How to Find Your Celebrity Look Alike

You know your own face better than anyone alive. What you cannot do is see it fresh, the way a stranger does on a first meeting. That is the small gap a celebrity look alike search fills: it takes the geometry of your face, compares it with a library of famous faces, and hands back the twelve closest.

This guide covers the whole process, from choosing a photo to reading the results without overinterpreting them. It also explains why the answer changes depending on which picture you use, which is the part most people find confusing.

How face matching actually works

Strip away the jargon and the process is a short assembly line that runs entirely inside your browser tab.

1. Face detection. A small neural network scans your image and draws a box around any faces it finds, along with the position of the eyes, nose, and mouth corners.
2. Alignment. The face is rotated, scaled, and cropped into a tidy 112 by 112 pixel square, with the landmarks lined up in consistent places.
3. Embedding. A second network (an SFace model) converts that square into 128 numbers. Those numbers describe the shape and texture of the face: how wide the jaw sits relative to the cheekbones, how far apart the eyes are, how the brow line falls.
4. Comparison. The tool measures the angle between your 128 numbers and the 128 numbers of every celebrity in the library. That measurement is called cosine similarity. Small angle, high similarity.
5. Ranking. The closest matches are sorted and shown as a match score out of 100, rescaled so the numbers are easier to read than raw similarity values.

Two things are worth noticing here. First, nothing is uploaded. Detection, alignment, embedding, and comparison all happen on your own device using WebAssembly, and the photo is dropped the moment you close or reload the page. Second, the library holds a small number of images per celebrity, often just one, so you are being compared against a specific photograph rather than against a person’s entire face history.

For a longer explanation of the privacy side, see our privacy page.

Before you upload: what makes a good photo

The matcher can only work with what it can see. These conditions give it the cleanest possible input:

• Face the camera. Both eyes visible, chin roughly level, head not tilted more than a few degrees.
• Even light. Open shade, a window, or soft indoor light. A single hard lamp on one side flattens one half of your face into shadow.
• One person in frame. Group shots confuse the crop, especially when two faces overlap.
• No sunglasses. The model needs to see your eyes. Regular glasses are fine if there is no hard glare across the lenses.
• No heavy filters. Smoothing, face slimming, and eye enlarging all rewrite the texture the embedding is trying to read.
• Face at least 15% of the frame. A full-body holiday snap gives the detector a face the size of a thumbnail. Crop it first.
• A relaxed expression. A neutral or slightly smiling face reads more reliably than a mid-laugh action shot.
• A recent-ish photo. Not because older pictures are worse, but because the celebrity library is built from mid-career portraits. A photo from a similar life stage compares more fairly.

One honest caveat: pictures that look good to you are not always good for the matcher. A dramatic portrait with strong side light might be your best photo and a poor input for a face comparison.

Photo type, and how well it tends to work

Photo type	How well it works	Why
Front-facing selfie in daylight	Very well	Even light, full frontal pose, face large in frame
Passport or ID style photo	Very well	Neutral expression, flat lighting, clean background
Group photo with one clear face	Well, if cropped	Cropping removes the second face and the clutter
Mirror selfie	Mixed	Camera often covers part of the face, lighting is uneven
Sunglasses or deep shadow across the eyes	Poor	The landmarks the model relies on are hidden
Beauty filter or heavy smoothing	Poor	Skin texture has been rewritten before the model sees it
Side profile or extreme angle	Weak	The model was trained mostly on frontal and near-frontal faces
Childhood or teenage photo	Not comparable	The library uses mid-career portraits of adults

Step by step

1. Pick a photo that meets most of the conditions above. Your camera roll almost certainly has one.
2. Open the tool on the home page and either upload the file, paste an image URL, or start with a sample photo.
3. Wait a few seconds while the models run. On a laptop this is close to instant. On an older phone it can take a moment, because the download and the math both happen locally.
4. Read the top twelve. Look at the whole list rather than the first row. The order near the top is often tight, and ranks three through eight can be separated by a hair.
5. Run a second photo and compare the two lists. This is the most useful thing you can do with the tool, and most people skip it.

The upload cap is 20MB, which is generous for a single photo. If your file is bigger, resize it rather than fighting the limit.

Why your top match changes with every photo

This surprises people, so it is worth spelling out. Four things move the result around.

Your face is not a fixed shape. Expression, head angle, and the focal length of the lens all change the proportions the model measures. A phone held close exaggerates the nose and narrows the jaw. The same face shot from two metres away looks different to a camera and to the embedding.

Lighting rewrites texture. The embedding encodes more than geometry. Shadows change how the model reads the contour of your cheeks and the depth of your eye sockets.

The library is a set of photographs, not platonic faces. If a celebrity’s reference picture is a 1990s press shot at an awards ceremony, you are being compared with that specific lighting setup and that specific haircut.

The top scores are close together. When your nearest neighbours sit within a couple of points of each other, tiny changes flip the order. A different photo does not mean the tool is broken. It means the signal near the top is genuinely tight.

How to compare two photos of yourself

If you want a result that feels more stable, treat it like a small experiment.

• Run both photos and write down the top five names and scores from each.
• Look for overlap. A celebrity who appears in both lists is a more durable result than a one-off number one.
• Keep conditions as similar as you can: same time of day, same light, same distance from the camera.
• Do not compare raw scores across photos as though they were measurements on a ruler. A 74 in one photo and a 79 in another does not mean the second photo made you more like someone.
• Try a third photo if the first two disagree completely. The truth is usually somewhere in the middle of the three.

What to do with the result

The most fun use is social. Screenshot the list, post the top three, and ask your friends whether they see it. If you want to bring someone in without sending your photo anywhere, generate a challenge link: it encodes the matched celebrity’s name and the score in the URL fragment and nothing else. No image, no name, no personal data. A friend can open it and run their own match to see whether they beat your score.

Other things people do with a match: build a film list from the actor they were matched with, test a family resemblance (run the same photo of each sibling), or check whether the result holds across a decade of photos.

What this method is not

Face embeddings are not a scientific test of resemblance. They measure similarity in a 128-number space defined by a model trained for face verification, and that is a narrower thing than “looks like”. Two faces can score highly because of broad structural similarities that a human would never notice, and two faces that friends think look identical can score modestly if the photos differ in angle or light.

The match score out of 100 is a similarity score, not a probability, not a confidence level, and not a claim about how alike you truly are. It is a number that goes up when two faces sit close together in the model’s representation and down when they do not.

Above all, this is entertainment. Our tool is not face recognition for identification, not biometric verification, and not something to use to work out who an unknown person is. Use it on yourself, use it on your friends with their permission, and enjoy the arguing.

Find your celebrity look alike and see what the top twelve say. Then try a second photo and watch the list shuffle.