FIELD NOTES / FACE MEASUREMENTS
Understand a symmetry result
Left-to-right balance is one photo measurement, with several sources of variation.
AI-assisted product documentation, checked against our published method and the technical sources below. Diagrams and worked numbers are illustrative, not personal test results or a clinical review.
Start with the camera
A head turn changes how far each side is from the lens. Tilt, uneven light and expression can move the apparent outline or the landmarks even when the underlying face is unchanged. Our method corrects eye-line roll before comparing paired points, but that does not remove every perspective effect. Start with a neutral front-facing image and a consistent camera position when investigating a large difference.
Inspect the approximation
The symmetry score comes from a two-dimensional mirror-error approximation around a visible midline. It is not a medical measurement of anatomical symmetry. Look at the plotted points and check whether hair, shadows or facial expression affect their placement. A local difference can have a straightforward image explanation. The score’s formula and version help you compare like with like rather than relying on a single unexplained number.
Choose a useful next step
Try another neutral photograph taken under similar conditions before drawing a conclusion. Ordinary grooming, hairstyle framing or an expression you like can be explored without treating natural variation as a defect. The tool cannot diagnose sudden physical changes or symptoms. If your concern is a new health symptom, a photo score is not the appropriate way to assess it; seek suitable professional care rather than using this comparison as reassurance.
Separate orientation, pose and landmark error
A facial symmetry analyzer answers a narrower question than whether a face is symmetrical in three dimensions. Here it measures agreement between paired visible points in one image. Use the checklist to decide whether the photograph supports a comparison before interpreting the score.
- Face the camera directly. Eye-line roll correction can level a tilted image, but it cannot reconstruct the hidden side of a turned head.
- Keep the expression neutral. A raised brow or uneven smile can move paired landmarks without representing a persistent structural difference.
- Check both cheek edges and eyes for hair, glare, blur and shadows. Do not infer a personal trait from a point placed on an obstruction.
- Remember that an original and horizontal flip contain the same captured information. They do not form an independent second measurement.
- Read the quality status and raw formula. A higher symmetry score here means smaller normalized mirror error, not greater health or social value.
Worked example: normalize the mirror error
Suppose an illustrative mesh has a cheek width of 200 pixels and a mean mirrored-pair error of 2 pixels. The error ratio is 0.01. The current engineering formula maps it to 95. These teaching numbers are not measurements from a real person and are not a normal-range threshold.
| Illustration | Mean error ÷ cheek width | Formula score |
|---|---|---|
| A | 2 ÷ 200 = 0.01 | 95/100 |
| A scaled ×2 | 4 ÷ 400 = 0.01 | 95/100 |
| B | 6 ÷ 200 = 0.03 | 85/100 |
The score is 100 × (1 − 5 × mean error ÷ cheek width), clipped from 0 to 100. Scaling all input coordinates equally preserves the ratio mathematically. Re-running the detector on a resized image can still move its estimated points, so this example does not promise identical results after image editing.
Run a neutral-photo check
- Repeat the same original file three times and save the quality status, mirror score and method version. Do not treat a rejected run as zero.
- Take another directly front-facing neutral photograph with the same light and camera position. Compare the point placement before judging a score difference.
- If the result varies, list the changed capture conditions. There is no validated cutoff here that can separate natural facial variation from image or detector error.
This check is for a photograph, not symptoms or a diagnosis. You do not need to make both sides look identical. If you simply prefer one hair part or portrait angle, that preference can be useful without translating it into a claim that your anatomy improved.
Use this with the right tool
Sources and scope
These sources explain detection and photography. They do not validate our weights, reference targets, attractiveness labels or styling suggestions. Our own scoring rules are documented on the method page.
- NIST · Perspective distortion
Defines appearance changes caused by insufficient camera-to-subject distance.
- Ward et al. (2018) · Nasal distortion in short-distance photographs
A mathematical camera-distance study, not a validation of this tool or a recommended treatment.
- Google AI Edge · Face Landmarker overview
Documents the landmark model and its outputs; it is not a beauty-rating model.