Dragonborn AI Art Prompts: How to Stop Getting a Human in a Costume

Why does AI keep giving your dragonborn a human face?

Image models learn what "dragonborn" looks like from their training data, and that data is a mess: convention cosplay photos (humans in prosthetics), painted miniatures, and fan art that ranges from lizard-headed to nearly human. Faced with contradictory examples, the model averages — and the average of "dragon" and "person" is a person with scaly makeup.

The word itself works against you. "Dragonborn" decomposes into dragon and born, so the model is constantly tugged toward its two strong neighbors: full dragons (millions of clean training images) and humans (the default subject of every portrait it has ever seen). Whichever direction the rest of your prompt leans, the model follows. Write "handsome dragonborn with a braided beard" and you've voted for the human. Write "dragonborn breathing fire over a hoard of gold" and you've voted for the dragon.

That's why this race punishes vague prompts harder than tieflings or half-orcs, which at least start from a humanoid face. A dragonborn done right has no human facial features at all — no nose bridge, no lips, no eyebrows, no hair. The model won't remove those on its own. You have to name every replacement explicitly: what sits where the nose would be, what covers the skin, what grows where hair would grow.

Which race anchors must come first in the prompt?

Models weight early tokens more heavily than late ones, so the anatomy goes first — before class, before gear, before mood. Five anchors do the work:

• Prominent reptilian snout — the single highest-value phrase. "Snout" is what separates a dragon head from a scaled human face. "Elongated muzzle" also works.
• Full facial scales — and make the coverage explicit: "scales covering the entire face and neck, no visible skin." Without "entire," you get scale patches on human cheeks.
• Slit-pupil eyes with a named color — amber, molten gold, pale ice-blue. Round human eyes instantly re-humanize the face.
• Horns with a direction — "a pair of backswept horns" or "short paired brow horns." Direction matters; bare "horns" often lands as a headband-like nub.
• No hair — say it outright. Hair is the strongest single human tell, and models love adding it.

Stacked into an opening clause, it looks like this:

A red dragonborn warrior with a prominent reptilian snout, full facial scales covering the entire face and neck, slit-pupil amber eyes, and a pair of backswept obsidian horns — no hair, no visible human skin.

That's the skeleton every prompt in this guide builds on. The dragonborn library page keeps the full anchor list, and the prompt generator front-loads these automatically when you pick dragonborn as the race — you fill in subtype, class, and mood, and it handles the ordering.

How does naming a subtype like "red dragonborn" change results?

"Dragonborn" alone leaves scale color to chance, and chance usually picks muddy brown-grey. Naming a subtype hands the model a concrete color anchor it can actually paint — and because D&D art of "red dragonborn" and "silver dragonborn" is reasonably consistent, the phrase pulls the whole image toward better reference material. The Civitai dragonborn LoRAs document the same behavior: prompting the subtype color is how their users control the output.

The standard subtypes give you ten reliable colors:

• Chromatic: red, blue, green, black, white
• Metallic: gold, silver, bronze, copper, brass

Two upgrades make the color read as a creature instead of a paint job. First, add a second tone for the underbelly and throat: "deep crimson scales with a pale ochre underbelly at the throat" looks biological; a single flat color looks like a rubber suit. Second, pair the eyes deliberately — molten gold eyes against crimson scales, ice-blue against white, acid-green against black.

Metallics need one extra word about finish. "Silver dragonborn" can come out as grey skin; "burnished silver scales that catch the light" forces the metallic sheen. For gold and brass, a golden, warm palette in the style section of the prompt reinforces the read; for white and silver, something cold and grim keeps the scales from washing out.

How do you get horns, crests, and the tail right?

Horns

Give every horn a shape and a direction. Phrases that render reliably: "a pair of long backswept horns," "ram-curled horns hugging the jaw," "a row of short spikes along the brow ridge." Naming a material or finish — polished obsidian, weathered ivory — keeps them from blending into the scales.

