Ethnicity prompts are appearance vocabulary that nudges an adult character’s skin tone and features toward a described look. They matter because every checkpoint has a strong default face, so without a deliberate descriptor you get the same look every time. Treat these as neutral, respectful descriptors of adult subjects only, weight them when the model ignores them, and never target a real person or lean on stereotypes.
A note before the tag bank: use this responsibly
Everything below describes adult, fictional, original characters, eighteen and over, full stop. Ethnicity in a prompt is descriptive vocabulary for skin tone and facial features, nothing more. Keep it neutral and respectful. Do not use it to caricature or stereotype any real group, do not pair it with costume-y cultural cliches to make a character an ethnic prop, and never prompt toward the likeness of a real, named person or celebrity. Diversity in your output is a good thing to prompt for on purpose. Fetishizing a real ethnic group is not the goal and it produces worse, flatter art anyway. With that framing set, here is how the vocabulary actually behaves in Stable Diffusion.
The practical reason to get this right, beyond the ethical one, is that respectful descriptive prompting also produces better images. When you describe a person as an individual with specific features and a specific complexion, the model renders a person. When you lean on lazy stereotype associations, the model reaches for the flattest, most cliched version of a look it saw in training, and the result feels like a costume rather than a character. Precision and respect point the same direction here.
What these prompts control and why they matter
Every checkpoint you download has a default human baked into it. Train a realistic model heavily on one region’s photography and its untagged output will drift toward that region’s average face and skin tone, over and over. This is the single reason people complain that all their generations look the same. The model is not broken, it is showing you its center of mass. Descriptive appearance vocabulary is how you move off that center on purpose.
Skin tone and facial features are the two things these descriptors actually control. A good descriptor shifts complexion, and it nudges feature vocabulary the model has learned to associate with a look: nose shape, eye shape, lip fullness, brow, hair texture. Understanding that these are appearance nudges, not identities, keeps your prompting honest and also more effective, because you can reach the same result by describing the individual features directly when a broad descriptor misfires.
There are really two ways to steer appearance, and knowing which to reach for saves a lot of rerolling. The broad-label route uses a regional descriptor as shorthand for a bundle of features, which is fast but blunt and easy for the model to ignore. The feature-led route names the actual attributes you want, complexion, eye shape, nose, lip fullness, hair texture, which is more verbose but far more controllable and far less likely to trip a stereotype. Strong prompting usually starts broad and then refines with feature tags wherever the label misses.
This matters most in NSFW work because skin is usually the largest surface in frame. Getting complexion, undertone, and texture right is the difference between a real-looking body and a plastic one, and it pairs directly with your body type prompts and skin texture prompts.
The copy-paste descriptor bank
Grouped by what the tokens do. Combine one skin-tone tag, optionally one regional descriptor, and one or two feature tags.
Skin tone and undertone
fair skin, light skin, porcelain skin, pale complexion
light tan skin, olive skin, medium skin tone, warm beige skin
tan skin, sun-kissed skin, golden brown skin, bronze skin
deep brown skin, rich dark skin, ebony skin, deep melanated skin
cool undertone, warm undertone, neutral undertone, golden undertone, olive undertoneRegional and ethnic appearance descriptors (neutral appearance vocabulary)
East Asian features, Southeast Asian features, South Asian features, Central Asian features
Middle Eastern features, Mediterranean features, Latina features, Hispanic features
Nordic features, Slavic features, Scandinavian features, Eastern European features
West African features, East African features, Afro-Caribbean features
mixed ethnicity, multiracial features, ambiguous ethnicityIndividual facial features (the precise controls)
monolid eyes, almond eyes, hooded eyes, round eyes, wide-set eyes, upturned eyes
full lips, defined cupid's bow, thin lips, wide smile, heart-shaped lips
high cheekbones, soft round face, angular jaw, strong brow, soft nose, aquiline nose, button nose
thick dark eyebrows, straight black hair, textured curly hair, tight coils, wavy dark hair
freckled skin, even complexion, dimples, defined jawline, broad nose bridgeThe feature block is your escape hatch. When a regional descriptor gets ignored or comes out as a costume, describe the actual features and the model will usually comply, because feature tokens are more concrete than a broad geographic label. Combining two or three feature tags with a skin-tone tag gives you a specific, individual face rather than an averaged stereotype, and it is the single most reliable way to get intentional diversity into a set.
