Muse Muse

I was savagely judging the vulgar herd at the café I just walked out of, as if I myself were completely free of their vulgarity. No. No, that’s not it.

To qualify as one of these vulgar people, you need to meet the following conditions:

Condition one: “No seriousness allowed.”
Condition two, in my opinion: “Taste must not be too refined.”
Condition three: “No crisis about making a living; even if there is one, you may choose to actively forget it.”

What’s tricky is, I don’t know why I carry this strange air of authority. Maybe it’s like when R is about to tell a dirty joke to S or Q — they’d grin at each other, eyes glinting, full of mischief and daydreams — while I would somehow steer it toward the tone of a gynecology lecture, and they’d be itching to take notes like schoolkids.

And if I’m lucky you might just give it a little kiss
And that’s the message that we deliver to little kids
And expect them not to know what a women’s clitoris is

That’s from “The Real Slim Shady.” Q was nodding along to it one day, and I happened to have printed the lyrics, so I pointed at that part (purely to tease her). She asked what a clitoris is. I said: the clit. She went silent with a dead-serious face, like a small animal that just met a divine beast. I thought, if you want to tease these little animals, maybe just say “I miss your little hairy cunt every day,” or tell her to write “cunt” — because the moment she sees the proper word, it turns back into class notes. (I just discovered Mac’s default Simplified Chinese input doesn’t even have the character 屄. What kind of hygiene campaign is this?)

Okay, preamble more or less done. Years of a sterile emotional structure from my wife — zero tolerance for my wandering mind — forced me to hide my lecherous side, like that Kuaishou likes incident I once wrote down. I had to bury it, and I carried that concealment into everyday conversation. Plus I’ve kept the big beard for years, which is supposedly a hard sell for girls who haven’t seen much of the world.

Around March–April, full of hope for the Studio, R and ZJ and I were daydreaming, and I floated an idea: find the Japanese actress who looks most like a given face. Long ago I saw something similar on some porn site, but it only let you try once — I uploaded a colleague’s photo, the trial ended, paywall up, and the result wasn’t that accurate anyway. I wasn’t impressed.

Then the day before yesterday, resumes sent into the void, I thought: idle anyway, why not try it. So I scraped missav’s actress pages. Making anything free and public these days is no small feat — sigh, no good deed goes unpunished. For a while I’d felt that service-style agents were bad for privacy, and I’d long wanted to tinker with opencode + local LLM, except the best local model I have is Qwen3-35B. So this time I just went with opencode’s free model, muse spark 1.3 — and the thing built it in one evening. Which plunged me into deep suspicion about whether this thing has any commercial value at all. I hung it on a temporary Vercel domain, the technical writeup is here. I call it muse.face, and I hope it brings me inspiration driven by primal desire.

The creepy part is how accurate it is. I fed it promo images from FANZA emails — unless the actress isn’t in the library, everything clears 0.6 similarity. I personally took part in the University of Greenwich Super-Recognisers research, studied for two or three years as a subject — my own face recognition is above average, a so-called Super Recogniser. And even to me, some of these actress faces look confusingly similar — yet muse.face tells them apart cleanly!

So naturally, my first query was little Q’s photo. All women of age whom I know — consider yourselves warned~. Sadly, Q wears glasses in the photo, and muse gave a limp sub-0.3 match. Honestly the library of 4,138 actress faces probably just doesn’t contain anyone who looks much like her. A married woman like Q, strictly speaking, doesn’t meet the entry bar for the formal industry at all — she’d sooner appear in the uncensored amateur zone. Getting that data would take another level of technique. I already have the idea, and the agent has the plan.

After that I collected, one by one, former colleagues, classmates, acquaintances, my wife’s besties — the thrill of transgression is delicious. A few interesting cases, in order:

Sample A: my wife’s ex-bestie. Highest similarity of the lot, the biggest prize of this inspiration project. Saved.
Sample B: ex-colleague. Low accuracy. I blame the training set being mostly work covers with too few candids — next I plan to have the agent crawl actresses’ social media for a fresh wave of RAG fodder.
Sample C: ex-colleague. Same problem, low accuracy. That’s when it dawned on me (obvious in hindsight) — what I call the “beauty redshift of the Japanese adult film industry.” In plain words: C’s face could never get into the label-backed industry; the odd amateur gig for quick cash, sure.
Sample D: ex-colleague. The closest I’ve come to the edge of physical cheating — but she’s in debt and possibly dealing with an endocrine condition, so my interest cooled rationally. muse did find an extremely similar actress though!
Sample E: classmate. Fits the redshift elimination theory.

Anyway, the deeper it gets the more fun it is. Writing this far, I’m not sure I dare publish it on the blog directly.