Why is my rating inaccurate?
Where the AI goes wrong, how to check your landmarks, and how to fix them manually.
If a score looks wrong, it is almost never the formula. The AI places dozens of landmarks automatically, and it can be a few millimetres off. Those millimetres move ratios, and the ratios move your rating. The fix is to adjust the points yourself — and if you want full accuracy, use Manual review instead of trusting the AI placement.
Why is my rating inaccurate?
Automatic landmark detection is fast and usually close. Close is not good enough for measurements that turn on a millimetre. Hairlines get placed on the hair instead of the skin, jaw angles drift into the neck, chins sit on a shadow. One bad point can throw several ratios at once, and a point that is off on one side but not the other looks like asymmetry you do not have.
The maths itself is deterministic: given a set of coordinates, the same scores come out every time. When a rating looks off, the inputs were off — the photo, or the points the AI chose. Both are yours to fix, and fixing them is expected, not a workaround.
Where the AI goes wrong
Our detection model places dozens of named points on each photo. It is trained on a lot of faces, but a handful of points are consistently harder than the rest, and small errors on them have outsized effects:
| Landmark | Typical error | What it moves |
|---|---|---|
| Trichion (hairline) | Placed on the hair, not the skin edge; lost under a fringe | Top third, facial thirds, face height ratios |
| Gonion (jaw angle) | Drifts up toward the ear or down into the neck on soft jaws | Bigonial width, jaw frontal angle, gonial angle, lower third |
| Menton / soft-tissue chin | Sits on the shadow under the chin rather than the true lowest point | Lower third, face height, chin ratios |
| Zygion (cheekbone) | Hard to see on lean or full faces; placed too high or too far in | Face width ratios, cheekbone height |
| Canthi (eye corners) | Inner corners hidden by lashes; outer corners on the lid instead of the corner | Canthal tilt, eye separation, eye aspect ratio, symmetry |
| Alar base and nose tip (side) | Confused by nostril shadow or a slight head turn | Nasal projection, nasolabial angle, nasofacial angle |
| Tragus / porion (side) | Hidden by hair; needed to level the head | Frankfort plane, every side angle |
Errors also compound. A gonion that is 4% low does not just change one ratio; it shifts the lower third, the jaw angle, the bigonial width and the neck ratio at once, and a point that is off on one side but not the other shows up as asymmetry you do not have.
Fixing landmarks on an existing analysis
- 1
Open the analysis and go to the Options tab
The Edit Landmarks card shows your front photo (and side photo, if you uploaded one) with every point drawn on it. - 2
Click the point that looks off
A zoomed editor opens centred on that landmark. Drag it, or nudge it with the arrow keys for fine control. Keys 1, 2 and 3 switch zoom level, R resets the point, Enter confirms. - 3
Work through the risky points first
Hairline, jaw angles, chin, cheekbones and eye corners on the front; ear, nose tip and chin on the side. Do not judge from the overview; zoom in on each one. - 4
Save your analysis
A save bar appears once anything has changed, with a count of modified landmarks. Saving recalculates every ratio and the pillar scores that depend on them. Compare before and after; if a single ratio jumped a long way, look again at the landmark that drives it.
Manual review on a new scan
Your first scan runs automatically so you can see results quickly. From your second scan onwards you are asked to choose a mapping mode:
- Automatic: the AI maps the landmarks and the analysis is created straight away. You can still edit afterwards as described above.
- Manual review: the AI places a first pass, then you are walked through every landmark one at a time. Each step shows the point's name, how to find it, and a reference image. Adjust or confirm, move to the next, then check the whole set on a review screen before the analysis is created.
Pillar assessments can be re-rated too
Angularity, Dimorphism and Features assessments are scored from your photo by a vision model rather than from landmarks. If one looks wrong, open the assessment and choose Retake. You are shown a morph that runs from the lowest to the highest expression of that trait; drag it until it matches your face and save. The pillar score updates from your rating. Every assessment also has a guide in this Help Center explaining exactly what is being judged, so you can rate against the same criteria the model uses: Angularity, Dimorphism, Features.
Is it the photo, or the landmarks?
- Everything looks a bit off in the same direction (long face, big nose, close-set eyes): photo distortion. Reshoot at six feet with the back camera. See Taking the right photos.
- One or two ratios are wildly out while the rest look reasonable: a misplaced landmark. Find the point that ratio depends on and fix it.
- Asymmetry flaws you do not recognise: either the head was turned or tilted in the photo, or a point on one side is misplaced. Check both.
- Side score much lower than front: usually a head that was not level or not at a true 90°. The Frankfort correction can only do so much.
What accuracy means here
The measurements
Given correct landmarks on an undistorted photo, the ratios and angles are exact geometry and are as reliable as a cephalometric tracing. That part you can trust completely.
The ideal ranges and weights
Where each ratio should sit, and how much it counts, comes from published aesthetic and cephalometric research, with ideal ranges adjusted for gender and ethnicity. This is well founded but it is still a model of what people find attractive, not a physical constant. That is why we describe the Harmony score as a guide to your proportions rather than a rating of you.
Confirmed versus possible
In your strengths and flaws, items are listed as confirmed when the measurement is clearly outside (or inside) the ideal range, and as possible when it is borderline, where a small landmark shift could change the verdict. Treat “possible” items as prompts to check the relevant landmarks before acting on them.
Still looks wrong?
Watch the video walkthrough to see correct placement on a real face, or post a screenshot of your landmarks in the Discord and someone will point out what is off. If you believe there is a bug, tell us which ratio and which analysis, and we will look at it.
Contact supportStill need help with accuracy?
Send us a message and a person will get back to you.