Where the bands come from. The ideal ranges are not invented. They are fitted on SCUT-FBP5500 — 5,500 frontal faces, each rated by 60 volunteers (330,000 ratings, IEEE ICPR 2018) — restricted to the 2,750 male faces. For each measurement the band is the median and interquartile range of the best-rated 20% of faces, and its importance weight is the effect size (Cohen's d) between the best- and worst-rated fifth. The PSL number is a position on a normal curve, which is how the rating communities define it: 4 sits at the midpoint, one point is one standard deviation, and the band cutoffs are theirs (Sub3 below 1.5, LTN 1.5–3.5, MTN 3.5–5, HTN 5–6, Chadlite 6–7, Chad 7+). That is why PSL 5 means ~84th percentile and PSL 6 ~98th, and why the shares beside each band come out as they do.

Where the PSL scale and the data disagree. The canon weights facial dimorphism heavily. The attractiveness data does not. Of the twelve features the community treats as decisive, five are not supported: wider fWHR, a squarer jaw, a low gonial angle and hooded eyes all rate worse with real raters, and canthal tilt — the single most-cited PSL feature — has essentially no correlation at all (r = −0.01). So this score is honestly "percentile of measured attractiveness placed on the PSL ladder", not a reproduction of a forum verdict. It also tested the canon's other claim — that one bad feature caps the whole read — and rejected it: adding a weakest-link penalty to the model moved held-out r from 0.735 to 0.738, which is noise.

Two numbers. The solid marker is the measured score. The hollow ring is the community score: the measured score plus a bounded adjustment for where the forum rubric and the ratings data disagree. It is anchored rather than independently fitted, because an early version that scored all 32 measurements on the canon's own weights correlated negatively with human ratings (r = −0.20) and put young Tom Cruise at PSL 3.3. The two rubrics agree on 26 of 33 measurements and conflict on seven, so the honest gap is roughly a third of a point for a typical face and at most about one and a half — not the three that version invented. It is still a stated prior, not a validated model: there is no dataset of PSL-labelled faces to check it against.

What that is worth. The per-feature band score agrees with the human mean rating at r = 0.43 on held-out faces. The hand-written rubric it replaced scored r = −0.09, i.e. slightly anti-correlated. The joint model shown on the right reaches r = 0.73 as shipped (0.75 in-fit, before dropping race to a population mean). Two measurements are excluded from scoring because they were measured to be invalid: the upper third (the mesh top is the forehead, not the trichion) and symmetry (which behaves as a template-fit residual, not asymmetry).

Method. MediaPipe FaceLandmarker gives 468 mesh vertices plus 10 iris points in 3-D, averaged over five rotation-jittered passes. Landmarks are converted to one isotropic unit, the canonical 468-vertex face model is fitted (Horn quaternion / Kabsch), that rigid transform is inverted, and the ratios are measured in the canonical frontal plane.