Counting what is inside AV titles with AI

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! We found something big!
研究所の所長 顔見 識朗(かおみ しきろう)
所長
What is it, what is it — so noisy first thing in the morning.
研究員 藍子(あいこ)
藍子
We counted what is inside AV titles — 23,292 of them, with AI. It judges the thumbnails one frame at a time and works out how many minutes each act gets.
研究所の所長 顔見 識朗(かおみ しきろう)
所長
Oh, that does sound interesting. So, what did you find?
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
Stop teasing me.
研究員 藍子(あいこ)
藍子
Let us start with the basics. The most common was fellatio. It appears in 96% of titles, and in titles that have it, the median length is 13.3 minutes.
研究所の所長 顔見 識朗(かおみ しきろう)
所長
Hmm, well, that is about what I expected.

Time breakdown by act (23,292 titles)

ActTitles containing itLength in those titles (median)Share of all time
フェラ96%13.3 min9.8%
騎乗位85%6.5 min5.2%
バック83%6.0 min4.1%
正常位78%5.5 min3.5%
オナニー76%4.5 min3.3%
手コキ72%5.5 min3.5%
クンニ65%3.3 min2.0%
背面騎乗位43%2.3 min0.9%
パイズリ27%3.0 min0.7%
しゃがみ騎乗位19%2.0 min0.4%
研究員 藍子(あいこ)
藍子
Now for the main event. When we re-tallied just the VR titles — 1,795 of them…
研究所の所長 顔見 識朗(かおみ しきろう)
所長
Out with it, then…
研究員 藍子(あいこ)
藍子
VR was almost all cowgirl! It appears in 90% of the titles, with a median length of 7.8 minutes. Missionary was only 23%, and doggy style only 24%!
研究所の所長 顔見 識朗(かおみ しきろう)
所長
Whaaat?!
研究員 藍子(あいこ)
藍子
In non-VR titles it is missionary 83%, doggy style 88%, cowgirl 85%. It is almost exactly reversed.
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
It is all decided by where the camera sits. In VR, the viewer’s eyes become the camera. A position where your partner is on top, looking down at you, simply makes the better picture.
研究所の所長 顔見 識朗(かおみ しきろう)
所長
I see — so how you shoot it decides what is in it.

Time breakdown for VR titles only (1,795 titles)

ActTitles containing itLength in those titles (median)Share of all time
騎乗位90%7.8 min9.9%
フェラ85%5.0 min5.8%
オナニー78%4.2 min3.8%
手コキ58%3.5 min2.3%
しゃがみ騎乗位51%3.2 min2.3%
背面騎乗位46%2.3 min1.4%
パイズリ41%3.0 min1.2%
バック24%2.7 min0.6%
正常位23%3.0 min0.6%
クンニ10%2.0 min0.2%
特任研究員(統計学) 分散 さゆり(ぶんさん さゆり)
分散
Hold on. Isn’t that sample biased?
研究員 藍子(あいこ)
藍子
Ugh… We only looked at titles the lab has been analyzing. They are mostly titles on sale and standalone titles — not a random sample of everything on FANZA.
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
Also, model-wise, you see, the judgment runs on thumbnail stills — one frame every 10 to 20 seconds. We only count stretches of 40 seconds or more as a segment, but misses happen all the time. And things that happen at the same time, like fellatio and handjobs, get counted in both!
特任研究員(統計学) 分散 さゆり(ぶんさん さゆり)
分散
So read these as a "tendency", not exact numbers. VR having so much cowgirl is a big enough gap that I do not think bias explains it away.
謎のAV女優 ナナシ
ナナシ
…In VR, the gaze is everything.
謎のAV男優 黒子(くろこ)さん
黒子さん
…Cowgirl is, honestly, a blessing.
研究所の所長 顔見 識朗(かおみ しきろう)
所長
Today’s lesson, then. Where the camera sits decides what a title contains.
研究所の猫 ピクセル(通称ピク)
ピク
Meow (sleeping on the keyboard)

How we counted

  • We used the thumbnails FANZA publishes for each title (stills cut from the whole title every 10 to 20 seconds)
  • An image recognition model judged the act in each frame, and stretches lasting 40 seconds or more count as one segment
  • A title counts as "containing" an act if the act appears for at least one minute
  • Acts that happen at the same time are counted separately, so totals can exceed the length of the title
  • The titles are ones the lab has been analyzing; they are not a random sample of all FANZA titles
  • The judgments are made by machine and can be wrong. These numbers are for seeing tendencies

You can also see the breakdown for each title

For titles we have analyzed, the title page shows the time for each act, as "What is in this title (AI estimate)". You can also open them from Search by act time breakdown, which lines up titles by how long each act runs. You can also find them by searching titles by title name or product code.

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