Midjourney D&D Prompts That Work in V7: Parameters, Race Anchors, and Copy-Paste Examples

What changed about Midjourney prompting in V7?

Four changes matter for character art.

Prose beats keyword stacking. V7's headline improvement was prompt understanding. "A dwarf blacksmith with a soot-streaked red beard, lit by forge fire" now lands more reliably than dwarf, blacksmith, red beard, soot, forge, dramatic lighting. Write sentences describing what a viewer would see; drop filler tokens like 8k, ultra detailed, and masterpiece — they steer nothing in V7.

Personalization is on by default. V7 was the first model to enable profile personalization automatically (you unlock it by rating image pairs for a few minutes). It nudges every generation toward your historical taste. Useful for a house style, but if your portraits keep drifting in one aesthetic direction, personalization is a likely culprit — you can toggle it off per prompt or in settings.

Draft mode exists. Draft renders at roughly ten times the speed for half the GPU cost. Iterate on wording in draft, then enhance the keeper to full quality. For portrait work, where you often burn ten generations finding the right face, this halves the real cost of experimenting.

--cref is gone. Character Reference was a V6 feature; V7 replaced it with Omni Reference (--oref), covered below.

One more thing: V8.1 became Midjourney's default model in June 2026. Everything here about prompt structure carries forward, and --oref, --sref, and raw mode remain part of the current toolset — but if you want the exact V7 behavior this guide describes, pin it with --v 7 at the end of your prompt.

How should you structure a D&D portrait prompt in Midjourney?

Order matters because Midjourney weights early tokens more heavily. A reliable sequence for character portraits:

1. Framing and medium — "Bust portrait of...", "Half-body painting of...". A bust portrait crop gives the model more pixels per face and conveniently hides hands.
2. Race, class, and gender — "a female tiefling warlock".
3. Permanent physical anchors — skin tone, eyes, hair, horns, scars. Two or three specifics, phrased concretely.
4. Gear and materials — one or two items with a material each: "dented steel plate", "oiled brown leather".
5. Lighting and environment — a named light source does more than any adjective; dramatic rim light alone can carry a portrait.
6. Art style — "painterly digital art", "dark fantasy oil painting".
7. Parameters — always last.

Put together:

Bust portrait of a middle-aged human knight, sun-weathered brown skin, cropped grey-flecked beard, a pale scar through one eyebrow, dented steel plate armor with a faded blue tabard, lit by dramatic rim light against a stormy sky, painterly digital art --ar 3:4 --style raw --stylize 100 --v 7

Aim for 40-80 words before the parameters. Shorter prompts hand too many decisions to the model; much longer ones bury your anchors under detail the model will start ignoring.

Which parameters matter for character portraits?

--ar (aspect ratio). Default is square. For portraits use --ar 3:4 or --ar 2:3; --ar 4:5 if you want minimal crop from square. Stay 1:1 only when you're making VTT tokens.

--style raw. V7's default aesthetic layer adds its own beautifying interpretation. --style raw strips most of that, sticking closer to your literal wording — which is exactly what you want when you've specified a scar, a skin tone, and a horn shape and need them respected. Raw output looks less "finished" with vague prompts, so it rewards the detailed structure above.

--stylize. Runs 0-1000, default 100. Low values follow the prompt literally; high values hand Midjourney artistic license. For character art, license means drift: at --s 500+ faces converge toward Midjourney's idealized default and your specified traits fade. Working bands: --s 0-50 for strict trait accuracy (slightly plainer rendering), --s 100-150 as the everyday sweet spot, --s 200-250 only when you want painterly flair and can tolerate some trait loss.

--no. V7's negative prompt: --no beard, helmet, hat removes the things Midjourney loves adding to fantasy portraits. It also helps token prep — --no shadows, busy background with a stated solid backdrop makes background removal cleaner.

Draft mode. Not a parameter but a mode toggle on the website: half cost, ~10x speed. Do your face-hunting there, then enhance.

A typical portrait tail reads: --ar 3:4 --style raw --stylize 100 --v 7.

