← Back to app

How Desvela Works

Our measurement methodology, sources, and known limitations

What we measure

Desvela analyzes your face across five independent zones. Each zone produces its own score from 0–100. There is no composite “attractiveness score” — we believe a single number destroys the diagnostic signal that makes analysis useful.

Zone 1 — Foundation: Bilateral facial symmetry. Symmetry is one of the strongest cross-cultural markers of perceived attractiveness, linked to developmental stability.

Zone 2 — Proportions: Spatial relationships between features — the distance ratios that determine whether a face “looks balanced.” Based on the 36/46% golden ratios identified by Pallett, Link & Lee (2010).

Zone 3 — Sexual Dimorphism: How strongly your face signals femininity (or masculinity). Measured through jaw taper, chin angle, eye geometry, and face width-to-height ratio. Low dimorphism scores are often the single biggest driver of unfavorable attractiveness perception.

Jaw Taper vs Gonial Angle: Jaw Taper measures the frontal contour angle of the jaw — how sharply the jawline tapers from the cheekbones to the chin. This is measured from a front-facing photo using MediaPipe landmarks at the infraorbital, gonion, and chin points. This is different from the clinical gonial angle, which is measured from a lateral (side-view) cephalometric X-ray and typically produces values around 120–130°. Our frontal measurement produces values around 155–165° because the geometry of viewing the jaw from the front is fundamentally different from viewing it from the side. We use “jaw taper” rather than “gonial angle” to avoid implying equivalence with clinical measurements. Both measure the same anatomical structure — the mandibular angle — but from different views, producing different numbers. Neither is wrong; they are different measurements.

Chin Angle (Gonion Divergence): Chin Angle measures the gonion divergence angle — the angle at the chin formed by lines to each jaw corner. This is documented in cephalometric literature as Gol-Me-Gor (left gonion → menton → right gonion). A smaller angle indicates a sharper, more V-shaped jawline. A larger angle indicates a wider, rounder lower face. This metric is distinct from Jaw Taper, which measures the angle at the jaw corner rather than the chin. Reference: Xiao et al., “Three-Dimensional Analysis of Mandibular Angle Classification and Aesthetic Evaluation of the Lower Face in Chinese Female Adults,” J Craniofac Surg, 2018.

Zone 4 — Feature Detail: Granular feature-level metrics including lip proportions and philtrum-to-chin ratio. This zone delivers the most specific, actionable recommendations.

Zone 5 — Framing: Hair, profile, and accessory optimization. Coming in a future update.

How we measure

We use Google’s MediaPipe FaceLandmarker to detect 468 facial landmarks in real time. All processing happens on your device — your photo never leaves your phone.

MediaPipe places landmarks on your skin surface, not on bone. This is an important distinction: clinical cephalometric analysis uses X-ray imaging to measure skeletal geometry directly. Consumer apps — including ours and every competitor — measure soft-tissue surface positions. Our angles and ratios are calibrated specifically for MediaPipe’s output, not copied from clinical literature values.

For example, our jaw taper baselines are set at 161° (the observed MediaPipe mean), not the 120–130° you’ll find in cephalometric textbooks. Those textbook values describe skeletal angles measured from lateral X-rays. MediaPipe measures the frontal soft-tissue contour angle, which is consistently ~30° more obtuse.

Personalized baselines

Every metric is scored relative to a personalized baseline that you select. A jaw taper of 158° may be average for one population group but a significant deviation for another.

You choose your comparison reference when you start the analysis. This selection determines which baseline means and standard deviations are used to score your measurements. We currently support baselines for Caucasian, Northeast Asian, Southeast Asian, South Asian, African, Middle Eastern, Latino, and Mixed backgrounds.

Our baselines are calibrated from a combination of observed measurements across our test dataset and directional patterns confirmed by peer-reviewed anthropometric literature. When our sample size for a group is small, we widen the standard deviation to avoid over-penalizing faces that fall outside a narrow observed range.

Our baselines have been validated against 424 faces from the Chicago Face Database (Ma, Correll & Wittenbrink, 2015), a peer-reviewed academic face stimulus set with independent attractiveness ratings across Asian, Black, Latino, White, Indian, and Multiracial populations.

