Millions of colours, a handful of shades

Why your eye needs help here

Two patches of a cheek can look identical side by side and be half a value apart when you mix them. Worse, judgement drifts: the shade you called “mid-warm” an hour ago is not the shade you will call “mid-warm” now, because your eye adapts to whatever it just looked at. A photograph does not drift. Measuring it does not either.

Grouped where perception lives

Every pixel converts into CIE L*a*b*, a colour space built so that equal numerical distance means roughly equal visible difference. Grouping in Lab produces shades that look distinct to a painter. Grouping in RGB — which is what most posterise filters do — produces shades that are mathematically tidy and visually arbitrary.

Boundaries that mean something

Raw grouping leaves a speckled mess wherever a smooth gradient crosses between two shades, and camera noise makes it worse by looking like real detail. Two passes fix that: an edge-preserving filter removes noise before grouping, and a refinement pass then asks of every pixel whether it should join its neighbours — a question weighted down wherever a genuine boundary runs, to help preserve features such as a jawline or eyelid.

Then it tells you how well it did

The flattened map is compared against your untouched original in CIEDE2000, the current standard for perceived colour difference, and reported plainly: the average difference, the share of the picture within a just-noticeable difference, and the worst five per cent. If the answer is not good enough, raise the shade count and look again.