Halfling and Gnome Portrait Prompts: How to Get Adults, Not Children

Why does AI draw halflings and gnomes as children?

Image models learn visual correlations, not lore. In their training data, the features that define halflings and gnomes — short stature, round faces, large expressive eyes, smooth skin — co-occur overwhelmingly with photographs and art of children. When your prompt supplies those cues and nothing to contradict them, the model resolves the ambiguity toward its statistical center: a kid.

Worse, child cues cascade. Once the model commits to one — say, a small body — it pulls in the rest of the cluster: bigger irises, a button nose, poreless skin, a shorter jaw. That's why a prompt that only says adult halfling still comes back looking twelve. One weak adult vote loses to four correlated child votes.

There's also a data-scarcity problem. "Halfling" appears in far less training art than "elf" or "knight," so the token itself is weak, and the model leans harder on the proportion cues you gave it. "Gnome" has the opposite issue: the token is saturated with garden-gnome imagery — the ceramic old man with the red cone hat — so male gnomes at least render old, but they render as ornaments, and female gnomes fall back into the child cluster.

The strategy that follows from the mechanism: don't argue with the model about age in the abstract. Feed it concrete visual features that never appear on children, and starve it of the words that trigger the cascade.

Which age anchors force adult faces — and which words don't work?

Age anchors work in tiers. The strongest are features that flatly cannot appear on a child's face:

• Facial hair — long grey sideburns, stubbled jaw, braided white beard (gnomes). The single most reliable anchor for male characters.
• Grey in the hair — grey-streaked brown curls, silver-shot auburn hair. Works for any gender and keeps the character middle-aged rather than ancient.
• Skin texture — weathered, sun-creased, deep crow's feet, laugh lines around the eyes. These override the smooth-skin default directly.

Mid-tier anchors support the strong ones but rarely win alone: middle-aged, in her late forties, veteran, world-weary. Numeric ages help slightly more than the word "adult" but still behave like soft votes.

What fails outright: adult, mature, grown-up, and negations like not a child. Diffusion models handle negation poorly in positive prompt text — "not a child" mostly reads as "child."

What actively backfires: cute, tiny, little, youthful, button nose, doll-like, chibi, and adorable. Each is a direct deposit into the child cluster. Diminutive is sneaky — it's a stature word, but models treat it as an age word.

The working rule: stack at least three strong anchors per prompt, and repeat them verbatim in every regeneration. If you're keeping the character across sessions, that fixed wording doubles as your consistency spec. The prompt generator bakes age and build fields into the trait spec for exactly this reason.

How do you signal small stature without triggering child proportions?

The cleanest answer: don't. In a bust portrait or head-and-shoulders close-up, height literally cannot appear in the frame. Let the face carry the race — curly hair, sideburns, weathered features for a halfling; the prominent nose and wild hair for a gnome — and skip stature words entirely. This is the highest-hit-rate approach for character sheets and VTT tokens, where tight framing is what you wanted anyway.

When you need a full-body shot, you have two safe tools:

1. Name the proportions, not just the height. Three feet tall alone gets you a toddler, because three-foot humans in training data are toddlers. Pair it: three feet tall with adult head-to-body proportions, stocky compact build. Adults are roughly seven heads tall; toddlers are about four — "adult proportions" points the model at the right template.
2. Show scale through props instead of adjectives. A tankard held in both hands, a human-scale doorway behind the character, a shortsword worn like a full sword, tools that look oversized in the grip. Scale props read as "small adult in a big world," while the words small and tiny read as "child."

Build words help too: stocky, barrel-chested, broad-hipped, sturdy are adult body descriptions that happen to fit both races. A child is small and slight; a halfling farmer is small and solid. Give the model the second one.

How do halflings and gnomes differ visually in a prompt?

They're both Small races around three to four feet tall in D&D terms, but their portrait anchors barely overlap, and mixing them is why so many results look like generic "fantasy little person."

Halfling anchors

A halfling reads as a comfortable, weathered rural adult: round warm face, curly hair, long sideburns on men (per D&D lore halflings rarely grow full beards — sideburns are the canonical facial hair), bright practical eyes, and bare, leathery, curly-haired feet if the framing shows them. Dress them in traveler's practicality: wool, coarse linen, a well-worn leather vest, earth tones. Think prosperous farmer or tavern regular, not woodland sprite.

Gnome anchors

A gnome is sharper and stranger: a prominent nose (the load-bearing feature — name it), bright wide eyes, wild wispy hair that defies gravity, pointed ears, ruddy or tan skin, and for men a full beard, often braided or singed. Gnomes carry their profession on their body: brass goggles pushed up on the forehead, ink-stained fingers, pockets full of gears, a jeweler's loupe on a chain.

One gnome-specific warning: never write pointed hat, red cap, or garden anywhere near the word gnome unless you want the ceramic lawn ornament. Anchor the fantasy context instead — workshop, arcane laboratory, tinker's bench — and the garden-gnome pull fades.

What do finished halfling rogue and gnome artificer prompts look like?

The halfling rogue is one of D&D's most-played pairings — 1.8% of the 100,000-plus characters in FiveThirtyEight's 2017 analysis of D&D Beyond data, one of the strongest race-class correlations in the dataset. It's also the hardest version of this problem: rogue prompts want hooded, small, sneaky, and nimble, and three of those four feed the child cascade. Keep the hood down or pushed back so the age anchors on the face stay visible, swap small and sneaky for wiry and watchful, and let the gear say rogue.

