Why does the same face score differently on two scans?

Is a two-point difference a real change?

No. A score of 58 followed by a 56 is the same face read twice, and BecomeTen labels it that way in the report: within the noise band, shown neutrally, neither progress nor regression. This is the most common support question we get, and the honest answer is that the number has a resolution, and that resolution is roughly two points on the 0–100 scale for photos taken under similar conditions.

Every measurement works like this. A bathroom scale drifts a few hundred grams between one morning and the next without anything having changed; a blood-pressure cuff gives two readings a few points apart in the same sitting. Nobody concludes from the second reading that the first was wrong. A face rating from a photograph is looser than either of those instruments, and pretending otherwise would be false precision of exactly the kind the methodology rules out.

Where does the variance come from?

Five sources, in roughly descending order of how much they move a score. The first three are yours to control; the last two are properties of the rater.

The first two rows are the reason the photo setup guide exists. Hold light, angle, distance and expression constant and most of the variance goes with them.

Why does the model itself vary on the same photo?

This is the part most rating apps do not say out loud. BecomeTen scores a photo with a large vision-language model against a fixed rubric, and such models are not deterministic: run the identical image twice and the output can differ by a point or two, and a trait that sits near the boundary between two bands can land on either side. How AI face rating works explains the pipeline; the short version is that the model gives a structured opinion, not a measurement, and opinions have some scatter.

Two things limit that scatter. The rubric is fixed, with named traits, named bands and the same calibration anchors on every run, so the model is not free to invent a new scale each time. And the headline score is recomputed by the server from the area scores rather than taken from the model's own summary, which removes one layer of drift. What remains is small, and it is the reason the noise band is two points rather than zero.

What do observers actually agree on?

If a model wobbles, it is fair to ask whether attractiveness is stable enough to be scored at all. The perception research says yes, at the level of the face, and not at the level of the decimal.

That is the shape of a face rating too. Which tier you are in, and which two or three areas are your weak points, is stable across scans and across human raters. Whether you are a 61 or a 63 is not, and BecomeTen does not pretend to know. The PSL scale explained covers why forum ratings vary far more than this, with no rubric and no anchors at all.

What does a real change look like at a checkpoint?

A real change is slow, lands in one or two areas, and survives two checkpoints. Skin tone and clarity move first, typically inside the first eight to twelve weeks of a routine with evidence behind it. Jawline definition moves with body composition over a few months. Hair density, if it moves, takes longer still. Harmony barely moves at all, because bone does not.

So at a 6–8 week checkpoint on matched photos, the signature of progress is a rise of three or more points in the area you were working on, with the others holding still, repeated at the next checkpoint. The signature of noise is the opposite: every area shifting together, which is lighting; a spike that falls back, which is water or sleep; or a jaw that jumped while the skin sat still, which is usually angle. How to track progress has the full reading table.

How does PRO progress tracking compare scans?

Every photo is graded before it is scored, on lighting, angle and sharpness. On a re-scan the change from baseline is shown only when both photos pass; if either fails, the comparison is withheld and the report says which condition to fix. Where both pass but differ in light or angle, the change is shown with a low-confidence note. And a change within two points is always labelled noise.

With PRO, at €9.99 a month or €59.99 a year for 20 scans per rolling 30 days, each scan is plotted per area over time against your baseline, with the same rule applied at every point. The model sees both photos side by side and is instructed to report only differences it can see. Unchanged is a valid result. Regressed is a valid result. It may never report a change in bone between scans, and it will not soften a drop that is real. A tracker that only ever reports improvement is not tracking anything.

What should you not do with a two-point drop?

The useful response to a small difference is to do nothing, keep the setup constant, and look again at the next checkpoint. The rater is honest about its own resolution so that you can be patient with it.