Drow Portrait Prompts: Getting the Skin Right Without Saying "Dark Elf"

Why does "dark elf" confuse image models?

Because "dark" is doing three jobs at once, and the model has to guess which one you meant:

• Dark as lighting. The most common misread: a fair-skinned elf in a dim, moody scene. The model satisfied the word by turning down the lights.
• Dark as wardrobe and vibe. Black leather, black hair, brooding expression — a goth elf with pale skin.
• Dark as somebody else's dark elf. Training data is full of competing concepts that all answer to the same name: Elder Scrolls dunmer (ash-grey skin, red eyes), Warcraft night elves (violet skin, glowing eyes, huge ears), and a long tradition of anime dark elves with tanned human skin and white hair. Whichever one dominates the model's training wins, and it's rarely the D&D drow.

D&D canon is specific: drow skin runs from dark grey through jet black, like polished obsidian, often with a blue or violet cast — paired with stark-white hair and, most commonly, red eyes. None of that is recoverable from the phrase "dark elf".

The fix is the same discipline that works for every hard-to-render race: stop relying on the label and describe what a camera would see, starting with the skin. Use "drow" as the race word — it's a more precise token than "dark elf" — but treat it as a hint, not a spec. The prompt generator applies this automatically when you pick drow: the composed prompt leads with explicit pigment, hair, and eye anchors instead of trusting the name.

Which exact skin tones work for drow, and which words fail?

The phrases that render reliably all follow one pattern: a concrete, non-human pigment bound to the word "skin", placed in the first clause of the prompt.

The working palette, roughly from canon-classic to stylized:

• "obsidian-black skin with a faint violet sheen" — the closest to book canon. The sheen phrase matters: it tells the renderer the darkness is a reflective surface quality, not a shadow.
• "jet-black skin with a cool blue undertone" — same idea, cooler cast.
• "deep plum-purple skin" or "dusky violet skin" — the most popular gallery rendering, slightly stylized but unmistakably non-human.
• "charcoal blue-grey skin" or "slate-grey skin with a purple cast" — reads more dunmer-adjacent but holds well under varied lighting.

What fails, and why:

• "dark skin" — resolves to a human complexion or plain underexposure. Never non-human.
• "black skin" alone — swings between a human tone and a literal void-black silhouette with no facial shading.
• "shadowy" / "dusky" alone — pure lighting words; the model dims the scene and moves on.
• Letting the color float. "A purple drow" often paints the clothes or the background purple and leaves the face beige. Bind it: "a drow with deep plum-purple skin."

This is the same anchor-word principle that governs human complexions — vague brightness adjectives fail, named pigments hold — covered in full in the skin tone prompt guide.

How do you anchor hair, eyes, and Underdark context?

Skin alone gets you a dark-skinned elf. What makes a drow read instantly is contrast: near-white hair against near-black skin. That two-tone signature is the strongest identity signal in the whole prompt, so spend words on it.

• Hair: "long stark-white hair" is the default and the most reliable. "Silver-white" and "moonlight-white" also hold. Canon allows pale yellow and rare silver or copper, but white is what models have seen, so white is what renders cleanly. Give it behavior — "pulled back from pointed ears", "loose over one shoulder" — so the hairline renders instead of helmet-ing.
• Eyes: bright red is the canonical common color and it renders well: "piercing crimson eyes" or "blood-red eyes". Pale lilac and silver are canon alternatives that read more regal. Skip "glowing red eyes" unless you want the glow to become a light source on the cheekbones.
• Features: "sharp angular features, pointed ears, white eyebrows" — the white eyebrows detail is small but does outsized work selling the pigment as skin rather than shadow.

Then place the character somewhere only a drow would be. Underdark context both reinforces the race and hands you lighting that flatters dark skin: "a vast lightless cavern", "the cold blue glow of bioluminescent fungi", "faint purple faerzress light", "spider-silk banners", "carved black stone". A magical glow or moonlit treatment gives you cool rim light that separates obsidian skin from a dark background instead of swallowing it.