Crests and frills

Where a human has hair, a dragonborn has structure: "a bony crest sweeping back from the brow," "a frilled ridge running from skull to nape," "small jaw spines." Including one of these does double duty — it fills the visual hole where hair would be, which removes the model's temptation to add hair on its own.

The tail

Whether dragonborn have tails is a genuine canon split: the 2014 rules describe them as tailless, but later official books added tailed variants — Wildemount's draconblood, for one — and fan art goes both ways. The model has seen both, so it flips a coin unless you decide. In a half-body or full-body shot, write "a thick draconic tail" or "tailless" explicitly. In a head-and-shoulders close-up you can ignore the question entirely — which is one more reason tight framing is the easy mode for this race.

Wings

Standard dragonborn don't have wings, but "dragon" drags them in constantly. Add "wingless" in the prompt, or suppress them with the negative side of your tool: --no wings in Midjourney, "wings" in a Stable Diffusion negative prompt. DALL-E has no negative field, so the in-prompt "wingless" is your only lever there.

Why should the rest of the prompt stay minimal?

Every clause in a prompt competes for the model's attention. When you follow five race anchors with six clauses of filigreed armor, heraldry, background architecture, and weather, the anchors get diluted — and diluted anchors are exactly how the human face creeps back in. Worse, elaborate gear descriptions are votes for the wrong subject: the model has seen "ornate gothic plate armor, noble bearing, flowing cloak" ten thousand times, and in nearly every example the wearer was human.

Practical budget: one clause of gear, one of lighting, one of style. That's enough to say who this character is.

• Gear: pick one or two pieces and let material do the talking — a scratched bronze pauldron, crossed leather straps, a wool traveling cloak. Straps and open collars beat gorgets and full helms because they leave the neck scales visible, and visible scales below the jaw are what sell "this whole body is draconic."
• Lighting: warm directional light makes scale texture pop. Firelight is the classic pairing for chromatic reds and golds; a dramatic rim light traces the snout and horn silhouette so the profile reads unmistakably non-human even in a dark image.
• Style: painterly digital painting handles scale texture more gracefully than photorealism, which tends to render scales as either plastic or disease.

If a detail doesn't help the viewer identify the character in two seconds, it's spending attention you need for the snout.

What do working dragonborn prompts look like in Midjourney and DALL-E?

Both tools take descriptive prose, so the anchor-first structure carries over directly — the differences are parameters and phrasing tolerance.

Midjourney gives you --no for suppressing the failure modes and --style raw for less beautification (default aesthetics push faces back toward pretty and human). A working v7 prompt:

Head-and-shoulders portrait of a red dragonborn paladin, prominent reptilian snout, full facial scales in deep crimson with a pale ochre underbelly at the throat, slit-pupil molten-gold eyes, a pair of long backswept dark horns, wingless. A scratched bronze pauldron over a leather strap. Warm firelight against a dark background, painterly digital fantasy art. --ar 2:3 --style raw --no wings, hair, beard, human skin

Once a generation nails the head, feed it back with --oref (v7's omni-reference, which replaced --cref) to hold the face across new poses. Full parameter details are in our Midjourney D&D prompt guide.

DALL-E / ChatGPT has no negative prompt and no style flags, so state the non-human anatomy as an instruction and repeat the coverage claim:

A bust portrait of a silver dragonborn sorcerer. Non-human anatomy: a full reptilian head with an elongated snout, burnished silver scales covering the entire face and neck with no human skin visible, pale ice-blue slit-pupil eyes, a bony crest sweeping back from the brow instead of hair, wingless. She wears a simple slate-blue robe with a silver clasp. Soft diffused light, muted cool palette, painterly digital art.

ChatGPT's advantage is conversation: if the first render is too human, reply "make the snout longer and remove all hair — fully reptilian head" and it revises. More on that workflow in the ChatGPT portrait guide.

Should Stable Diffusion users just use a LoRA?

