How AI facial analysis works, and where tools actually differ
Every face analysis tool does the same three things. They differ at the third step, which is the one almost none of them explains.
Last updated 2026-09-21
Step one: find landmarks on the face
A detection model places reference points on the image. The widely used ones place around 478, covering the eye openings, the lips, the line of the nose, the jaw and the surface of the cheeks.
This is where marketing overstates. 478 landmarks are not 478 measurements. They are coordinates, not conclusions. How much can be measured depends on which points get paired with which, and on how many of those pairs mean something anatomically — far fewer than the point count. For DooDee that is 51 values.
The quality of this step determines everything after it. If points land badly because of lighting, hair across the face, or camera angle, every value computed from them moves too.
Step two: compute ratios and angles
From the coordinates the system computes distances and angles, then divides by a reference distance on the same face to produce a ratio. That division is necessary because a photograph has no scale: a closer camera enlarges every distance in the frame at once, while the ratios between them hold.
This step is arithmetic. It is checkable, and two tools using the same landmarks get the same answer.
Step three: compare against something — and here is where tools diverge
A raw number means nothing until it is compared with something, and there are three available somethings:
| Compared against | What it can say | Limit |
|---|---|---|
| Published population norms | Where you sit relative to that group | Few values have them, and the group has to match you |
| Yourself (left-right, or a previous scan) | Difference within one face, and change over time | Says nothing about anyone else |
| An ideal the maker chose | A number that looks like a score | An opinion rather than a measurement, and almost never sourced |
DooDee uses the first two. Of 51 measured values, 12 have published reference figures, from a cohort of 240 Thai adults aged 18–35. The rest compare you with yourself.
Why the same face gives different results
- Direction of light — anything read from shadow, like under-eye intensity or jawline definition, moves with the lamp more than with the face.
- Head rotation — a few degrees is enough to move a symmetry figure. If symmetry changed between two scans, the camera angle is the likeliest explanation.
- Camera distance — a close lens enlarges what protrudes, like the nose, relative to everything else, and a ratio does not undo that.
- Expression — a smile genuinely changes mouth width and cheek position. That is not error.
The fix is not a more accurate tool. It is shooting under the same conditions each time you intend to compare.
A short checklist before trusting any tool
- Does it say what it compares you against, and who that group is?
- Does it say what it cannot measure? A tool claiming to measure everything from one photograph is claiming more than a photograph carries.
- Does it report millimetres from a single image? Then it is guessing your camera distance.
- Does it separate "compared with other people" from "compared with yourself"?
- Where does your uploaded photograph go, and can you delete it?
Read next
- What one photograph can measure — 51 values, and 13 it cannot reach
- Do I look good? — no tool can answer that, and here is what one can measure instead
- The golden ratio and the face — where it came from, and why it is not a standard
- Before a cosmetic consultation — what to ask, and when to walk out
- Why foreign facial norms do not transfer to Thai faces
- What is DooDee — AI facial analysis that also tells you what it cannot measure