Which celebrity do I look like?

01Start here: Google's Art Selfie

Go to artsandculture.google.com/camera/selfie, take a selfie, and it will match you to museum portraits from the last several centuries. It's free, requires no sign-up, and it's built on the same face-embedding technology that powers every commercial face-recognition system.

Verified live on 2026-04-24: the tool is at /camera/selfie (the art-selfie-2 URL that circulated on some older blog posts returns a 404 now). Google also runs a separate /camera/gen-selfie tool that's a different, newer generative-image experiment, not the same thing.

Why Art Selfie is a better tool than most lookalike apps even for "celebrity" searches:

• Its reference set pulls from the Google Arts Project portrait database, tens of thousands of works across museums worldwide, broader and more visually varied than any celebrity-actor database.
• It matches against portraits across ethnicities, eras, and continents. Most free celebrity tools draw 90% from Hollywood A-listers.
• The match quality is consistently high because Google's face-embedding model is one of the better ones in production, and they refresh it. Most lookalike apps run a quietly outdated open-source model.
• The result is often more interesting than "you look like Brad Pitt." You might be told you resemble a specific 18th-century Dutch portrait subject, and you may well actually resemble them.

If your goal is a sharable, unexpected, culturally interesting answer, this is the tool. Almost everyone's reaction after trying it is "huh, I see it."

02If you want a Hollywood-celebrity match specifically

StarByFace.com is the most commonly used free celebrity-lookalike tool, and it gives you a ranked list of matches with confidence scores. It's fine for entertainment. Limitations:

• Its reference set is dominated by American and British film and TV actors, with some K-pop coverage
• The matching algorithm is decent but not state-of-the-art
• Results vary wildly with photo quality: a clean frontal photo returns consistent matches across multiple attempts; a selfie in sunglasses returns a different match every time

Y-Star (iOS/Android) has better East Asian celebrity coverage, particularly K-pop idols and Chinese film actors. If that's the match you're hoping for, use this instead of StarByFace.

FaceApp has a paid "celebrity" filter, but it's more about face-morphing you toward a celebrity than matching you to one. Fun but cosmetic, not analytical.

Gradient is paid and more curated. The match set is smaller but the results feel more considered. Better for an Instagram-ready reveal than for honest similarity data.

Avoid anything that asks for email, makes you "share to unlock" the result, or shows an ad farm after the upload. A real face-matching tool doesn't need your contacts.

03The science of why this works at all

Every tool mentioned above is doing the same thing under the hood. Each face (yours, and every reference face) becomes a face embedding: a list of 128 or 512 numbers that encodes the geometry and texture of the face. The tool measures the distance between your embedding and every embedding in its database; the smallest distances are returned as matches.

The embedding technique was introduced in Google's 2015 paper "FaceNet: A Unified Embedding for Face Recognition and Clustering" (Schroff, Kalenichenko & Philbin, arXiv:1503.03832). The ArcFace improvement came in 2019 (Deng et al., arXiv:1801.07698). Every modern face-recognition system inherits from one of these.

The consequences for your result:

• Your photo is doing more work than the tool. A clean, well-lit frontal photo pulls a consistent, meaningful embedding. A selfie at a weird angle in low light produces a noisy embedding and a noisy match.
• Hair and makeup register as part of the face. Change your hair, you move in embedding space. That's why "I look like five different people across five photos" happens.
• The tool's reference set defines its world. A tool with only American actresses cannot tell you that your closest match is Aishwarya Rai; it'll tell you Anne Hathaway or Penélope Cruz and call that good.

04The genuinely interesting research

In 2022, Dr. Manel Esteller's team at the Josep Carreras Leukaemia Research Institute in Barcelona used photographs from a 20-year art project by Canadian photographer François Brunelle ("I'm Not A Look-Alike!", ongoing since 1999) as the raw material for a DNA study. Brunelle had photographed ~250 pairs of unrelated look-alikes across multiple countries; Esteller's team sequenced 32 of those pairs.

Result: 9 of the 16 pairs analysed shared significantly more SNPs with each other than with unrelated controls: meaningful genetic similarity despite no known family relationship.

