How every numberis made.
This page is the specification. If a value in a report can't be traced to a row here, that's a bug. Each report is stamped with this version and a SHA-256 fingerprint of its result.
1. Pipeline
Capture gate thresholds
Camera distance is estimated from IOD in pixels and the EXIF 35 mm-equivalent focal length assuming population-mean IOD ≈ 62 mm; below 70 cm we warn, citing Ward 2018 (≈30% nasal-base widening at 12″) and Derakhshan 2024 (12–19% midface stretch). Sharpness is the Laplacian variance of a 256 px face crop; exposure is the fraction of clipped face pixels.
2. Uncertainty
σ per landmark coordinate = 0.0396 × IOD px (Kartynnik 2019). Well-defined points (canthi, iris centres, lip midline, mouth corners, subnasale, nose tip, alar edges, lid margins) get 0.5σ; nasion and sublabiale 0.75σ; silhouette points (zygion-ish, gonion-ish, menton, mesh top, brow) the full σ.
σ(a / b) = (a / b) · √((σa / a)² + (σb / b)²)
σ(angle over baseline L) = σ · √2 / L
Why the intervals are the size they are
Each metric also has a floor (ci_floor) reflecting landmark-definition ambiguity (e.g. canthal tilt ±3°). Ratios of paired facial measures are positively correlated, and independence is assumed, so the intervals are conservative by construction. The published landmark error is dominated by loosely-defined cheek and forehead vertices — which is why silhouette-based metrics (jaw, face width) are Tier B and canthal metrics are Tier A.
3. Metric table
4. Norm provenance
Reference means and SDs are transcribed from the young-adult tables of Farkas et al. (1994; 2005 international study, 1,470 subjects, 25 groups), Jayaratne et al. 2012 (Southern Chinese, n=103, 3D), Al-Sebaei 2015 (Saudi, n=168), and metric-specific sources listed per page.
Values shown with prov. have not yet been audited line-by-line against the primary printed tables. They are close, cited, and marked. The base table lives in scripts/gen-metrics.mjs.
How the pooled population is built
Ratios are derived from a single base table of linear measures so every derived value is internally consistent. The pooled population is an unweighted average across the six groups with between-group spread folded into its SD — wide by construction.
5. Why "ideal ranges" are not shown
Sources and what we do instead
Farkas 1985; Jayaratne 2012; Al-Sebaei 2015 (92% of Saudi men have a nose wider than ICD); Khoshab 2022 (PRISMA review of 27 studies across six ethnic categories: no ethnic/gender group consistently approximates the canons or the golden ratio). Canthion reports each canon as a descriptive check with its conformity statistic — never as a target.
6. Skin
Face crop at 512 px, sRGB → approximate CIE L*a*b* (D65), regions from convex hulls of landmark groups shrunk 15–20% to avoid edges.
Definitions
- Redness = cheek/nose a* − forehead a*
- Under-eye = infraorbital L* − mid-cheek L*
- Texture = mean local σ of L* in 5×5 windows
- Spots = connected components on a*/L* thresholds, size 6–300 px, aspect < 2.5
- Uniformity = σ of regional mean L*
- Shine = fraction of T-zone with L* > 92 and chroma < 12
Erythema is flagged unreliable when mean cheek L* < 45 (melanin dominates a*). All values are relative to this photo's lighting.
7. Morph
Similarity Procrustes alignment of MediaPipe's canonical face mesh to the user's 468 points; target = user + α·(canonical − user); piecewise-affine warp over the official 898-triangle tesselation plus a border ring. α is user-controlled, default 0.25, labelled literally. A symmetric generic face — not a population mean.
8. Protocol grading
Drawn from a fixed library by deterministic triggers on the findings. Two interventions reach grade A: daily broad-spectrum sunscreen (Hughes 2013, N=903, 4.5-year RCT) and topical tretinoin (Weiss 1988 and successors). Mewing is included at grade D so the report says, in writing, why it is not recommended.
Full library (18 items)
- A Daily broad-spectrum sunscreen — source
- A Topical retinoid (tretinoin) — evening — source
- B Body composition and facial adiposity — source
- B Sleep — source
- B Camera distance, height and lens — for photos of you — source
- A Sun avoidance and physical shade — source
- C Brow grooming and hair framing
- C Facial exercise — source
- D Mewing / orthotropics — not recommended — source
- D Chewing gum 'for jawline' — not recommended as a strategy
- B Antioxidant / niacinamide serum (morning, under sunscreen)
- B Barrier-repair moisturiser
- B In-clinic options for texture and pigmentation (chemical peels, laser resurfacing, microneedling)
- C In-clinic options for the under-eye (tear-trough filler, lower blepharoplasty)
- info About surgical and injectable recommendations in general — source
- A The two things that actually keep teeth white
- B Peroxide whitening (strips, trays, dentist-supervised gel)
- B In-clinic options for tooth colour, shape and alignment
9. Determinism
CPU inference is bit-stable on the same device; landmarks are rounded to 0.01 px; all downstream math is closed-form. Each report shows a SHA-256 fingerprint of the numeric result. "Re-run" re-detects from scratch and compares. Across devices, last-decimal differences are possible; the fingerprint is a same-device guarantee.
10. Known limits
- Mesh vertices ≠ anthropometric landmarks. "Zygion" and "gonion" here are silhouette points and are so labelled.
- 2D photogrammetry carries systematic bias vs calipers (~9.3 mm bigonial, ~3.3 mm bizygomatic — Lim 2022) with high repeatability.
- Depth-spanning features (nose width vs ICD, midface) are inflated 12–30% at selfie distance. We estimate distance and warn; we cannot correct.
- Computational photography (skin smoothing, face-slimming, HDR) is undetectable in general and corrupts skin metrics first.
- Population norms are means of specific samples, not of "populations"; the user chooses the label.
- Nothing here predicts what any specific person will find attractive; ~40% of that variance is idiosyncratic (Leder 2016; Germine 2015).
11. Changelog
- 2026.08-r1 — initial public methodology. 18 geometric metrics, 6 skin metrics, 11 bone-level measurements taken as photographic equivalents, 18 protocol items. Norm tables marked provisional pending audit.
- Planned: line-by-line audit of norm tables against Farkas 2005 Table 2 and Farkas 1994; per-population mean shapes for the morph from a consented dataset; controlled-capture longitudinal mode; TrueDepth/WebXR metric capture to unlock profile metrics.