What does it mean for faces to "look alike"?

AV女優AI顔診断研究所(AIAVMATCH LAB)・2026/11/01

研究員 藍子(あいこ)
藍子
Director, listen to this. I used our AI to compare two actresses people always say "look alike" — and it came back "not very similar"…
研究所の所長 顔見 識朗(かおみ しきろう)
所長
What, did the AI get it wrong?
研究員 藍子(あいこ)
藍子
Well, I can’t quite say it’s wrong either. I started wondering what "looking alike" even means in the first place.
研究所の所長 顔見 識朗(かおみ しきろう)
所長
Hmm… What it means to look alike, eh. A deep question. Let us investigate that today.

What "looking alike" means to AI

特任研究員(画像認識AI) 畳込 ニューロ(たたみこみ にゅーろ)
ニューロ
Model-wise, you see, we convert a face into a sequence of 128 numbers. The distance between the eyes, the shape of the nose, the outline — all of it becomes numbers.
研究所の所長 顔見 識朗(かおみ しきろう)
所長
As many as 128?
特任研究員(画像認識AI) 畳込 ニューロ(たたみこみ にゅーろ)
ニューロ
Yes. Then we measure how much the two sequences of numbers point in the same direction. That is the "similarity score".
研究員 藍子(あいこ)
藍子
The key point is that we crop out just the face before measuring. So hairstyle and clothes barely matter.
研究所の所長 顔見 識朗(かおみ しきろう)
所長
I see. So the AI is looking at the framework of the face.

What "looking alike" means to people

特任研究員(メイク講師) 盛田 映子(もりた ばえこ)
盛田
But you know, when people feel two faces "look alike", what matters most is the hairstyle and makeup. In other words, the vibe.
研究員 藍子(あいこ)
藍子
And that shows up in the data, too. When we measured hairstyles title by title, 2,830 people wore a different hairstyle depending on the title (out of the 2,971 people we could measure across two or more titles).
特任研究員(メイク講師) 盛田 映子(もりた ばえこ)
盛田
See? Between a title with long black hair and one with a bright short cut, the same person looks like someone else entirely.
研究所の所長 顔見 識朗(かおみ しきろう)
所長
I understand. Long ago, I took a photo of a TV personality to the salon and asked them to give me the same hairstyle. And my vibe ended up quite similar, I tell you.
研究員 藍子(あいこ)
藍子
Oh! Really?
研究所の所長 顔見 識朗(かおみ しきろう)
所長
…No, it did not, actually.
特任研究員(メイク講師) 盛田 映子(もりた ばえこ)
盛田
Of course not. A hairstyle can bring the vibe closer, but it cannot bring the facial framework closer.
研究所の所長 顔見 識朗(かおみ しきろう)
所長
I see. People look at the vibe; AI looks at the framework. That is why the two accounts diverge.

Faces that "resemble everyone" really exist

研究員 藍子(あいこ)
藍子
And there is one more big discovery. Want to know?
研究所の所長 顔見 識朗(かおみ しきろう)
所長
What is it, stop teasing me
研究員 藍子(あいこ)
藍子
When we made lists of "similar actresses" from faces in titles, one actress appeared in the top 6 of 241 people’s lists! Out of 1,000 people.
研究所の所長 顔見 識朗(かおみ しきろう)
所長
What?! She resembles one in four people?!
特任研究員(画像認識AI) 畳込 ニューロ(たたみこみ にゅーろ)
ニューロ
In the AI world, this is called a "hub". When you average many faces, the components common to everyone grow stronger, and the result ends up close to every face.
研究員 藍子(あいこ)
藍子
So we subtracted the overall average and corrected for the "resembles everyone" part as well. Then the maximum of 241 appearances dropped to 19, and the people who had never appeared in any list went from 36% to 0%.
研究所の所長 顔見 識朗(かおみ しきろう)
所長
"Resembling everyone" is a hair’s breadth from "resembling no one", isn’t it.

One photo is never enough

特任研究員(メイク講師) 盛田 映子(もりた ばえこ)
盛田
Also, profile photos are all about looking your best. Lighting, makeup, retouching — it is all in there.
研究員 藍子(あいこ)
藍子
We also ran an experiment: gather faces from title footage and identify the same person. With just one face, the match rate was 24%. But bundle 40 faces of the same person, and it rose to 71%.
特任研究員(画像認識AI) 畳込 ニューロ(たたみこみ にゅーろ)
ニューロ
The variation in angle, expression and lighting cancels itself out as the count grows.
研究所の所長 顔見 識朗(かおみ しきろう)
所長
So: never judge from a single photo.

How to read the "similarity score"

特任研究員(統計学) 分散 さゆり(ぶんさん さゆり)
分散
Hold on. That "similarity score of X%" is not an absolute number, you know.
研究員 藍子(あいこ)
藍子
Right… Scores between strangers average around 0.1, and even the top 5% are around 0.3. We convert where a pair falls within that distribution into a percentage.
特任研究員(統計学) 分散 さゆり(ぶんさん さゆり)
分散
In other words, "90% similarity" does not mean "90% the same face" — it means "quite close, as strangers go". Do not misunderstand that.

Today’s lesson

謎のAV女優 ナナシ
ナナシ
…When people say "you two look alike", I wonder if it is meant as a compliment.
謎のAV男優 黒子(くろこ)さん
黒子さん
……(nods silently)
研究所の所長 顔見 識朗(かおみ しきろう)
所長
Here is today’s lesson. "Looking alike" is not one thing. AI looks at the framework; people look at the vibe. Looking with both is best.
研究所の所長 顔見 識朗(かおみ しきろう)
所長
…But then, what exactly is this "vibe" made of?
研究員 藍子(あいこ)
藍子
Want to know?
研究所の所長 顔見 識朗(かおみ しきろう)
所長
Quit teasing.
研究員 藍子(あいこ)
藍子
Next time, we take "vibe" apart! Hair, makeup, expression, gaze… we will put numbers on which of them matters most.
研究所の所長 顔見 識朗(かおみ しきろう)
所長
I look forward to it.
研究所の猫 ピクセル(通称ピク)
ピク
Mew (walks across the keyboard and types "fffff")

Try your own "look-alikes"

  • Diagnose from a face photo: find actresses whose facial framework is close
  • Search by facial features: specify your preferences across 9 axes, such as eye size and face outline
  • Search by hairstyle: choose, title by title, the hairstyles that shape the vibe
  • The "compare with in-title faces" feature on actress pages: a way of comparing that is not affected by retouching of profile photos

研究所のほかの記事

自分の好みの顔を、AIで探すなら顔写真で診断するか、顔の特徴で探すをどうぞ。

WEB SERVICE BY FANZA