Our scoring formula

Each metric is scored using a Gaussian decay function that penalizes deviation from the baseline mean:

base_score = 100 − (deviation / mean × 200)
ethnic_adjustment = 0.3 + 0.7 × exp(−0.5 × z²)
final_score = base_score × ethnic_adjustment

The 0.3 floor ensures that even extreme outliers retain a minimum score rather than being zeroed out. The Gaussian adjustment means small deviations are barely penalized while large deviations are progressively more costly.

Zone scores aggregate their constituent metrics using a weighted formula: 70% average of all metrics + 30% lowest individual metric score. This reflects how human perception works — the weakest feature dominates the overall impression more than the average.

What we can’t measure

Transparency about limitations is as important as the analysis itself.

Symmetry scoring hidden: Zone 1 symmetry scores are not displayed to users because MediaPipe’s landmark detection does not reliably measure bilateral asymmetry under real-world selfie conditions. The model smooths landmarks toward symmetry, producing inflated scores on studio photos and hallucinating severe asymmetry on phone selfies due to uneven lighting, hair occlusion, and JPEG compression. Validated against 424 CFD faces (r=0.05 correlation with attractiveness) and 22 real selfies. A skin homogeneity metric is planned to make Zone 1 meaningful.

Hairline detection: Landmark 10 (our upper reference point) sits on the upper forehead, not at the actual hairline. This means faces with high or receding hairlines may score normally on vertical proportion metrics even when the forehead is visually dominant.

2D frontal only: We analyze a single front-facing photo. Profile analysis, true skeletal angles, and three-dimensional projection cannot be measured from a 2D frontal image. Features like nasal projection and the profile E-line require side-view analysis, which is planned for a future update.

Skin analysis: We do not currently measure skin texture, pigmentation uniformity, or skin health. Zone 1 runs on symmetry alone until skin analysis is implemented.

Male calibration: Our initial release is optimized for female facial analysis. Male baselines exist but have not been validated to the same standard.

Sources

  1. Pallett PM, Link S, Lee K. "New 'golden' ratios for facial beauty." Vision Research, 2010. — Basis for 36/46% golden ratio metrics.
  2. Rhodes G. "The evolutionary psychology of facial beauty." Annual Review of Psychology, 2006. — Foundational research on symmetry and attractiveness.
  3. Farkas LG et al. "International anthropometric study of facial morphology." Journal of Craniofacial Surgery, 2005. — Cross-ethnic facial proportion norms.
  4. Le TT et al. "Proportions of the aesthetic face in Vietnamese-Americans and North American Caucasians." Archives of Facial Plastic Surgery, 2002. — Southeast Asian vs Caucasian proportional differences.
  5. Park DH et al. "Anthropometric analysis of the Korean face." Journal of Craniofacial Surgery, 2020. — East Asian canthal tilt and proportion data.
  6. Wen YF et al. "Cephalometric norms of Chinese adults." Journal of Dental Research, 2015. — East Asian FWHR and jaw taper norms.
  7. ICD pooled analysis — Intercanthal distance across populations.
  8. QOVES Studio methodology — Computational aesthetics framework using cephalometric data.
  9. Ma DS, Correll J, Wittenbrink B. "The Chicago Face Database: A free stimulus set of faces and norming data." Behavior Research Methods, 2015. — 424-face validation set with cross-ethnic attractiveness ratings.
  10. Xiao Y, Luo M, et al. "Three-Dimensional Analysis of Mandibular Angle Classification and Aesthetic Evaluation of the Lower Face in Chinese Female Adults." J Craniofac Surg, 2018. PMID: 29762450. — Gonion divergence angle (Gol-Me-Gor) and mandibular classification.

Our commitment

Desvela is designed as a diagnostic tool for personal exploration, not a judgment of worth. Beauty has subjective and cultural dimensions beyond what any measurement can capture.

We publish this methodology page because we believe transparency earns trust. If you have questions about our approach or find errors in our methodology, we welcome feedback.