Here's the whole identity block assembled — usable as the front half of any drow prompt:

A drow woman with obsidian-black skin with a faint violet sheen, long stark-white hair pulled back from pointed ears, piercing crimson eyes and white eyebrows, sharp angular features, lit by the cold blue glow of bioluminescent fungi in a vast Underdark cavern

How do you stop the model lightening the skin in bright scenes?

Skin-lightening is not your imagination — bias toward lighter skin is a documented, measured behavior of image models, and it hits drow twice. First, the model's prior pulls every face toward lighter tones. Second, bright light gives it an excuse: it reinterprets your pigment as a shadow that the light is now lifting.

Five counters, in order of effectiveness:

1. Choose lighting that agrees with the character. Cool, directional, low-ambient light — moonlight, candlelight at a distance, fungal glow — flatters obsidian skin. Avoid "golden hour", "bright sunlight", and "high key" phrasing; warm flooding light is where drift starts. If you need brightness, ask for contrast instead: a cold, grim palette keeps the exposure honest.
2. Describe the light as falling on the dark skin. "Cold silver rim light tracing her obsidian skin" makes the pigment part of the lighting instruction. The model can't lift the shadow without contradicting the sentence.
3. Repeat the pigment, bound to a feature. Once up front, once later: "...obsidian-black skin... her jet-black hands resting on the staff." Redundancy costs a few tokens and buys a whole batch of consistency.
4. Use negatives where you have them. In Stable Diffusion: pale skin, fair skin, tan skin, human skin tone. Midjourney: --no pale skin.
5. Keep stylization moderate in Midjourney. High --stylize values hand aesthetic control back to the model, and the model's aesthetic is light-skinned. Stay at or below the default 100 when the pigment is drifting.

If a surface scene in daylight is the whole point — a drow ranger above ground — write the light as overcast or dusk and keep counters 2 and 3 in place.

How do drow priestess, ranger, and assassin builds change the prompt?

The identity block from above stays fixed; the class layer changes clothing, props, mood, and scene. Race anchors first, class dressing second — same ordering rule as every other fantasy race.

Priestess of Lolth

The iconic drow authority figure. Visual signature: spider motifs and ceremonial wealth. "Black and silver vestments embroidered with spiderweb patterns", "a spider-shaped silver diadem", "a gown of dark spider-silk", imperious expression, chin slightly raised. Scene: a temple of carved black stone lit by braziers — warm firelight against cool skin reads dramatic, but keep it rim-light rather than flood (see the previous section). A priest framing with three-quarter pose sells the rank.

Drow ranger

The archetype one famous exile made popular: a drow living on the surface. Gear reads practical and weathered — "a forest-green hooded cloak over dark leather armor", "twin curved scimitars at the hips", "light snowfall". This is the build most likely to trigger skin-lightening because the scene is outdoors, so specify overcast or moonlit conditions and keep the pigment repeated. Build it on a hunter base and the wilderness props come along naturally.

Drow assassin

Lean into what the model already wants to do with drow — shadow and menace — but keep the face lit. "A dark hood framing her face, only her white hair and crimson eyes catching the light", "twin daggers", "muted charcoal leathers". An assassin works best in a tight bust portrait where the hood, hair, and eyes make the whole composition.

What do full drow prompts look like per generator?

The trait content is identical everywhere; only the format changes.

Midjourney

Descriptive prose, identity block first, parameters last. --style raw keeps the pigment literal instead of prettified; --ar 2:3 for portrait framing:

Fantasy character portrait of a drow priestess with obsidian-black skin with a faint violet sheen, long stark-white hair swept back from pointed ears, piercing crimson eyes and white eyebrows, black and silver vestments embroidered with spiderweb patterns, a spider-shaped silver diadem, cold blue fungal glow with warm brazier rim light, carved black stone temple in the background, digital painting, three-quarter portrait --ar 2:3 --style raw --no pale skin

For a recurring character, save your best render and feed it back with --oref (Omni Reference — it replaced v6's --cref in v7) alongside the same written spec; the combination holds the two-tone signature across a session. More parameter detail in the Midjourney D&D prompt guide.

DALL-E / ChatGPT

Plain sentences, and state explicitly that the skin is not a lighting effect:

A painted fantasy portrait of a drow ranger with deep plum-purple skin — the color is his natural skin tone, not shadow — long white hair tied back, red eyes, and sharp angular features. He wears a weathered forest-green hooded cloak over dark leather armor with twin curved scimitars at his hips. Overcast moonlit forest, cool silver light tracing his skin, oil painting style, half-body framing.

