How should a POM match score?

Drag a slider to change how much something counts, or change the product itself. The POMs are a fixed sample, each chosen to differ from the others in one way.

Matching POMs for Women’s Coat

AI fields are read from the sketch and name — pick a value to simulate another extraction. Brand is typed by the designer in the tech pack header.

Best matches

matches partly doesn’t match no info

Photo credits

Reference photos are from Pexels (free to use). They show the garment each POM stands for; the brands are not the POMs’ brands.

How the score works

Each POM collects points for everything it shares with the product. A closer match earns more of that point.
Green means it matches, amber means it only partly matches, red means it doesn’t, and grey means there was nothing to compare. A “?” marks the same value when the AI was unsure.
A part’s points are its importance ÷ the total importance × 100, so importance is never itself a score: all six at 10 is 16.7 points each, not 60 in total.
The percentage is the points a POM collected out of 100. Each pill shows points earned / points available; “Score breakdown” shows how they were worked out.
Confidence is the lower of the product’s and the POM’s. A label that matches earns 50% + 50% × confidence, a similar garment earns 50% + 35% × confidence, and a different label earns 50% − 50% × confidence.
Confidence only applies to labels. The sketch earns (similarity − 0.5) ÷ 0.5, and a shared brand earns the same share as the labels overall.
A missing brand is left out and its points go to the other parts, so it never counts against the POM. A product can list several brands; a POM sharing any one of them gets the brand points.
A POM with no sketch earns half the “looks similar” points, so it can’t outrank a POM with a good sketch. An unknown label also earns half.
A POM for another gender is hidden when the product and the POM are both at least 70% sure of their gender, whatever the weights.
A POM in a different size group loses 5 points, so it still shows but ranks lower. Size comes from the name and tags; if neither states one, it counts as regular.
Sketch similarities here are samples, set from one real test: a trench product scored 0.76 against a trench POM and 0.62 against a bomber POM.
These are the weights the API uses today; drag them to test other ideas.