Glowster

Why your face score can change between uploads. And what we do about it.

You're not imagining it

A common worry about AI face scoring is that the same picture, submitted more than once, can come back with a different number each time. If a score changes when nothing about your face changed, the app is not measuring anything. It is generating a plausible-sounding number.

This matters beyond pride. If your baseline is random, your progress is unmeasurable. A +4 after a month of disciplined skincare might just be the dice rolling differently.

Why AI scores can drift

Three engineering shortcuts can cause it. First, an app can send your photo to a language model with a vague prompt ("rate this face 1-10") and no fixed scoring criteria, so the model improvises a different judgment each time. Second, sampling randomness: the same settings that make a chatbot phrase things differently each conversation make a score land differently each upload. Third, inconsistent re-processing of your image: different crops, different compression, different inputs.

What it takes to make a score repeatable

We built Glowster around one promise: upload the exact same photo file again and you get the same scorecard. Five mechanisms make that true. (1) Deterministic preprocessing: every upload is resized and normalized the same way, so the model sees identical input for identical photos. (2) An anchored, versioned rubric: the model scores against fixed written band definitions for each sub-factor - not vibes - and every score is stamped with the rubric version that produced it. (3) We don't suppress the model's own sampling randomness - we don't control that - but we don't need to: score stability comes from the anchored rubric and the next mechanism instead. (4) Multiple independent scoring passes with the median taken - and more passes automatically when they disagree - so residual model noise gets cancelled out rather than eliminated at the source. (5) The app does the math, not the model: the model only scores fourteen named sub-factors, and your feature scores and overall are computed from them by a published weighted formula. The headline number is arithmetic over those sub-scores, never a figure the model improvised - which is exactly the drift failure mode above.

When we change anything about the rubric or the model, the version number changes, and we re-validate consistency against a fixed photo set before shipping. Your old scores stay comparable to your new ones, or we tell you they aren't.

Why this is the whole product

A consistent score is what makes the rest meaningful. Your 90-day plan targets your lowest-scoring features - that only works if those scores are signal. Each check-in shows a delta - that only matters if the delta is real. A score that drifts can only sell you the first scan. We can show you the second one moving.

Test us.

Run your scan, save your number, then scan the same photo again at your next check-in. If we're wrong about consistency, you'll know - that's the point.

Get my repeatable score →