Stable Diffusion

Comma-separated tags, most important first, with negatives doing real work:

drow woman, obsidian black skin, violet sheen, long white hair, red eyes, white eyebrows, pointed ears, dark hooded leathers, twin daggers, underdark cavern, bioluminescent blue glow, bust portrait, fantasy digital painting, detailed face
Negative prompt: pale skin, fair skin, tan skin, human skin tone, blurry, extra fingers

Stock checkpoints are thin on the drow concept, which is why several dark-elf LoRAs exist on Civitai — reach for one if the tags alone keep drifting. Or skip the assembly: pick drow in the generator, set character type, lighting, and framing, and paste the composed prompt into whichever tool you use.

Frequently asked questions

Why does my drow render as a regular human with white hair?

The skin anchor is missing, vague, or buried. If the pigment phrase isn't concrete and early — "obsidian-black skin", "deep plum-purple skin" — the model defaults to a light-skinned face and keeps only the easy traits like white hair and pointed ears. Move the skin phrase into the first clause, bind it to the word "skin", and repeat it once later in the prompt attached to a visible body part.

What is the difference between prompting a drow and an Elder Scrolls dunmer?

Dunmer are ash-grey with red eyes and often dark hair; drow are obsidian-black to deep purple with stark-white hair. The word "dunmer" pulls Elder Scrolls concept art, angular Morrowind styling, and grey tones, while "drow" leans toward D&D illustration. If your drow keeps coming out grey and gaunt, the model may be blending in dunmer training data — strengthen the purple-black pigment phrase and the white hair anchor.

What eye colors do drow have besides red?

Bright red is the most common in Forgotten Realms lore, but pale shades that appear nearly white — lilac, blue, pink, and silver — are also frequent, and rarer drow have green, amber, or violet eyes. In prompts, red renders most reliably and reads most recognizably. Pale lilac or silver eyes photograph beautifully in close-ups but occasionally drift toward blind-looking white, so add "bright" or name the color twice if that happens.

How do you prompt a male drow portrait?

Same anchors — obsidian or purple skin, white hair, red eyes, angular features — but state "male drow" or "a drow man" in the first few words. Gallery training data skews heavily female for drow, so an unspecified gender usually returns a woman. Adding masculine grooming details helps hold it: "white hair cropped short", "sharp jawline", "a scarred brow". Expect to reroll slightly more often than for female drow.

Do I need a LoRA to generate drow in Stable Diffusion?

No, but it reduces rerolls on stock checkpoints. Base models saw relatively little drow imagery in training, which is why dark-elf LoRAs exist on Civitai. Without one, front-load explicit tags — obsidian black skin, white hair, red eyes, pointed ears — and use negatives like "pale skin, fair skin, human skin tone". With a LoRA, the same tags become far more consistent, especially for the purple-black skin range.

What is a good negative prompt for drow portraits?

Target the two failure directions: skin drifting light and the character drifting human. A compact set that covers both: "pale skin, fair skin, tan skin, human skin tone, rosy cheeks, sunburn". Add "grey skin" if you're getting dunmer-style ash tones instead of purple-black, and the usual quality negatives like "blurry, extra fingers" as needed. Keep the list short — long boilerplate negatives dilute the ones doing real work.

Can a drow have hair that isn't white?

Lore allows stark white, pale yellow, and rarely silver or copper. In prompts, white and silver-white are by far the most reliable and are also what makes the race read instantly, because the white-on-black contrast is the drow's strongest visual signature. Pale yellow renders fine but weakens recognition; anything darker usually costs you the drow look entirely and pushes the render toward a generic elf.

How do I keep the same drow consistent across multiple images?

Lock a written trait spec — exact pigment phrase, hair description, eye color, one unique detail like a scar or a silver ear cuff — and reuse the identical wording in every prompt. Then add your tool's reference feature on top: Omni Reference (--oref) in Midjourney v7, character reference in Leonardo, or a LoRA in Stable Diffusion for long-running characters. The repeated wording matters as much as the reference image.