What happened to --cref, and how does --oref replace it?

Character Reference (--cref) was V6's consistency tool, and half the D&D prompt guides online still recommend it. It doesn't work in V7. Its replacement, Omni Reference, is broader: --oref <image URL> (or drag an image into the Omni Reference slot on the website) tells V7 to carry a subject — character, creature, or object — from your reference image into the new generation.

Strength is set with --ow, running 0-1000 with a default of 100. Midjourney's own guidance: low values around 25 suit style transfers where only rough identity should survive; higher values around 400 preserve specific details like a face; and very high omni-weight competes badly with high --stylize, so don't max both. Practical band for "same character, new scene": --ow 200-400.

Limits worth knowing: one reference image per prompt (though it can contain more than one subject), it costs roughly double the GPU time of a standard V7 job, and it isn't available in draft mode.

The workflow for a recurring character: generate a clean, front-lit, neutral-background portrait first, save it, then feed it back with --oref while re-stating your character's anchors in the text. The reference holds the face; the repeated wording holds everything the reference can't see. For the full cross-tool playbook — including Leonardo and Stable Diffusion equivalents — see the consistent AI characters guide.

Can you still weight prompt parts with :: in V7?

No. Multi-prompts and :: weights — the old trick of writing tiefling::2 armor::1 to emphasize the race over the gear — are listed in Midjourney's documentation as incompatible with V7. Guides that tell you to weight your D&D prompts this way are describing V6 behavior.

What replaces weighting in practice:

• Position. Early tokens carry more influence. If the race keeps getting lost, move its anchors to the front of the prompt, immediately after the framing phrase.
• Prose emphasis. V7 reads descriptive weight in language: "massive curling horns dominating the silhouette" versus "small horns half-hidden in her hair" is the new ::2 versus ::0.5.
• --no for negatives. The old negative-weight trick (green::-0.5) maps to --no green.
• Reference weights. The numeric dials that still exist live on references, not prompt text: --sw scales style reference strength and --ow scales omni reference strength, both 0-1000.
• Repetition. Restating a critical trait once in different words ("obsidian-black skin... her jet-dark features") measurably improves adherence when a single mention keeps failing.

If you're on V6 deliberately for multi-prompt support, the old syntax still works there — but for portrait work, V7's better prompt comprehension plus these levers covers everything :: used to do.

How do you lock one art style across a whole campaign with --sref?

Style Reference solves a different problem from Omni Reference: not "same character" but "same look." --sref accepts an image URL or a numeric style code and pushes every generation toward that aesthetic — which is how you make twelve NPC portraits generated weeks apart hang together like one commissioned set.

Three ways to use it:

• An image you like: --sref <URL> borrows its palette, rendering, and mood.
• A style code: --sref 1234567890 applies a fixed, shareable aesthetic. Codes are deterministic — write yours down and every portrait for the campaign inherits the same look.
• Random exploration: --sref random rolls a new code each run and reveals which code it used, so you can keep the winners.

Strength is --sw, 0-1000 with a default of 100. In V7, --sw hits harder with numeric codes than with image URLs, so start codes at --sw 50-100 before pushing higher. V7 also versioned the style system itself: --sv 6 is V7's default and --sv 4 selects the legacy pre-June-2025 code model — random codes only work with those two, which matters if you're reusing old codes that suddenly render differently.

A campaign recipe that works: pick one code, one lighting family, and one palette, and vary only the character description between prompts. It's the same batching logic as a party portrait set — every variable you freeze is one less way for the set to look mismatched.

What do working Midjourney prompts look like for D&D races?