Reference grid: descriptor to what it nudges to caveat
| Descriptor | What it nudges | Caveat |
|---|---|---|
| fair skin, cool undertone | light complexion, pink-neutral base | can over-brighten, verify texture holds |
| olive skin, Mediterranean features | warm medium complexion, defined features | drifts back to default on weak models |
| tan skin, Latina features | golden-brown complexion, warm tone | may add cliche wardrobe, keep neutral |
| East Asian features, monolid eyes | eye and face shape shift, lighter tone | broad label, use feature tags to refine |
| South Asian features, deep warm skin | warm deep complexion, strong brows | often ignored, weight up and add features |
| deep brown skin, West African features | deep complexion, fuller features, coils | many models bias light, weight strongly |
| mixed ethnicity, ambiguous ethnicity | blended features, in-between tone | unpredictable, lock with seed and reference |
Read this as directional, not deterministic. Different checkpoints respond differently, and a descriptor that lands cleanly on one model gets swallowed by another’s default on the next. The caveat column is the part to memorize, because each descriptor has a characteristic way it fails, and knowing it lets you pre-load the fix into your negative prompt or your feature tags.
Full example prompts
Deep skin tone, realistic finish
Positive:
photorealistic portrait of an adult woman, (deep brown skin:1.3), rich even complexion,
warm undertone, (West African features:1.2), full lips, high cheekbones, textured coily hair,
natural skin texture, soft studio light, 85mm, shallow depth of fieldNegative:
washed out skin, grey skin, ashy skin, plastic skin, lightened skin, overexposed,
deformed, blurryThe negative matters here. Many realistic models bias toward lightening, so lightened skin, washed out skin in the negative keeps the complexion where you asked for it. The feature tags, full lips and high cheekbones and coily hair, give the model concrete targets so it does not simply darken its default face and call it done.
Warm medium complexion
Positive:
portrait of an adult woman, (olive skin:1.2), warm undertone, Mediterranean features,
defined brows, soft nose, wavy dark hair, natural makeup, golden hour light, cinematicNegative:
pale skin, cool grey tone, cakey foundation, orange oversaturation, uneven skin toneFeature-led approach when the label fails
Positive:
portrait of an adult woman, light tan skin, (monolid eyes:1.3), (almond eyes:1.1),
straight black hair, soft round face, even complexion, natural light, realisticHere we skip the broad regional label entirely and describe features directly, which is often the more reliable route. When a checkpoint keeps swallowing your regional descriptor, this feature-led rewrite is the first thing to try before you reach for weighting or a different model.
Common failure modes and the fix
The model ignores the descriptor and reverts to default. You ask for deep skin and get medium, or ask for specific features and get the checkpoint’s house face. This is the model’s center of mass winning. Fix it three ways: weight the descriptor up, (deep brown skin:1.4), describe the individual features instead of relying on a broad label, and if the model simply cannot render a look, switch to a checkpoint that can. Some Stable Diffusion checkpoints are far more ethnically flexible than others, and no amount of weighting rescues a model that never saw the range. Weighting technique itself is covered in the prompt weighting guide.
Stereotyped or costume-y output. You describe an appearance and the model wraps it in cultural cliche: traditional dress, props, an exoticized backdrop you never asked for. This is the training data’s laziest association surfacing. Fix it by keeping your prompt strictly about the person, complexion and features, and adding the unwanted cliches to the negative. Explicitly prompt the setting and wardrobe you actually want so the model has no gap to fill with a stereotype. This is also just the respectful way to work, and it produces a modern, individual character instead of a caricature.
Feature blending on mixed or ambiguous descriptors. Broad or blended descriptors produce inconsistent, sometimes muddy results because the model is averaging across a wide space. If you want a specific blended look, define it with concrete feature tags rather than the word mixed, lock a seed, and use a reference image. An img2img or reference pass gives the model a concrete target to hold instead of an average to guess at.
Complexion looks flat or plastic. A correct skin tone with no texture still reads fake. Layer in natural skin texture, visible pores, subsurface scattering and run a detail pass. Skin realism is its own skill, and it carries directly over from the realistic AI work fundamentals.
Tone shifts under colored lighting. A warm grade or a colored gel can pull a complexion off target, making deep skin look muddy or fair skin look sallow. Grade neutrally first, confirm the complexion is right, then add color. Do not let the color grade do the skin tone’s job, because a warm push that looks great on one tone wrecks another.
Features fight the body. Occasionally a strong facial descriptor conflicts with a body-type tag and the model splits the difference oddly. Keep the face and body descriptors coherent, describe one person consistently, and if a mismatch appears, resolve the face with a detail pass while holding the body from the base generation.
Keeping ethnicity consistent across a set
Complexion and features are part of a character’s identity, so consistency here is really character consistency. The same tools apply.
Lock the descriptor block exactly. Same skin-tone tag, same feature tags, same order, same weights, in every prompt. Paraphrasing is how a character’s complexion quietly drifts a shade lighter over a set, and a lighter drift is especially common because most models pull that way by default.
Lean on a reference. The most reliable way to hold a specific face and tone across many images is a reference image plus the character consistency techniques you would use for any recurring character. A trained LoRA or a solid reference pins the look far better than text alone, and it is the honest answer to the feature-drift problem. Text descriptors set the neighborhood, a reference nails the address.
Control lighting per image. Because color grade shifts perceived tone, grade each image in the set consistently. If image one is warm and image five is cool, the same complexion will read as two different people. The consistent photo set workflow covers holding lighting steady across a series.
