The internet has been asking the same question since the first photo booth: who, exactly, do I look like? Most celebrity look alike apps answer with a coin flip dressed up in a progress bar. We wanted something more honest. Eleven’s celebrity look alike finder runs your photo through ArcFace, the same family of models that decides whether your passport matches your face at the airport, and ranks you against fifty hand-picked celebrities by mathematical similarity rather than vibes. What comes back is a card with a top match, four runners-up, and a breakdown of which features the model actually noticed. No account required. No tracking pixels. No upsell on the result page. Just a celebrity face match, reasoned out from the geometry, served in roughly the time it takes to pour a glass of water.
Two faces become two vectors in 512 dimensions. The closer the angle between them, the closer the resemblance. There is no closer reading than that.
How it works.
A photo goes in. The model does the math. A verdict comes out. Three moves, in order.
Upload. Drop a single front-facing image into the panel. JPG, PNG, or HEIC, up to five megabytes. The model wants clear light and visible eyes; anything portrait-grade will do. EXIF metadata, the camera details and geolocation tag your phone quietly attached, is stripped the moment the file lands on our server. No account. No tracking. You confirm two things: that you’re eighteen, and that the face is yours.
Analyze. Once the upload completes, your face is cropped and passed through Buffalo_L, an open-source ArcFace implementation built by InsightFace. It converts your features into a 512-dimensional vector: a list of 512 numbers that, taken together, describe your face the way coordinates describe a point on a map. Those numbers are then compared to the celebrity vectors stored in pgvector, our Postgres-based similarity index. The comparison itself is one equation, repeated fifty times. Compute time runs under 200 milliseconds. The full upload-to-result loop takes around two seconds on a normal connection.
Result. What you get back is a verdict, not a guess. Top match in large type. Four runners-up beneath it. A similarity score for each. Then the breakdown grid: face shape, eyebrow score, smile score, all derived from a 106-point landmark map of your features. Share the card. Opt the photo into the public gallery for that celebrity. Or close the tab and the photo is gone. The default is gone.
The science.
There are two ways a computer can look at a face. It can classify, which means picking from a fixed menu: male or female, smiling or not, this celebrity or that one. Or it can recognize, which means converting the face into a kind of fingerprint and comparing fingerprints against each other. Recognition is the better tool. It is what opens your phone. It is what the e-gate at Heathrow uses to wave you through without a stamp. And it is what we use, turned inward toward fame instead of identity.
The fingerprint, in our case, is a 512-dimensional embedding. Think of it as the face’s coordinates in a very high-dimensional space. Each dimension captures something, perhaps the angle of a brow, the ratio of a jawline, the curve of a lip, without a human ever having to label what it captures. The model learned those features on its own, from millions of training images, by being asked over and over: same person, or different? After enough rounds, it figured out which arrangements of pixels tend to belong to the same face.
To compare two faces, we use cosine similarity. It is, simply, the angle between the two vectors. Identical twins read close to zero. Strangers read close to ninety degrees. Your celebrity look alike, the one the model says you most resemble, is whichever of the fifty in our library lands at the smallest angle to you.
Why fifty celebrities and not five thousand? Because curation matters more than scale at this stage. We picked faces that read distinctly across the model’s geometry, span the major demographic combinations, and hold up under the breakdown grid. A bigger library would dilute the experience without making it more honest. A celebrity face match is only as good as the faces it can match against.