Mostly, yes. Stock Stable Diffusion checkpoints know "dragonborn" even less than Midjourney does, and the community answer is downloadable: Civitai hosts several dragonborn LoRAs — concept LoRAs for SD 1.5-era checkpoints like DreamShaper and newer SDXL versions — trained specifically to produce the reptilian head. Their trigger token is typically just dragonborn, and subtype colors (red dragonborn, silver dragonborn) steer scale color exactly as they do in prose prompts. Check each model card for its base checkpoint and recommended weight before mixing it with your usual portrait checkpoint.

With a LoRA loaded, a tag-format prompt looks like this:

dragonborn, red dragonborn, reptilian snout, scaled face, slit pupils, amber eyes, backswept horns, wingless, bust portrait, leather armor, bronze pauldron, firelight, dark background, fantasy, detailed digital painting
Negative prompt: human, human face, skin, hair, beard, wings, full dragon, deformed

Two habits matter more here than anywhere else. Front-load: SD's attention drops off across the prompt, so the race tags stay in the first handful of tokens. And keep the negative prompt doing the species work — human face, hair, beard in the negative is often worth more than another positive tag.

No LoRA available (a hosted service, say)? Fall back to the prose anchors from earlier and accept a lower hit rate. Checkpoint recommendations, token limits, and resolution settings are covered in the Stable Diffusion D&D portrait guide.

Frequently asked questions

Why does my dragonborn come out as a full four-legged dragon?

Too many dragon-associated words and no humanoid anchors. Add "humanoid posture, standing upright, wearing armor" or name the class and gear early — clothing is a strong humanoid signal. Cut scene words like "hoard," "lair," or "breathing fire," which pull toward creature illustrations of true dragons rather than character portraits.

Do dragonborn have hair or beards?

No. Dragonborn are fully scaled with no hair anywhere, which is exactly why AI keeps adding it — hair is the default for every portrait the model has seen. Write "no hair" explicitly and describe what replaces it, such as a bony crest or frilled ridge sweeping back from the brow. If a beard-like shape appears, add "beard" to your negative prompt or regenerate.

How do I make a female dragonborn actually read as female?

Skip hair and soft human features — they just re-humanize the face. Instead use build and silhouette: a slimmer jaw and snout, smaller or more elegant horns, lighter frame, and posture or costume cues like a fitted robe or jewelry. Stating "female dragonborn" plus one or two of these cues works in most generators without sacrificing the reptilian head.

Can I show a breath weapon in a portrait prompt?

Yes, but keep it small or the image tips into action-scene territory. Phrases like "faint ember glow between the jaws," "wisps of smoke curling from the nostrils," or "crackling blue light along the teeth" add the element without opening the mouth wide, which models handle badly. Match the effect to the subtype: fire for red, lightning for blue, frost for white.

What about gem dragonborn like amethyst or sapphire?

Models don't know Fizban's gem subtypes by name, so describe the look instead: "crystalline amethyst scales with a faceted shimmer, translucent purple crest." Treat it as a custom color scheme layered onto the same anchors — snout, full scales, slit pupils, horns. Expect more retries than red or silver, since there's little training art to lean on.

Do alternative terms like "lizardfolk" or "argonian" work better?

Sometimes, as fallbacks. "Anthropomorphic dragon humanoid" often outperforms "dragonborn" on models that barely know the D&D term. "Argonian" renders reliably thanks to Skyrim fan art but drags Elder Scrolls styling with it. "Lizardfolk" reads more feral and swamp-toned. Try them when the anchors alone keep failing, and keep the anatomical phrases in the prompt either way.

What eye colors look best on a dragonborn portrait?

Pick a color that contrasts the scales so the eyes read at portrait size: molten gold or amber against red and black scales, ice blue against white or silver, acid green against bronze and brass. Always specify slit pupils — round pupils are one of the strongest human tells in an otherwise reptilian face. Solid-color eyes without visible pupils also work for a more alien look.