Image models drift toward human faces, so non-human races need their anchors front-loaded — right after the framing phrase, before any gear. Name the subtype where one exists ("silver dragonborn", "wood elf"): it sets skin, scale, or coloring in one token. Five copy-paste examples:

Bust portrait of a wood elf ranger, angular sun-browned face with long pointed ears, moss-green eyes, ash-blond hair braided back, weathered leather armor under a forest-green hooded cloak, dappled woodland light, painterly digital art --ar 3:4 --style raw --stylize 100 --v 7

Head-and-shoulders portrait of a drow priestess, obsidian-black skin with a cool violet undertone, long white hair, crimson eyes, silver spider-motif jewelry over dark silk vestments, cold magical glow in a cavern, dark fantasy oil painting --ar 3:4 --style raw --stylize 150 --v 7

Half-body portrait of a tiefling warlock, deep crimson skin, thick ram-curl horns sweeping back from her brow, solid gold eyes without pupils, black robes stitched with faintly glowing violet sigils, candlelit study, dramatic digital painting --ar 2:3 --style raw --stylize 150 --v 7

Bust portrait of a silver dragonborn paladin, prominent reptilian snout, full facial scales in polished silver-grey, slit-pupil amber eyes, swept-back horn crest, engraved plate armor with a sunburst emblem, warm temple light, heroic fantasy illustration --ar 3:4 --style raw --stylize 100 --v 7

Three-quarter portrait of a female half-orc barbarian, olive-green skin, small lower tusks over a set jaw, black hair shaved at the sides, fur-trimmed hide armor, a rope-wrapped greataxe across her back, overcast tundra light, gritty painterly style --ar 2:3 --style raw --stylize 100 --v 7

The hardest races have dedicated deep dives: drow skin tones, tiefling horns and coloring, and dragonborn anatomy each carry ready-made anchor vocabulary you can drop into this structure.

Frequently asked questions

What aspect ratio is best for D&D character portraits in Midjourney?

Use --ar 3:4 or --ar 2:3 for character portraits; both give vertical space for head and torso without wasting canvas. Use --ar 4:5 if you want something closer to square, and plain 1:1 only when the image is destined to become a VTT token, since Roll20 and Foundry tokens are square.

Is Midjourney free for making D&D portraits?

No. Midjourney has no free tier or trial; the cheapest plan is Basic at $10 per month, which includes about 3.3 hours of fast GPU time. To stretch a subscription, iterate in draft mode, which renders at half the GPU cost, and only enhance the images you keep. If you need free generation entirely, a local Stable Diffusion install is the usual alternative.

Do these prompts still work in Midjourney V8?

Largely yes. V8.1 became Midjourney's default model in June 2026, and the fundamentals carry over: descriptive prose structure, front-loaded race anchors, --ar, raw mode, style references, and omni references all remain part of the current workflow. If you want the exact V7 behavior described in a guide, add --v 7 to the end of the prompt to pin the model version.

Why do all my Midjourney characters have the same face?

Two causes stack: high stylize values pull faces toward Midjourney's idealized default, and V7's on-by-default personalization nudges everything toward your rating history. Drop to --stylize 50-100, consider --style raw, and add differentiating anchors the model must honor: specific age, a crooked nose, asymmetry, a scar, a distinct ethnicity, or an unglamorous expression. Generic descriptions get the generic face.

Can Midjourney generate a transparent background for tokens?

Midjourney doesn't output true transparency. The workaround is prompting for a solid, flat-color background that contrasts with the character, adding --no shadows to keep the edge clean, and then removing the background in a free tool afterward. A centered bust or head-and-shoulders crop converts most cleanly into a circular token frame.

Does Midjourney block weapons or blood in fantasy prompts?

Weapons, armor, and combat poses in a fantasy context generally pass moderation without trouble. What gets blocked is gore: words like blood, severed, or graphic wound descriptions can stop a job. If a character prompt is refused, cut the gore vocabulary and imply danger through posture, lighting, and battle-worn gear instead — a notched blade and dented pauldron read as violent history without tripping the filter.

How long should a Midjourney character prompt be?

Aim for roughly 40-80 words of description before the parameters. Under that, the model fills gaps with defaults and you lose control of the face. Far over it, later details get diluted and the anchors that define your character carry less weight. Spend the budget on two or three permanent physical traits, one or two gear items with materials, a named light source, and an art style.