The red flags
- Sub-centimetre accuracy claims. No published method achieves that, including professional 3D scanning.
- A single accuracy figure with no breakdown by measurement.
- No mention of test conditions, sample size or how validation was done.
- Photos used to train models, disclosed nowhere.
- Health claims derived from your measurements without saying so.
What to ask for
- Per-measurement error, not an average. Waist and arm differ by more than double.
- Sample description. Young lean Western samples are the norm; ask whether bodies like yours were included.
- Whether ATA creation is included. Paying to create a recipient token account changes the cost.
- What happens to your photos. Ideally processed and discarded.
Benchmark numbers to judge against
Published smartphone app validation reports root-mean-square errors of 2.5 cm at the arm, 4.6 cm at the hip, 5.1 cm at the thigh and 6.1 cm at the waist.
A professional 20-camera scanner in the same study gave 2.9 cm at the arm and 9.2 cm at the waist. If an app claims better than that with a phone, be sceptical.
Ask what the error is on the waist specifically. That is where the honest answers are, and where the dishonest ones give a single average.
Why the average is a trap
An average across arm, thigh, hip and waist will look respectable because most measurements are decent. Waist, the one that matters most for fit and is hardest, gets buried.
Always ask for the worst number, not the mean.
What the research has not covered
There is no published benchmark for photo-based AI estimation of the measurements that determine garment fit: chest, shoulder, neck, inseam, thigh, calf, arm and back length. Existing validation has focused on waist, hip and limb circumferences for nutritional assessment.
So if a provider quotes a precise figure for inseam accuracy, there is likely no study behind it.
Privacy as a trust signal
A tool that tells you plainly what it stores, what it discards and who can see your data is more likely to be honest about everything else. Vague privacy language is a reasonable warning.
Note the nuance: body measurements are not special-category data under GDPR unless used to uniquely identify someone, but inferred health data would be. See the privacy detail.
A practical checklist
- Per-measurement error published.
- Validation against a tape measure, not against another app.
- Sample composition stated.
- Photos discarded or explicitly retained.
- No health inference without disclosure.
- Prices stated before you commit.
Deeper detail in how accurate these tools are.