Writing 'asian woman' gives you a generic averaged-out face. Specify Japanese, Korean, Chinese, or Thai and the output transforms. Here's exactly what changes and how to prompt each.
A friend (yes, a real friend) told me ximages was broken because "Japanese and Chinese come out looking the same."
I asked to see his prompt. He'd written asian woman.
Bro. That's not Japanese. That's averaged Asian — the model literally averaged every face across East and Southeast Asia. Of course the output looks like nobody in particular. Specify the country.
This post breaks down what actually changes between these four country tags, and how to write prompts that get you each specific vibe.
bigASP and SDXL organize training data by caption tags. The asian tag is a massive grab-bag — Japanese AV stills, Korean idol photos, Chinese livestream caps, Thai pageant shots — all lumped together. The model's mental "prototype" for that tag is an average across all of it.
asian woman = 60% East Asian average + 30% Southeast Asian + 10% Central Asian
That average has no character. Switch to a specific country and the model uses a much narrower subset. The features become legible.
Easiest of the four to get good output from. Japanese adult photography is heavily represented in SDXL training data (no need to elaborate on why).
Default characteristics:
jk uniform or office lady tags hit 90%+ accuracyTemplate:
young japanese woman, mid 20s, long black hair, fair skin,
slight smile, sitting on bed, white blouse, soft natural daylight,
evenly lit bedroom, natural skin pores
Watch for:
anime eyes, doll eyes, oversized eyesjapanese girl skews young — use woman for adult content alwaysModern Korean aesthetic: pointed chin, double eyelid, fair skin, long straight hair.
Default characteristics:
kpop idol tag triggers stage/performance aestheticTemplate:
young korean woman, late 20s, long straight black hair, v-line jaw,
double eyelid, natural makeup, looking at camera, slight confident smile,
white t-shirt, modern apartment, soft natural daylight,
natural skin pores
Watch for:
unique facial features or specific traits to differentiatekorean girl group triggers multi-person scenes (5-7 people). Use singular phrasingThe hardest of the four to get right. The chinese tag includes historical drama, modern fashion, internet celeb, art photography — wildly different aesthetics in the training set.
Default characteristics:
cheongsam for qipao, hanfu for traditional dress, otherwise must say modern chinese womanTemplate:
young chinese woman, late 20s, long black hair, warm skin tone,
gentle smile, standing in a modern living room,
white sweater, cream pants, soft afternoon daylight through window,
natural skin pores
Watch for:
modern chinese woman or a city tag like shanghai womantraditional hanfu or cheongsam, qipao work wellThe most underrated tag. Southeast Asian features are distinct enough to be clearly different from Japanese/Korean/Chinese output.
Default characteristics:
Template:
young thai woman, mid 20s, long dark brown hair, sun-kissed honey skin,
bright smile, walking on a beach, bikini, palm trees in background,
late afternoon golden sunlight, beach atmosphere, natural skin pores
This is the only country template where I keep golden sunlight in — Thailand's tropical light actually looks like that, so it doesn't read as cinematic-cliché.
Watch for:
ladyboy will trigger trans subject matter — pay attention to whether you want this| Dimension | Japanese | Korean | Chinese | Thai | |---|---|---|---|---| | Face shape | Soft with subtle definition | V-line (thinnest) | Moderate dimension | More dimensional | | Skin tone | Cool pale | Cool pale | Warm pale | Warm honey | | Default mood | Quiet, contained | Confident, outward | Subtle, traditional | Tropical, vibrant | | Makeup level | Natural, light | Heavier, defined | Moderate | Moderate | | Best setting | Bedroom / school | Modern apartment | Living room / period | Beach / outdoor | | Failure mode | Anime face drift | Identical "famous face" | Drifts to period costume | Skin tone pulled too pale |
Want Asian features with European bone structure? Specify the mix:
half japanese half european / asian-european mixed race /
half korean half american
The model genuinely tries to blend features. Avoid writing just mixed race — too generic, model picks two random ethnicities.
❌ oriental dated colonial term, tags low-quality images
❌ exotic tagged on low-quality data
❌ asian girl underage risk (use woman not girl, always)
❌ chinese doll CG / figurine training samples
Beginner verification trick: write a template, then generate 4 images changing only the country tag (one each for Japanese / Korean / Chinese / Thai). Look at them side by side. If you can instantly tell which is which, your prompt is working. If all four look similar, your other descriptors are diluting the country tag — drop some descriptors so the country tag's weight comes back.
Related: 60-second AI girlfriend beginner guide · 12 prompt traps every beginner falls into · 12 popular outfit prompts