Necromancer, Warlock, and Witch Portrait Prompts: Making Dark Casters Look Distinct

What visually separates a necromancer from a warlock, witch, and wizard?

Image models can't draw the difference between studied, innate, and borrowed magic — but they can draw the physical evidence each one leaves. Map the fiction to things a camera could see:

• Wizard (studied): the magic lives in gear. A spellbook on a chain, component pouches, ink-stained fingers, reading spectacles, layered scholar's robes. An archmage is this plus rank: finer fabric, a formal staff, authority in the pose.
• Sorcerer (innate): no props at all. The magic lives in the body — faintly luminous eyes, light tracing the veins of one hand, hair drifting as if underwater. A sorcerer prompt with a spellbook in it stops reading as a sorcerer.
• Warlock (borrowed): the magic looks imposed from outside. Glowing pact sigils branded along the forearms, eyes the wrong color for a human, a gifted amulet or rod they didn't make, one consistent otherworldly color repeated in eyes, sigils, and energy.
• Necromancer (death specialist): funerary motifs. Bone charms and clasps, a raven-feather mantle, gaunt pallor, grave dust on black robes, sickly green or violet light.
• Witch (folk practitioner): humble materials and domestic ritual. Rough wool and linen, dried herbs braided into hair, small charms of feather and twine, a candle instead of a glowing staff.

The discipline that makes this work: pick one signature set per character. A figure with a spellbook, pact sigils, bone charms, and a herb bundle reads as noise, and the model will average it into that generic black-robed wizard again.

Which spell-effect phrases actually render (glowing eyes, eldritch swirls, runes)?

Image models render light reliably and "magic" vaguely. "Crackling with arcane energy" gives the model nothing to draw; a colored light source with a position does. The working formula is effect + color + position + what it illuminates:

• Weak: "casting a spell" → Strong: "sickly green light pooling in her raised palm, casting upward shadows across her face"
• Weak: "eldritch energy" → Strong: "ribbons of violet light coiling around his forearm, throwing purple highlights along his jaw"
• Weak: "magical aura" → Strong: "a faint cold glow outlining her silhouette against the dark"

Specific effects, ranked by how consistently they render:

1. Glowing eyes — near-perfect success, but control the intensity: "faintly luminous amber eyes" reads eerie; "eyes blazing with white light" blows out the whole face.
2. Energy wreathing a hand — very reliable when phrased as flame or light: "one hand wreathed in pale green flame".
3. Floating runes — "faintly glowing runes drifting around her raised hand" works; expect elegant gibberish glyphs, which is fine at portrait scale. Runes on skin render better as "luminous sigils traced along his forearms".
4. Smoke and wisps — "thin wisps of dark vapor curling from his fingertips" renders well and adds motion without hiding anything.

Because the spell is now a light source, it should agree with your scene lighting — a green palm-glow in a warm sunlit scene confuses the model. The magical glow lighting setup makes the spell itself the key light, which is the cleanest version of this.

How do you prompt death motifs without tripping content filters?

Necromancer prompts fail moderation more than any other caster because the obvious vocabulary — corpses, blood, rot — is exactly what filters target. Midjourney aims for PG-13 and runs a dynamic filter that blocks gore terms and their synonyms; ChatGPT is stricter still and refuses at the prompt level, before any image is generated, whenever it reads violence into the wording. Both, however, will happily draw skulls and skeletons as objects. The filter isn't reacting to death imagery — it's reacting to violence and viscera.

So write the motifs as ornament, architecture, and antiquity instead of biology:

• corpse, dead body → "ancient skeletal remains", "a figure of yellowed bone in tattered ceremonial robes"
• blood → "deep crimson", "dark red ritual pigment"
• rotting, decayed → "weathered", "crumbling", "centuries-old"
• undead servant → "a silent skeletal attendant"
• graveyard gore → "a moonlit mausoleum", "rows of funerary urns", "carved tomb reliefs"

Ornamental framing also just produces better portraits: "a skull carved from yellowed ivory hanging at her belt" gives the model a concrete prop with material and placement, where "surrounded by corpses" gives it a moderation flag and a compositional mess. Keep the human subject clearly alive and unharmed, put the death imagery in props and setting, and even DALL-E cooperates. If ChatGPT still refuses, our ChatGPT portrait guide covers rephrasing patterns for fantasy content in more depth.

