Face Symmetry Test

Stages run in a fixed order, and each one has a job that the next depends on.

Correct head pose first. Raw coordinates are useless until tilt comes out, because a rotated head reads as lopsided even when it isn't. Roll is estimated from the line joining the outer eye corners, and the whole mesh derotates until that line lies flat. Yaw gets estimated from the ratio of left and right facial widths. Past a small threshold the image is rejected rather than corrected, since rotation about the vertical axis destroys information that cannot be recovered.

Fit the midline second. The facial axis is not the vertical line through the nose tip, which is itself often deviated. It is fitted by least squares as the axis minimizing total squared distance between every landmark and the mirror image of its partner. The fit runs over the full set of bilateral pairs: outer and inner eye corners, alar rims, mouth commissures, gonial angles, brow peaks and temple points.

Measure paired deviation third. Each left landmark reflects across the fitted axis, and the Euclidean distance to its right counterpart becomes that pair's raw deviation in pixels.

Normalize and score last. Pixel figures mean nothing alone, since a larger photograph produces larger numbers. Every deviation divides by interpupillary distance, the standard scale reference for facial metrics. The mean across all pairs then maps onto a 0 to 100 curve, where zero deviation gives 100. The same photograph processed twice returns the same figure, and that figure falls into one of four bands.

Bands describe populations, not people. A number is a geometric fact about one photograph, in the same category as a shoe size, and the band carries no ranking with it. The rows underneath the total say more.

Regions rarely agree with each other. Eyes commonly pair closely while the jaw does not, or the mouth turns out to be the only area contributing meaningful deviation. Knowing which area drives the aggregate tells you something the aggregate hides.

Most people also score higher across the upper face than the lower. The orbital region develops under the constraint of the eye and tends to pair tightly, while the mandible varies more and responds to habitual use. Both readings collapse, though, if the photograph itself is crooked.

Photography corrupts this measurement more easily than any other on the site, because the algorithm cannot separate a genuinely uneven face from an even one shot badly.

1. Square yourself to the camera. Yaw is the fatal error. Five degrees of turn shortens one half in projection and can cost several points. Look at the lens, not at the preview beside it.
2. Set the camera at eye level. Pitch tilts the apparent horizontal of every pair at once.
3. Light both sides equally. Directional light darkens one half, and landmark detection loses precision in shadow. Two sources, or one large diffuse window facing you, does the job.
4. Hold a neutral, relaxed expression. Expression is uneven in almost everyone, and a half-smile dominates the mouth region.
5. Clear both temples of hair. An uneven parting registers as facial difference.
6. Leave the image unmirrored. Front cameras flip by default. Flipping changes nothing in the score, but it confuses you about which side is which.

Even a perfectly captured photograph leaves several questions permanently out of reach.

It cannot tell you whether you're attractive. Perceived attractiveness draws on skin quality, expression, proportion, coloring and dozens of other signals, and the measured contribution of correspondence within that mix is small.

It cannot separate structural difference from postural or photographic difference. A face reading 88 might carry a genuinely different jaw on each side, or might have been shot with a two-degree turn. Three photographs compared against each other is the practical test. Tight clustering points to structure.

It cannot detect anything medical. Facial unevenness is ordinary variation in almost every case, and this page has no capacity to tell ordinary variation from anything else. A sudden change, particularly one appearing over hours or days or arriving with weakness, needs immediate medical attention rather than a web tool.

And it cannot tell you to be more even. Correspondence is a property faces hold in degrees, the way hands hold varying finger lengths. The measurement is interesting. The idea that a higher number means a better face has no support in the geometry, and this tool does not make it.

Correspondence, not quality. The test derotates the mesh, fits an axis by least squares, reflects every bilateral pair and scales the result against interpupillary distance, then places the mean on a 0 to 100 curve. The bands describe populations rather than people, the regional rows carry more information than the total, and yaw, pitch, directional light, expression and hair are the five things most likely to corrupt a reading. Mild asymmetry stays the human default throughout, and no score here says anything about health, worth or beauty. Run the face symmetry test three times and read the regions.