How AI face analysis actually works (2026 guide)

Step 1 — Landmark detection

The first thing any face-analysis engine does is find the face and then find points on it. Modern systems mark 68, 140, or even 468 landmarks — the corners of the eyes, the tip of the nose, the outline of the lips, the jawline. These points are the raw geometry every downstream metric is built from.

AscendMe uses a 140-point model because the extra landmarks (midface, cheek apex, mid-brow) matter for the metrics we actually score, but doubling the count to 468 mostly adds noise without moving accuracy on aesthetic measurements.

Step 2 — Normalization

Two photos of the same face at different distances, angles, and lens focal lengths will give completely different landmark coordinates. Before anything can be measured, the engine has to normalize — align the face to a canonical pose, correct for perspective, and scale so that inter-pupillary distance (or another stable reference) is fixed.

This is the step where most consumer 'attractiveness testers' silently fail: they measure raw pixels, which means the same face scores differently depending on whether it was shot on a phone at arm's length or a DSLR at three meters.

Step 3 — Ratios and angles

Once the geometry is stable, the engine computes ratios and angles that map to well-studied aesthetic traits.

• Facial thirds — the vertical thirds from hairline to brow, brow to nose base, nose base to chin
• Facial fifths — five equal vertical fifths across the face at eye height
• Canthal tilt — the angle between the inner and outer corner of the eye
• Midface ratio — cheekbone width vs bigonial (jaw) width
• Chin projection — the horizontal projection of pogonion relative to the lower lip
• Lip fullness and vermillion ratio — upper vs lower lip height

Step 4 — Calibration to a reference population

A ratio in isolation means nothing. A 33% upper facial third is 'ideal' by the classical canon, but the classical canon was calibrated on a narrow slice of European faces. Any honest engine benchmarks measurements against a reference population that matches the user's reported age, sex, and regional background.

Without that calibration, the tool systematically flags typical features of one demographic as 'flaws' — which is both bad science and offensive.

Step 5 — Compositing into a score

Once each sub-metric is calibrated, they get weighted and composited into a headline number (AscendMe uses 0–100; PSL tools use 1–8). The weighting is a modelling choice and the honest thing to do is expose the breakdown, not just the composite.

The composite is a headline. The breakdown is where the roadmap starts — because two people with the same score can have completely different priorities.

Step 6 — Turning the analysis into a plan

The last step — and the one most 'AI face' tools skip entirely — is translating the sub-scores into an ordered set of interventions. Skin quality is very responsive to routine and sleep. Facial fat can be shifted in a quarter. Bone-level structure can't be changed non-surgically at all, and a good report should say so instead of upselling.