How does magic intensity change the portrait, from subtle to overwhelming?

Magic intensity is a dial, and most prompts leave it cranked to maximum by accident. Three useful bands:

• Subtle: one quiet tell — faintly luminous eyes, a single lit rune on a ring, a wisp of vapor from a closed fist. The image reads as a character who happens to cast. Best for recurring NPCs and anyone whose face you care about.
• Moderate: an active effect contained to one region — a hand wreathed in green flame, sigils glowing through a sleeve, runes orbiting a raised palm. The classic character-sheet look.
• Overwhelming: the environment reacts — full-body aura, hair and robes lifted by unseen wind, dust rising, light flaring off every surface. Reads as a boss reveal or a book cover, not a portrait.

The tradeoff nobody states: effects compete with the face for pixels. At high intensity the model spends its detail budget on light ribbons and drifting debris, and facial features get approximated — the same reason busy armor produces melted faces. Two mitigations: keep active effects at chest height or below ("light pooling in her lowered palm"), or push the magic behind the subject as a rim light so it frames the silhouette instead of overlapping it.

If you build prompts in our generator, the magic-intensity field encodes exactly these bands, so you can render the same warlock at "quiet menace" for a roster portrait and full manifestation for a climactic scene without rewriting the character.

Which lighting and color palettes sell a dark caster?

"Dark fantasy" in a prompt mostly produces murk. What sells the mood is a specific low-key lighting scheme with one controlled accent color:

• [Candlelit](/library/lighting/candlelit) — warm amber pools with deep falloff. The witch's natural habitat: intimate, domestic, slightly conspiratorial.
• [Moonlit](/library/lighting/moonlit) — cold blue-silver wash with hard shadows. Default for necromancers; it makes pale skin look drained rather than airbrushed.
• [Magical glow](/library/lighting/magical-glow) — the spell is the key light. Under-lighting from a palm-held effect is the single most caster-coded lighting choice available: light from below inverts normal facial shadows, which reads as unsettling before the viewer knows why.
• [Dramatic rim light](/library/lighting/dramatic-rim-light) — subject mostly silhouette, one bright edge. Good for warlocks whose patron color can supply the rim.

On palette, restraint does the work. A cold and grim or muted, desaturated base keeps the frame quiet so a single accent — the green of the necromancer's palm-glow, the gold of a warlock's sigils — lands with full force. A rich saturated palette fights the mood; everything glows, so nothing does.

One warning from the drow playbook: models tie skin brightness to scene brightness, so a gaunt pallor or unusual skin tone holds better in these dim schemes than in daylight. State the skin tone explicitly anyway — "pale grey-tinged skin" — so the low light doesn't drift it.

What does a full example prompt look like for each dark caster?

Each prompt below follows the same order — framing, subject and signature, spell effect as light, scene lighting, palette, style — the structure our prompt formula guide covers in depth.

Necromancer (Midjourney; the two parameters keep it painterly and portrait-shaped):

Half-body portrait of a gaunt human necromancer, hollow cheeks and pale grey-tinged skin, layered black robes trimmed with small bone charms, a raven-feather mantle across his shoulders, one lowered hand with sickly green light pooling in the palm and casting faint upward shadows on his face, a cracked leather grimoire chained at his hip, moonlit mausoleum interior behind him, muted desaturated palette with a single green accent, dark fantasy digital painting --ar 2:3 --style raw

Warlock (DALL-E/ChatGPT; plain prose, nothing filter-adjacent):

Three-quarter portrait of a tiefling warlock with dusky violet skin and swept-back horns, thin golden sigils glowing softly along her forearms, faintly luminous amber eyes, a high-collared coat of dark brocade fastened with a star-shaped amulet, wisps of golden light curling from her fingertips, dramatic golden rim light against a deep indigo background, painterly dark fantasy style

