People sometimes assume a face test is thrown together in an afternoon: grab some pictures, done. The ones that feel right take more than that, and most of the work is editorial, not technical. Here's the process behind every test on this site, including the rules we learned the hard way.
Step 1: choosing the work
A test only works if players recognize the cast. So we start with works that have a broad, recognizable set of faces: a long-running series, a hit film, a game with a big roster, a K-pop group, or a family of animals. We track what people are searching for and what readers suggest, and we skip works we don't know well ourselves; a test built without knowing the characters produces wrong names, wrong groupings, and results that fans spot as off immediately.
Step 2: the cast
More characters make a better test, because a bigger cast gives the AI more distinct directions to place you in. We aim for twelve to sixteen per test and include supporting characters and mascots when they're recognizable, since a non-human face (a dog, a robot, a mascot) can be a genuinely good match for some players and is always a fun one. A test needs at least five or six solid characters to be worth publishing; below that it's a coin flip with pictures.
Step 3: stills, not headshots
This is the rule we learned by getting it wrong. Early on we tried using actors' promotional headshots for live-action tests. Matching was poor: red-carpet lighting, different hair, no in-character makeup. The face people recognize as the character is the one in the show, so we collect frames from the work itself, several per character, at different angles and expressions. Group shots are dropped (the model would average two faces), posters with text overlays are dropped, and anything that isn't clearly the character's face is dropped. A character with only one clean still can stay; a character with none is cut.
Step 4: one medium per test
Live-action and animation never share a test, and neither do photographs of real animals and drawings of them. An image model separates media before it separates faces, so a mixed cast would always favor whichever side your photo is on. The full reasoning is in why anime faces are hard for AI.
Step 5: review
Every candidate test gets a contact sheet: every still of every character on one page. A person goes through it and drops the failures: a still that's actually a different character, a frame where the face is turned away, an image that's from a different adaptation. Only then are the vectors computed and the test assembled.
Step 6: names and descriptions in ten languages
Character names get their official localized form in each language, checked by a person, because machine translation of a name produces nonsense (a character called "Snare" or "Root" where the real localized name is a transliteration). Result descriptions are written as projections, "you're the type who…", rather than character biographies, so the result feels like it's about you. And every test gets its own introduction, written for that work, rather than a template with the title swapped in.
Step 7: the gate, and going live
Before publishing, an automated check verifies the structure: every character has usable images, every language field is filled, no untranslated text is left behind. If it passes, the test goes live, and from then on we watch how people play it: which characters win most, whether the results people share look like matches fans agree with, and whether anyone reports a wrong name. Tests get corrected after launch when they need it.
What we don't do
We don't store your photos (the analysis never leaves your browser; see the privacy checklist). We don't mix media. We don't publish a test for a work we don't know. And we don't claim the result means anything beyond resemblance: it's a measurement, and the fun is in what you do with it.