Comparing impression comedians and the real people with AI

AV女優AI顔診断研究所(AIAVMATCH LAB)・2026/11/02

研究員 藍子(あいこ)
藍子
Director! We compared impression comedians and the people they impersonate — with AI!
研究所の所長 顔見 識朗(かおみ しきろう)
所長
Oh, those acts are masterful, I tell you. Their faces must be dead ringers too, no?
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 and out with it.
研究員 藍子(あいこ)
藍子
Let me give you the conclusion first: measured by facial structure, all four pairs were pretty much within the range of complete strangers.
研究所の所長 顔見 識朗(かおみ しきろう)
所長
What in the world?!

How we compared them

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, it is the same method as last time in "What does it mean for faces to look alike". We convert a face into 128 numbers and measure how much the two sets of numbers point in the same direction.
研究員 藍子(あいこ)
藍子
For the photos, we used one photo per person that is available under a public license (CC). And the impersonation pairings are only the four that were listed as repertoire in the subjects’ own profile articles.
特任研究員(統計学) 分散 さゆり(ぶんさん さゆり)
分散
Also, some of the names are partially masked. We do not publish the photos either. The measurements were done here at the lab.

Results: the "closeness" of the four pairs

Impressionist → subjectCloseness (up to 0.3)ClosenessRank among stranger pairs
キンタ○ー。 → 前田○子
0.1421.6%
山本高○ → 織田○二
0.218.1%
ホ○ → 木村○哉
0.1127%
清水ミ○コ → 宇多田ヒ○ル
0.225.4%

The vertical line in the bar is the cutoff for the top 5% of stranger pairs (0.22). "Rank among stranger pairs" is the share of the 74 pairs other than the impersonation pairs that were closer than that pair.

研究員 藍子(あいこ)
藍子
The four pairs averaged 0.17. The 74 pairs of strangers averaged 0.06, and even the closest pair among them was 0.25.
研究所の所長 顔見 識朗(かおみ しきろう)
所長
So they are a little closer than the average stranger, but…
研究員 藍子(あいこ)
藍子
Right. But not runaway close. Even the two closest pairs (0.21 and 0.22) had pairs of strangers about as close.
研究所の所長 顔見 識朗(かおみ しきろう)
所長
So what I felt was a "dead ringer" was not the facial structure after all…

So what creates "looking alike"?

特任研究員(元撮影カメラマン) 画角 さだめ(がかく さだめ)
画角
Impressions are not the art of copying a face, you know. Voice, expression, gaze, mannerisms, signature poses. You pick a few of the things that make someone who they are, and exaggerate them.
特任研究員(メイク講師) 盛田 映子(もりた ばえこ)
盛田
Hairstyles, outfits and makeup too. It is close to bringing in a photo and asking, "Make me look like this person."
研究所の所長 顔見 識朗(かおみ しきろう)
所長
Don’t bring up my hair salon story again.
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
What AI measures is the facial structure that exists "before" that exaggeration. So being good at impressions and having a similar bone structure are two different things.
謎のAV女優 ナナシ
ナナシ
…"Looking alike" is not something the impersonator creates. It is completed inside the head of the person watching.
謎のAV男優 黒子(くろこ)さん
黒子さん
……(applauds silently)

What we cannot claim

特任研究員(統計学) 分散 さゆり(ぶんさん さゆり)
分散
Hold on. Everything so far is an "observation", not a conclusion — just to be careful.
研究員 藍子(あいこ)
藍子
Right… Only four pairs, and only one photo per person. And the size of the face in the photos varied quite a lot (48 to 385 pixels wide). The smaller the face in the photo, the more unstable the numbers become.
特任研究員(統計学) 分散 さゆり(ぶんさん さゆり)
分散
The angle, the lighting and the year the photo was taken also move the numbers. So we cannot say things like "this comedian does not resemble the person" from these numbers. All we can say is that, within the range measured from the photos of these four pairs, the structural closeness was not very different from that of strangers.
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
To really verify it, we would need dozens of photos per person, with the angle and expression aligned, and take averages. The same thing we do with the actresses in the face database.

Today’s lesson

研究所の所長 顔見 識朗(かおみ しきろう)
所長
Here is today’s lesson. What creates "looking alike" is not only the facial structure. Hair, makeup, expression, gaze… people feel that someone looks alike depending on what you choose to emphasize.
研究員 藍子(あいこ)
藍子
So next time, we finally 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
Meow (grooming himself on top of the keyboard)

How we measured

  • We used photos from Wikimedia Commons under public licenses (CC), one per person, 13 in total. The photos are not published on this site
  • From each photo we took the single largest face and turned it into 128 numbers with a face recognition model (the same model as the face diagnosis)
  • Closeness is how closely the two number vectors point in the same direction (cosine similarity). The closer to 1, the more they point in the same direction
  • For the comparison among strangers, we used the 74 pairs, out of the same 13 photos, that are not impersonation pairings
  • Some names are partially masked. This article is an observation for entertainment and is not an evaluation of any specific person

More articles and tools

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