Facial asymmetry
Bilateral asymmetry index
How far each point on one side of the face sits from where its partner on the other side would put it, averaged over every landmark and given as a percentage of the distance between the pupils.
Also searched as face symmetry test, asymmetrical face and facial symmetry.
How to read it
the face is highly bilaterally symmetric at the resolution of the landmark detector.
a typical degree of bilateral asymmetry. Every face is asymmetric; the noise floor shown is how much of this value the detector itself contributes.
above-typical bilateral asymmetry at this resolution. Before reading anything into it: head yaw of even a few degrees inflates this value more than most real asymmetry does. Lee et al. (2025), with 400 raters, found no significant relationship between symmetry and attractiveness once averageness was controlled.
A symmetry score from a photo mostly measures landmark jitter plus head yaw. It is here because people ask for it. The noise floor is printed beside it so it cannot be over-read.
Reference values by population
| Reference population | Female (mean ± SD) | Male (mean ± SD) | n | Source |
|---|---|---|---|---|
| North American White | 3.2 ± 1.2 | 3.2 ± 1.2 | — | Typical range for healthy adults measured with 2D landmarks incl. detector noise. Provisional; not a population anthropometric value. prov. |
| East Asian | 3.2 ± 1.2 | 3.2 ± 1.2 | — | Typical range for healthy adults measured with 2D landmarks incl. detector noise. Provisional; not a population anthropometric value. prov. |
| South Asian | 3.2 ± 1.2 | 3.2 ± 1.2 | — | Typical range for healthy adults measured with 2D landmarks incl. detector noise. Provisional; not a population anthropometric value. prov. |
| African / African diaspora | 3.2 ± 1.2 | 3.2 ± 1.2 | — | Typical range for healthy adults measured with 2D landmarks incl. detector noise. Provisional; not a population anthropometric value. prov. |
| Middle Eastern | 3.2 ± 1.2 | 3.2 ± 1.2 | — | Typical range for healthy adults measured with 2D landmarks incl. detector noise. Provisional; not a population anthropometric value. prov. |
| Hispanic / Latino | 3.2 ± 1.2 | 3.2 ± 1.2 | — | Typical range for healthy adults measured with 2D landmarks incl. detector noise. Provisional; not a population anthropometric value. prov. |
| Not specified — show pooled | 3.2 ± 1.2 | 3.2 ± 1.2 | — | Typical range for healthy adults measured with 2D landmarks incl. detector noise. Provisional; not a population anthropometric value. prov. |
Where these numbers come from
Transcribed from the cited anthropometric literature; SDs of ratios are propagated conservatively (independence assumed). Entries marked prov. await line-by-line audit against the primary tables. See the norm provenance note.
Why every value has an error bar
Every value Canthion reports carries a 1σ uncertainty propagated from MediaPipe Face Mesh's published accuracy (3.96% of inter-ocular distance mean absolute error per landmark, Kartynnik 2019). On a typical photo with 300 px between the pupils, that is about 12 px of wander on every point. Small ratios and angles inherit it, so their intervals are wide. A number without an interval is a number pretending to be more precise than it is.
References
- Lee, Li, Rafiee, Jones & Shiramizu (2025). Further evidence that averageness and femininity, rather than symmetry and masculinity, predict facial attractiveness judgments. Sci Rep 15:5498. doi.org/10.1038/s41598-025-86974-0
- Kartynnik, Ablavatski, Grishchenko & Grundmann (2019). Real-time Facial Surface Geometry from Monocular Video on Mobile GPUs. arXiv:1907.06724. arxiv.org/abs/1907.06724
Common questions
What is facial asymmetry?
How far each point on one side of the face sits from where its partner on the other side would put it, averaged over every landmark and given as a percentage of the distance between the pupils. Shown alongside the detector's own noise floor.
How is facial asymmetry measured from a photo?
It is computed from the face-mesh landmarks (midline): mean_i |p_i − reflect(p_partner(i))| / IPD × 100, midline fitted through midline landmarks. Every value carries a 1σ error bar propagated from the landmark detector's published accuracy, with a floor of ± 0.8, so a small difference inside the interval is not a difference.
What is a normal facial asymmetry?
Across the reference populations the anatomical norm is the same: 3.2 ± 1.2 for women, 3.2 ± 1.2 for men. This value is provisional and awaits line-by-line audit against the primary tables. Your own value is placed on the distribution you choose, as a percentile with its error bar.
What does a low or high facial asymmetry mean?
Below the mean: the face is highly bilaterally symmetric at the resolution of the landmark detector. Above the mean: above-typical bilateral asymmetry at this resolution. Before reading anything into it: head yaw of even a few degrees inflates this value more than most real asymmetry does. Lee et al. (2025), with 400 raters, found no significant relationship between symmetry and attractiveness once averageness was controlled.
How reliable is facial asymmetry from a photograph?
A symmetry score from a photo mostly measures landmark jitter plus head yaw. It is here because people ask for it. The noise floor is printed beside it so it cannot be over-read.
Your symmetry is measured on your own photo, with the detector's noise floor shown beside it, free.
Related measurements
The rest of the whole face set, then the nearest measurements elsewhere on the face. Every one is computed on the same photograph, in the same run, with its own error bar.
Also: how bone-level measurements are taken from a photo · methodology · glossary · all 35 insight pages