Bust portrait of a middle-aged halfling rogue, adult head-to-body proportions, weathered tan face with deep laugh lines and a thin scar through one eyebrow, grey-streaked brown curls, long sideburns, sharp wary hazel eyes, worn black leather armor with scuffed buckles, dagger hilt visible at the shoulder, dim candlelit tavern behind him, warm amber key light with deep shadows, painterly digital art

The gnome artificer is the flavor-defining gnome build, and it's actually the easiest small-race prompt to get right, because artificer props are age anchors: goggles, soot, singed hair, and decades of workshop wear all scream "has been doing this for forty years."

Head-and-shoulders portrait of an elderly gnome artificer, adult proportions, prominent hooked nose, bright copper eyes, cracked brass goggles pushed up on his forehead, wild white hair singed at the tips, deep crow's feet and soot-smudged ruddy skin, leather work apron over a patched wool coat, tiny clockwork gears braided into his beard, warm workshop lamplight, detailed digital painting

Both prompts follow the same skeleton: framing, race + class, proportion phrase, three or more age anchors, gear, lighting, style. Swap the class layer and the skeleton holds for any build.

How do the prompts change per generator?

The age-anchor vocabulary is universal; the packaging isn't.

Midjourney

Use the prose prompts above and add --ar 2:3 --style raw. Raw mode matters here more than for most subjects: Midjourney's default aesthetic pass beautifies and smooths faces, which erases exactly the wrinkles and weathering you stacked. You can append --no child, toddler as extra insurance, but treat it as a nudge — the positive anchors do the real work.

Stable Diffusion

Convert to front-loaded tags and use the negative prompt slot, which is the one place negation genuinely works:

portrait of a middle-aged halfling woman, adult proportions, weathered freckled face, laugh lines, grey-streaked auburn curls, green wool traveling cloak, leather satchel, warm firelight, painterly fantasy art, sharp focus
Negative prompt: child, kid, toddler, chibi, oversized eyes

Keep the negative list short and targeted — five terms aimed at the actual failure beat a fifty-term boilerplate wall, especially on SDXL-class models.

ChatGPT

ChatGPT's image model follows long prose literally, so the full descriptive prompts work as-is — it's the tool most likely to actually place the scar through the correct eyebrow. Hand it the finished description; don't ask it to invent one, or it will drift cute.

Flux

Flux has no negative prompting, so everything must be stated positively. Lean harder on the anchors: name the age band, the skin texture, and the proportions in one clause each. Flux's prose comprehension is strong enough that a halfling woman in her fifties with the compact build of an adult lands reliably.

If you'd rather not hand-tune per tool, the free generator composes the base prose prompt from trait fields — race, age, build, framing, lighting — and you adapt the packaging from there.

Frequently asked questions

Does writing 'hobbit' instead of 'halfling' help?

Often, yes. Training data for 'hobbit' is dominated by adult characters from the Lord of the Rings films, so the token itself pulls adult. The trade-off is that it also pulls Shire styling — waistcoats, green doors, movie-specific faces — which may not fit your character. If you use it, keep your own gear and setting descriptions strong so the film aesthetic doesn't take over.

Why does my gnome keep coming out as a garden gnome?

The word 'gnome' is saturated with lawn-ornament imagery, so the model defaults to a ceramic old man in a red cone hat. Avoid the words 'pointed hat,' 'red cap,' and 'garden' entirely, and anchor a fantasy context instead: a cluttered workshop, an arcane laboratory, class gear like goggles and gears. Naming an art style, such as painterly digital fantasy art, also pushes away from the ceramic look.

Can I just write 'adult halfling' and be done?

No. 'Adult' is a single weak signal, and the model weighs it against several correlated child cues that come bundled with small stature: big eyes, smooth skin, round face. Abstract age words lose that vote. Concrete features win it — grey-streaked hair, crow's feet, weathered skin, sideburns. Stack at least three, and the child interpretation becomes visually impossible rather than merely discouraged.

What negative prompts stop child faces in Stable Diffusion?

A short, targeted list works best: child, kid, toddler, chibi, oversized eyes. Add 'teen' if results still skew young. Keep it to a handful of terms — long boilerplate negative walls dilute each term's weight, and modern SDXL-class checkpoints need far less negative prompting than SD 1.5 did. Remember Flux ignores negatives entirely, so there you must state the age positively instead.

How tall are halflings and gnomes supposed to be?

In D&D's rules, halflings stand about three feet tall and gnomes run three to four feet; both count as Small creatures. For portrait art the exact number rarely matters, because most framings never show full height. If you do render full body, pair the height with 'adult head-to-body proportions' so the model doesn't map a three-foot figure onto a toddler template.

Can halflings have beards, or is that only gnomes?

In D&D lore, halfling men rarely grow full beards — long sideburns are their canonical facial hair, and mustaches are rarer still. Gnome men, by contrast, are famous for full beards, often braided, trimmed oddly, or singed from workshop accidents. This is a useful prompt distinction: sideburns instantly read halfling, a wild braided beard instantly reads gnome, and both double as strong age anchors.

Why do my halfling's eyes come out huge like an anime child?

Oversized eyes are part of the child-cue cluster: once the model leans young, it enlarges the irises automatically. Counter it by removing trigger words like cute or adorable, adding crow's feet or a squint to the eye description, and specifying an expression that reads adult, such as wary, shrewd, or tired. In Stable Diffusion, add 'oversized eyes' to the negative prompt directly.

Which art styles make small races look most adult?

Painterly and textured styles help because they render skin detail — oil painting, gritty digital painting, and charcoal all preserve the wrinkles and weathering your anchors request. Smooth, rounded styles fight you: anime, Pixar-style 3D, and soft watercolor all simplify faces toward youth. If you want a stylized look anyway, compensate by doubling the explicit age markers in the prompt.