Why Your Skin Scan Results Change Between Scans

Any photo-based measurement has a noise floor, and a tool that will not tell you its own is asking you to read meaning into numbers that may not carry any. So we measured ours.

The method: take a real face, apply controlled photometric perturbations to the same source image — exposure up and down, gamma up and down, colour temperature warmer and cooler, and additive sensor noise at three levels — and run every version through the full scoring pipeline. The skin has not changed at all. Any movement in the scores is the scanner reacting to the photo.

Eleven faces, ten versions each. The figures below are the median across those faces, on the 0 to 10 severity scale. They are regenerated from the committed results file rather than written by hand, and they measure invariance — whether the same face scores the same under different conditions — not accuracy, which is a separate question a harness like this cannot answer.

Under exposure, gamma and colour-temperature shifts — the everyday differences between a window in the morning and a bathroom at night — the scores barely move.

Oiliness shifts by about 0.05 points. Sensitivity by 0.11. Dryness by 0.16. The overall composite by 0.21. On a scale that runs to 10, and against a gate that only acts on a concern at 3.0, those are comfortably inside the noise.

Pigmentation (0.52) and acne (0.76) are looser. That is consistent with what we already say publicly about pigmentation being the weakest of the five axes, and it is why pigmentation is attenuated when the variation it finds shows no colour support.

Sensor noise is a different story from lighting, and the honest version is worth stating plainly. Most axes handle it well — oiliness moves by 0.015 and sensitivity by 0.05, better than under lighting changes, because grain does not shift the average brightness of a region.

Acne does not handle it. In the worst case across our set, the acne reading swung the entire width of the scale between the cleanest and the grainiest version of one face. The mechanism is specific: lesion candidates are found as local texture features, and once the whole image is grainy the extractor stops being able to tell a lesion apart from noise, so it finds nothing and the score falls toward zero.

This is why a scan taken in a dark room is the least trustworthy kind, and why the practical advice on our capture guide leads with light rather than with framing. It is also the clearest single argument for the confidence number: the pipeline can tell when an image is degraded, and it would rather report low confidence and fall back to a maintenance routine than hand you a plan built on a reading like that.

Three mechanisms, all of which exist because of the numbers above.

Displayed scores are damped. From your second scan onward, a single outlier reading is pulled toward your recent history rather than shown at full strength. A genuine change persists into the next scan and comes through in full; a one-off does not. The raw value is always what gets stored, so nothing is lost — only the display is smoothed.

Concerns have hysteresis. An ingredient is recommended for a concern once it crosses 3.0, and a concern that was active on your last scan is held active if it lands within 0.5 of that line. Without this, an axis reading 2.9 then 3.1 then 2.9 would add and remove an active ingredient from your routine on alternate days. A genuinely new concern still has to clear the full threshold.

Routines are carried forward. If you rescan within 24 hours and nothing material has changed — no concern crossing the line, no shared concern moving more than 1.5 points — you keep the previous routine and its start date rather than being handed a fresh plan. A 12-week schedule that resets every time you take a photo is not a schedule.

Same place, same time of day, same light. Facing a window is the easiest repeatable setup most people have.

Avoid scanning straight after cleansing, exercise or a hot shower — all three temporarily change both oil and redness, and none of them tell you anything about your baseline.

Give a change time to be real. Barrier repair shows up over weeks and pigmentation over months; two scans a day apart are measuring the same skin twice, not measuring progress.

Treat a single scan as one reading, not a verdict. The trend across several is the part that carries information.