Why most AI body scans guess, and what "Not visible" means

The test we ran on ourselves (August 26, 2026)

Sixteen scans through the live FlexScan grader: one person, one front photo scanned six times at the default settings, plus darker, brighter and heavily compressed versions of it, an arms-up front, and a side. Every response was kept and every grade was compared. The person was a licensed stock model; no user photos were involved.

• The same file scanned six times produced six different grade sets. Not one pair matched.
• Chest swung from A+ to B, and core from A to B-, on byte-identical input, minutes apart. Four grade steps.
• The grader invented grades for glutes in two of twelve front scans, on a photo where glutes are not visible. The "cannot see it" judgment itself was flipping.
• The advice moved with the grades: across six identical scans, the "focus on this" list named five different muscle sets.
• Lighting changes mattered less than that noise. The heavily compressed copy seemed to pull core and legs down one to two steps, but the re-run after the fix showed that was noise too. What survived: every re-saved copy (darker, brighter or compressed) came back one step lower on shoulders, every time.

What we changed

• We pinned the model's randomness (temperature zero, a fixed seed). Re-run six times on the same photo: seven of the eight muscle groups came back identical every time; one group moved by one step in one of the six runs.
• A one-step change is never reported as progress. The comparison screen requires a move bigger than the grader's own wobble before it draws an arrow, and it says so.
• Back is never graded from a front photo. It is the one judgment that was stable in all sixteen scans, and the rule is now explicit for every group: grade only what is clearly visible, otherwise write "Not visible".
• We ran this test on August 26, 2026, before and after the fix, and the grader's instructions and settings have not changed since. Re-running it is a manual step, not yet an automated release check.
• Comparisons only ever score the muscles graded in BOTH scans. If one photo showed six groups and the next showed eight, the extra two are left off the comparison, so its number cannot move just because the photo coverage did. (The score trend chart is different: each point averages whatever that scan could see.)

How the score is made

Each grade is a point on a 13-step scale, A+ at the top and F at the bottom. A scan never returns an F, though: the lowest grade it gives is D-. The groups that were graded are averaged, "Not visible" groups are left out entirely, and the average is shown out of 100: an all-A+ scan is 100, and the scale's F floor is 0, so the lowest score a real scan can get (all D-) is 8. The number is nothing more than the coach's letter grades averaged. It is not a body measurement, and it is why a change in which groups were visible can move it, and why comparisons use the shared groups only.

What "Not visible" means on your report

It means the coach was not shown enough to grade that group and refused to make one up. From a front photo that is normally back and glutes; from a side photo, often core. It is not a bad grade and it is not counted against your score. You can tap the muscle on the map and rate it yourself, or add the missing angle next time. The example scan is a real report with two groups left "Not visible" on purpose, so you can see exactly what it looks like before you sign up.

Why we do not print a body-fat percentage

A photo has no depth, no density and no view of what a shirt covers. Some apps publish a comparison of their estimate against DEXA and quote a range, which is the honest version of that number. We would still rather grade what can be seen than print a percentage that looks like a measurement. If you want body fat, use a method that measures it, and photograph yourself the same way every time so the two agree about direction.

What we still cannot promise

• A grade is an opinion about muscle development from pixels. Two coaches would disagree too.
• Pose, distance and compression move grades. Shoot the same way every time or the comparison is comparing photography.
• The model can change under us even when our code does not, so a result from August is not a guarantee for today. The test has to be re-run, not remembered.
• None of this is a medical assessment or a body measurement.