How Syva actually works.
We separate what can be measured from what must be judged, and we are explicit about which is which. No single ideal face, no black box — just landmarks, ratios, disclosed benchmarks, and honest confidence.
Geometry is deterministic. Aesthetics is a rubric.
Every geometric metric — canthal tilt, facial thirds and fifths, width-to-height ratios, jaw angles — is derived from the same fitted facial landmarks. Given the same landmarks, the same number comes out every time. There is no model opinion in a distance or an angle; it is arithmetic, and it is fully reproducible.
Your annotated overlay shows precisely which points were used for each measurement, so a result is auditable rather than asserted.
Qualities that can't be measured with a ruler.
Skin quality and overall harmony can't be reduced to a single distance, so they are scored by vision models trained against published norms. These reads are probabilistic, and we treat them that way: each returns a value alongside an explicit confidence.
When lighting, resolution, or pose weakens a read, confidence drops and the report says so — rather than presenting a shaky estimate as fact.
Compared to a reference distribution — and disclosed as such.
A raw ratio means little on its own. To make it interpretable, we place each metric against a normative distribution and report where you fall — for example, a canthal tilt in the 78th percentile of the reference cohort. Percentiles describe typicality within that group; they are not a ranking of people.
We disclose the cohort a benchmark is drawn from, and we deliberately avoid a single universal ideal. The reference is a tool for context, not a target every face should converge toward. Where a norm is sex- or age-specific, the comparison is scoped accordingly.
What this method cannot do.
Credibility means naming the edges of the tool as clearly as its capabilities. Four limits are worth stating plainly.
It estimates 3D structure from 2D images
A photograph flattens depth. We reconstruct plausible geometry from one or more angles, but projection, lens, and pose introduce error. Where an estimate is fragile, we lower its confidence rather than hide it.
Norms are culture-, sex-, and age-specific
Aesthetic reference distributions differ across populations and change over time. We benchmark against a stated cohort and never present a single, universal standard of a face.
Conformance is not the same as desirability
Scoring how closely a feature matches a distribution is a measurement, not a verdict on attractiveness or worth. Many admired faces sit well outside the average — by design.
A score is a starting point, not a diagnosis
Syva describes cosmetic conformance. It is not a medical device and does not diagnose, treat, or assess health. Clinical questions belong with a licensed professional.
See the method applied to a real report layout.