Verify in a strip. Put the set side by side and check that complexion and features hold. Drift is invisible one image at a time and obvious in a row. Fix any outlier with a targeted inpaint rather than rerolling, since a reroll risks losing everything else the frame got right.
Why checkpoints bias toward one look, and how to counter it
It helps to understand where the default comes from, because it tells you which lever to pull. A checkpoint’s bias is a direct fingerprint of its training data. A model fine-tuned mostly on one region’s photography learns that region’s average as the path of least resistance, so an untagged prompt lands there every time. This is not a value judgment by the model, it is statistics, and it means the fix is always to give the model a stronger, more specific signal than its default.
The three levers, in order of strength, are text, weighting, and reference. Text descriptors set the general neighborhood and are enough when the model has good coverage of a look. Weighting pushes harder when the descriptor is being swallowed, and moving from a broad label to concrete feature tags is often more effective than raising a number. Reference is the strongest lever: a reference image, an img2img pass, or a trained LoRA gives the model a concrete target it cannot average away, and it is the only reliable path when a checkpoint’s coverage of a look is genuinely thin.
This is also why checkpoint choice matters as much as prompt craft for diversity. Some Stable Diffusion checkpoints were trained on broad, varied data and render a wide range of complexions and features cleanly with light prompting. Others are narrow, and no weighting rescues a look the model never learned. If you find yourself fighting a checkpoint on every generation to reach a look it resists, the honest fix is a different checkpoint or a dedicated LoRA, not a heavier weight. For the deepest control over token strength when you do prompt, the prompt weighting guide covers how to escalate without breaking the rest of the image.
One final habit that pays off: build variety into your defaults. If every image in your library reverts to the same face, decide on a deliberate rotation of complexions and feature sets and lock each as a reusable block, so diversity is a choice you make once rather than a fight you have every prompt.
Where to go next
Ethnicity vocabulary is one input into a believable person, and it works best alongside the pieces around it. Pair it with the skin texture prompts so complexion reads real, the body type prompts for proportion, the hair prompts for texture that matches the look, and the makeup prompts on top. When you need the same person across many frames, the character consistency techniques and prompt weighting guide do the heavy lifting.
Frequently asked questions
How do I describe ethnicity in a prompt respectfully?
Keep it to neutral appearance vocabulary about skin tone and facial features of an adult, fictional character, and nothing else. Do not add cultural cliches, costumes, or props to turn a character into an ethnic stereotype, and never prompt toward a real named person. Describing complexion and features respectfully produces better art than fetishizing or caricaturing any real group.
Why do all my characters look the same even with different descriptors?
Every checkpoint has a default face and skin tone baked in from its training data, and untagged output drifts toward that center of mass. To move off it you must prompt a descriptor explicitly and often weight it up. If the model still refuses to render a look, it likely never saw enough of that range, and you need a more flexible checkpoint.
The model ignores my ethnicity descriptor. What do I do?
Try three things in order: weight the descriptor higher, for example deep brown skin at 1.4, describe the individual features directly instead of using a broad regional label, and switch to a checkpoint known for ethnic flexibility if the model simply cannot render the look. Feature tags like monolid eyes or full lips are more concrete and get ignored less often than geographic labels.
How do I avoid stereotyped or costume-like results?
Keep your prompt strictly about the person, complexion and features, and never let the model fill gaps with cliches. Prompt the setting and wardrobe you actually want explicitly, and put unwanted traditional-dress or prop associations into your negative prompt. This is both the respectful approach and the one that produces cleaner, more modern-looking images.
Why does deep skin come out looking washed out or grey?
Many realistic checkpoints bias toward lightening skin, so a deep tone gets pulled up and desaturated. Weight the skin-tone tag up and add lightened skin, washed out skin, and ashy skin to your negative prompt. Grade your lighting neutrally first and confirm the complexion is correct before adding any warm or colored grade that could shift it.
How do I keep the same complexion and features across a whole set?
Lock the descriptor block exactly, the same skin-tone and feature tags in the same order and weights in every prompt, and do not paraphrase. Use a reference image or a trained LoRA to pin the look, since text alone drifts. Grade lighting consistently across the set because color grade changes how a complexion reads.
Are mixed or ambiguous ethnicity descriptors reliable?
Not very, because the model averages across a wide space and results come out inconsistent or muddy. If you want a specific blended look, define it with concrete feature tags rather than the word mixed, lock a seed, and use a reference image. Giving the model a concrete target beats asking it to guess an average.
Why does skin tone look right in text but wrong on screen?
Colored lighting and color grading shift perceived complexion, so a correct skin-tone tag can still render muddy under a warm gel or sallow under a cool one. Grade neutrally, confirm the tone is right, then apply color. Also add natural skin texture and visible pores, because a correct tone with no texture still reads flat and plastic.