Witch (works as-is in Midjourney or DALL-E; for Stable Diffusion, convert to front-loaded comma tags):

Head-and-shoulders portrait of a weathered hedge witch in her sixties, deep-set grey eyes and sharp laugh lines, a rough wool shawl pinned with a carved bone clasp, dried herbs and small feather charms braided into her silver hair, lit by a single candle flame throwing warm amber light up one side of her face, dark cottage interior, earthy natural palette, oil painting style

The prompts transfer across tools because they're descriptive prose rather than parameter tricks — only the tail changes. Midjourney takes --ar 2:3 and optionally --style raw; DALL-E and Leonardo read the prose directly; Stable Diffusion wants the same anchors compressed into tags with the race and signature props first.

Frequently asked questions

Why does my necromancer look like a generic wizard?

Because "necromancer" plus dark robes is visually identical to "wizard" in the training data. The model needs death-specific anchors it can draw: bone charms or a bone clasp, a raven-feather mantle, gaunt pallor, grave dust, a moonlit tomb setting, and sickly green or violet light instead of neutral arcane blue. Add two or three of those and drop the staff, which is the strongest generic-wizard signal in the prompt.

What color should a warlock's eldritch magic be?

Any color works as long as it's singular and repeated. Pick one patron color and use it for the eyes, the sigils, and the energy effect in the same prompt — that repetition is what reads as a pact rather than random VFX. Green and violet read otherworldly, gold reads celestial or fey, deep red reads infernal. Two magic colors in one portrait usually just looks like a rendering error.

Will AI generators refuse to draw skulls and skeletons?

Generally no. Skulls and skeletal figures as objects or ornaments pass moderation in Midjourney and DALL-E without trouble. Refusals come from violence and gore context around them: blood, corpses, rot, dismemberment, or wording that implies a recently dead human. Keep the living subject unharmed, present bones as carved, ancient, or ceremonial, and the same imagery goes through cleanly.

How do I stop a witch portrait from looking like a Halloween costume?

Never write "witch hat" or "broomstick," and don't rely on the word "witch" alone — those pull directly toward costume-shop imagery. Specify age and texture instead: a weathered face in its fifties or sixties, rough wool or linen clothing, dried herbs, twine-and-feather charms, candlelight in a dark cottage. Terms like "hedge witch" or "folk healer" also steer the model toward grounded folk-fantasy rather than October clip art.

What framing works best for a spellcasting portrait?

Half-body or three-quarter framing, because a raised or lowered casting hand needs to fit in the frame with room for the light it throws. Head-and-shoulders close-ups give you better faces and hide hands entirely — worth it when hands keep rendering badly — but you lose the spell effect, so the caster identity has to come from eyes, sigils, and clothing instead.

How do I keep glowing eyes from taking over the face?

Control intensity with the adjective, not luck. "Faintly luminous," "a soft inner glow," or "embers behind the eyes" produce eerie but readable faces. "Blazing," "burning," or "beams of light" make the model treat the eyes as lamps, blowing out the surrounding skin and flattening expression. If the glow still dominates, name a second stronger light source in the scene so the eyes aren't the brightest thing in the frame.

Do these prompts work in Leonardo AI and Flux?

Yes. Both handle descriptive prose well, so the prompts transfer with the Midjourney parameters removed. Flux has no negative prompts, so anything you'd normally exclude has to be stated positively — say "a single green light source" rather than trying to negate stray colors. In Leonardo, pick a dark-fantasy or painterly model and the same anchor structure applies unchanged.

Should I put the class name in the prompt at all?

Keep "necromancer" and "witch" — they carry real visual weight in training data and cost nothing. "Warlock" and "sorcerer" are much weaker signals because general fantasy art doesn't distinguish them, so the D&D-specific meaning barely exists in the model. Include the word if you like, but treat it as a garnish: the sigils, props, and lighting are what actually determine which caster shows up.