and journalists can request access to our code repository through the contact page, then run from its root:
pnpm tsx packages/inference/scripts/benchmark-1000.ts
pnpm tsx packages/inference/scripts/benchmark-body-1000.tsLandmark caches are committed with the scripts, so the check repeats identically; source photos are not redistributed. The body set holds 1,000 full-length pose sets from Fashionpedia and VITON-HD, re-labelled by the current rules on 27 October 2026; a re-run reproduces all 1,000 (macro F1 1.000; 95% Wilson interval 99.62%–100.00%): again consistency, not accuracy.
Your photo
Your face photo is sent once and never stored. Your browser resizes it and removes location and camera metadata where supported, then sends it encrypted for quality checks. Landmarks are found in your browser, or on our server (currently by Amazon Rekognition) if it can't. The photo is held in memory for that request only, never written to disk or a database, and never used to train models. We keep the measurements, in a profile you can export or delete. The body shape detector doesn't upload its photo. See the privacy policy.
What a result can't tell you
- A result is a styling estimate, not medical, dermatological or identity information, and not an attractiveness score.
- One photo, one read: light, angle, expression, hair and glasses move landmarks.
- Landmark 10 sits below the hairline, so detector length ratios run lower than tape readings; the face shape calculator uses its own patterns.
- The seven face shapes are families with blurred edges.
We never score, rate or rank faces or identify anyone. Age, gender and emotion tools are separate and feed nothing into style advice. Your photo isn't kept after its read.
Statistics and citations
No production shape counts are published yet; face shape statistics will show them, with a sample window and stated biases, when they are. Journalists: cite this page, quality and evaluation and the press kit.
Method changes
Each method change is dated here.
Current versions: geometry 2026.09.1, catalogs 2026.09.4, recommender 2026.09.5. Saved looks made with an earlier version are marked, so you know when to compose a new one.
- Published the rule-check script and landmark caches; documented what the check does and doesn't show.
- Clothing recommendations ranked for your body silhouette, occasion and climate.
- Face symmetry, golden ratio, eye and lip tools report measurements only; every face score and rating removed.
- Colour palettes resolved to twelve sub-seasons; colours, pairs and beard styles ranked with reasons.
- Confidence comes only from measured photo quality; a photo whose quality can't be measured is flagged and never labelled High confidence.
Frequently asked questions
Is a FaceAuraAI result a medical or identity assessment?
No. Results are styling estimates from one photo's geometry. They are not medical, dermatological or identity facts, are never used to identify anyone, and never judge how anyone looks.
Why can two photos of me give different results?
Lighting, head angle, expression, hair and glasses move the landmarks, and a face between two shapes can tip either way. A straight-on photo in even light, hair back, glasses off, reads most consistently.
Is my photo stored or used to train models?
No. Your face photo is held in memory for one read, then discarded: never written to disk or a database, never used to train models. The body shape detector doesn't upload its photo.
What does the confidence label mean?
It shows how much to trust this read, judged from lighting, sharpness, head angle, landmarks measured directly, and how close the next shape scored. The number beside it shows the winning shape's score strength, not the chance it is right.
Why don't you publish an accuracy percentage?
A single accuracy figure would depend on photos we haven't seen and would overstate what one number can tell you. Our rule check shows consistency, not agreement with people. Each result carries its own confidence label and reasons.
Can I check your rules myself?
Yes: the rules are in the table above, and the classifier is deterministic. To re-run the script on the 1,003 cached landmark sets, request repository access through our contact page.
Which service processes my photo?
Your face photo goes once to our server for lighting and sharpness checks and is held in memory for that request only. Landmarks are found in your browser first, and on our server (currently by Amazon Rekognition) only if it can't. Body shape photos aren't uploaded.
