From our other sites
The story
On 5ch’s News VIP board, a thread popped up where one poster running a personal AI image-generation server started churning out illustrations on request. This time they tried a new, more photorealistic model — and found that simply specifying a “pretty face” tended to produce Western-looking features, meaning you needed negative prompts or additional LoRA training to reliably get Asian faces. Characters with little existing training data, like Blue Archive’s newer additions, proved especially tricky, leading posters to start trading prompt tips with each other.
What people said
Oh, a foreigner…
I just wanted a normal Asian bishoujo (pretty girl)
Yeah, that's me… sorry… it's annoying, isn't it
At least on 5ch you might as well talk like a real person
Fell asleep before I could reply, sorry about that
Far from annoying, I've used this a bunch of times, so I'm grateful
By the way, the punchline was that typing "Rena" from Blue Archive got me "Asuna" instead
Rena…?
Never heard the name, so I looked it up — she's a recent character
Since she's not in the training data, you'd need a LoRA for that
Yeah, she's a character who just debuted at the end of last year
I see, thanks
Haven't played Blue Archive in a while… does Mizu-oji (a character nickname, lit. "Water Uncle") still have "human rights" (gamer slang for being meta-viable)?
Easily still top-tier
Alriiight
guess I'll go check out at least the main story
Take your time and enjoy it!!
Sorry about that…
but the people using it now are skilled with prompts
they're managing to get Asian faces out of this same model… something I couldn't do…!
Oh, it can be done?
It'll be the middle of the night, but maybe I'll give it a shot after I get home from work
https://gzo.ai/i/W25vSG3.jpg
That's basically a straight-up photo
My guess at the prompt:
japanese woman, upper body, White blouse shirt, He tilted his head, black hair, long hair, light smile, closed mouth
something like that?
https://gzo.ai/i/SAsIg8U.jpg
Nice, that soft fluffy cute style is great 👍
For outfit prompts, the prompt selector has a ton of "outfit: ___" options ready to go, so definitely make use of those
Have a good day at work~
Gimme a salary for covering your bedtime duty, pwease (baby-talk spelling of "please")
Cleared out my failed attempts, tried all sorts of things, and finally got something postable
Illustration mode is like 10x harder, and I can't shake this "hmm, not quite" feeling
3D models aren't worth the trial-and-error cost, so I'm not a fan, ZE (playful speech tic)…
well, guess overcoming what you're bad at is necessary too…
yeah, the image quality does lean more Japanese now
But…
this is the "wholesome" version that hasn't been trained on NSFW content
is that really okay?
For real?
Hold on, let me look into it
Just specify a "pretty face" and… boom, Asian discrimination…
Mind picking out a fix for that on your end?
I'll give it a try
What the hell happened lmaooo
go ahead and throw it into PNG Info (to check the embedded metadata)
Prompt handling is key here
you need Japanese features in the positive prompt
and Western features in the negative prompt, or it won't work
The model? It's the original one!
https://imgur.com/xePzQP1.png
https://imgur.com/eE4Ue7q.png
https://imgur.com/V7MZOAK.png
Background and Key Points of This Topic
It’s well known that image-generation AI, due to biases in its training data, tends to output Western-looking faces unless given special instructions. The joking remark in the thread about “Asian discrimination” is just pointing at this structural quirk in the model, not making a discriminatory comment about any particular ethnicity. LoRA is a technique that trains a small amount of additional data onto an existing model to reproduce a specific character or art style — the more recently a character debuted, the less source data exists, making it harder to nail down the intended face. It’s also worth noting as background that “wholesome” and “NSFW-capable” models are often distributed and trained as separate lineages, chosen depending on the user’s purpose. On top of that, some posters in the thread noted that photorealistic, 3D-leaning models tend to show more visible flaws than 2D illustrations, making the trial-and-error process more costly.
*This article is excerpted and summarized from the 5ch (News VIP) thread “[AI Illustration] Running a Server with a 3D Model… In Progress [Server].”
Leave a Reply