Beauty apps give you a number.
This measures the number.
We analysed synthetic faces with Perfect Corp's YouCam engines — the same ones licensed by 800+ global beauty brands. Then we generated 81 plausible re-captures of each face for a single API unit, relit the original, and ran the identical analysis on every sampled variant. What survives that is worth trusting. Most of it doesn't.
Every subject is synthetic and went through the identical pipeline. Switching costs nothing — all three dossiers are pre-computed.
Subject A · synthetic subject · measured with YouCam Face Attribute & Ratio Analysis
That is the kind of number a face-rating app would hand you, and stop.
The same instrument, three faces.
Every subject below went through an identical pipeline. If a metric were simply reliable or unreliable, these rows would match. They do not — so the trustworthiness of a beauty score is a property of the photograph, not of the metric.
| METRIC | SUBJECT A | SUBJECT B | SUBJECT C |
|---|---|---|---|
| redness | SATURATED±0.0 | TRUSTWORTHY±2.2 | BORDERLINE±3.1 |
| oiliness | NOISE±8.1 | NOISE±5.8 | NOISE±10.6 |
| radiance | BORDERLINE±4.9 | BORDERLINE±2.6 | BORDERLINE±4.8 |
| texture | BORDERLINE±4.8 | NOISE±9.5 | NOISE±14.2 |
All subjects are synthetic, generated for this project. No real person is scored against a beauty canon here.
Run it on a real photograph
The gallery above is pre-computed, so browsing it costs nothing. This lane makes live API calls, so it is gated by a key — judge keys ship with the hackathon submission notes, one run each. A durable spend floor, read from the real API balance before any paid call, backstops the budget either way. Your image is held in memory for the length of the request and is never stored. The live run uses the relighting axis; the 81-frame capture axis takes longer than a serverless function may run, so it is pre-computed above.