The paper is "Look-alike humans identified by facial recognition algorithms show genetic similarities" in Cell Reports, 40(11), 2022. DOI: 10.1016/j.celrep.2022.111257. BBC News and The Guardian both covered it in August 2022.

What this tells you about your celebrity result:

If a tool matches you to a specific celebrity with high confidence, there's a plausible chance you actually share some genetic ancestry with them, particularly ancestry at the genes that determine facial structure. It's not a claim you can use in any practical way ("we're cousins!" will get you laughed out of the room), but the match isn't pure randomness either. Brunelle's photographs, picked by a human eye over two decades, turned out to be genetically meaningful when measured. A modern face-embedding tool picks matches the same way a careful photographer would, and then some.

05The photo you should upload

Most people take a selfie with their phone, hand raised awkwardly, backlit by a window, and then are disappointed by the result. A thirty-second setup change gets dramatically better matches:

1. Face a window. Daylight, not direct sun. Your face should be lit softly from the front.
2. Hold the phone at eye level. Not above; the up-angle distorts your face geometry and mangles the embedding. If you can, lean the phone against something and use the timer.
3. Neutral expression. No smile, no pursed lips, no raised eyebrows. The reference photos in celebrity databases are mostly neutral or softly smiling publicity shots; match that.
4. Hair out of your face. Celebrities in reference photos almost always have clean, styled hair. Your embedding will be closer to theirs if yours is too.
5. No sunglasses, no hats, no heavy filters. Filters are the #1 source of unexplained match drift.

Do these five things and any free tool will give you a more consistent, more meaningful answer.

06Where MyPhotoAI comes in (and where it doesn't)

If you're here for the match, go try Art Selfie. We don't compete with that; we don't have a celebrity-matching tool.

What MyPhotoAI does is adjacent: we take 5–15 of your phone photos and produce 50+ new, clean portraits of you in different professional styles. Not celebrity mimicry; just polished photos of your face, in good lighting, against controlled backdrops. Starts at $15 for 5 HD images; you preview watermarked results before paying.

Why that's relevant here: if the celebrity-match result was fun but also made you think "my actual photos of myself are not great," we fix that half of the problem. Separately and for real money, not as part of the match game.

07FAQ

Which tool gives the most honest answer?

Google Art Selfie, for most users. Its reference set is broader than any lookalike app's, and its face-embedding model is better-maintained than most free celebrity tools. For a specifically-Hollywood match, StarByFace is the default; it's fine as long as you don't treat a 70% confidence score as decisive.

Why does the tool say I look like someone I don't think I look like?

Usually one of three things: (a) your photo is at an unusual angle or lighting, (b) the tool's reference set doesn't include your actual closest match, or (c) you're better-looking or worse-looking than you think you are and the embedding is catching what the mirror doesn't.

Is it weird that the tool picks a different match across different photos?

No, that's expected. Face embeddings aren't fingerprints; they move slightly with hair, makeup, expression, and lighting. Try five photos and see which celebrity name comes up twice. That's your most stable match.

What if the match is genuinely unflattering?

The tool is not calling you ugly. It's reporting that your face geometry is closest to that celebrity's face geometry in its database, and that's a neutral statement about face shape, not about attractiveness. A highly photographed "attractive" celebrity is over-represented in reference sets, so they come up often even when the match isn't that close. A less-photographed but more precise match may be getting filtered out by the tool.

Can two siblings run this and get the same celebrity match?

Sometimes, not always. Siblings share a lot of facial geometry but not all of it; each one diverges from the shared baseline in different directions. It's common for siblings to get overlapping match lists but different #1 results. A fun experiment if you want to confirm the tools are actually measuring something.

Is my photo being stored / used to train AI?

Depends on the tool. Google Art Selfie says it deletes the uploaded image after the match and doesn't use it for training (check the current privacy notice for the authoritative version). StarByFace and most free apps don't publish clear retention policies. If privacy matters to you, use the Google tool or use an offline library like face_recognition on your own computer.