D&D AI Art Generator: Create Fantasy Characters and Maps

Why AI Art Changed Prep Night for Dungeon Masters

Every Dungeon Master eventually hits the same wall. You can describe a rain-slicked harbor district, a scarred half-orc warlord, or a cathedral carved into a glacier, and your players will nod politely — then ask which square they are standing on. Description is powerful, but it is not visual. A single image does more work in five seconds than three paragraphs of narration.

For years the answer was commissioning art. That worked beautifully for people with a budget and a month of patience. Everyone else made do with whatever they could find online, borrowing pieces in mismatched styles until the campaign looked like a collage stitched together from five different books.

Image generation tools changed the math. A DM can now type a sentence, wait thirty seconds, and get four candidate portraits for a villain. Nobody has to be an illustrator. Nobody has to wait three weeks. The real shift is not that art became cheap — it is that art became iterative. You can generate twenty versions of a face, discard nineteen, and keep the one where the eyes actually match the backstory.

That loop of generate, judge, refine is the entire skill. This guide walks through where these tools genuinely help, where they quietly fail, how to pick one, and how to build a workflow that keeps a long campaign looking like one coherent world instead of forty unrelated pictures.

What AI Art Generators Do Well — and Where They Break

Before choosing anything, get honest about what the technology is good at. Expecting a text-to-image model to do everything leads to disappointment; expecting it to do three things well leads to a much better campaign.

Where generation genuinely shines

Mood and atmosphere. Fog rolling through a graveyard, torchlight on wet stone, a sky full of unfamiliar constellations. Models are excellent at vibe.
Faces and costume concepts. Rich material detail — embossed plate, layered robes, ritual scars, mismatched jewelry — arrives fast and looks convincing at thumbnail size.
Volume. A cast of forty NPCs is suddenly practical. You can give every shopkeeper and guild contact a face without spending a year.
Iteration speed. Changing a character's hair, age, or armor color takes one edit, not a redraw.

Where generation quietly fails

Hands and small props. Fingers merge, daggers grow extra hilts, a lute turns into a plank with strings.
Readable text. Signs, scrolls, and book titles are usually gibberish unless the tool has a dedicated typography strength. Even then, treat text props as a layout job.
True top-down tactical maps. Models produce gorgeous isometric scenes and terrible grid geometry. Walls wander, doors lead nowhere, and nothing lands on a clean five-foot square.
Cross-image continuity. The same character rendered twice will drift in face shape, hair, and clothing unless you actively manage it.
Exact geography. If the river must run east to west past the keep, no prompt will respect that reliably. Draw the map first, then decorate it.

A useful rule: use generation for flavor, faces, and atmosphere. Use a mapping tool or plain drawing software for anything players will measure, count, or stand on.

Choosing the Right Tool for Your Table

There is no single best generator, only the best fit for how you run games. Score any candidate against this checklist before committing your prep time to it.

Style range. Does it handle painterly fantasy, ink sketch, pixel art, and 3D-render looks, or does everything come out with the same glossy sheen?
Reference images. Can you feed it a face, a costume, or a previous output and ask for a variation? This single feature decides whether consistency is easy or miserable.
Regional editing. Inpainting to repair a hand, outpainting to widen a scene into a banner — both save enormous time.
Aspect ratio control. Portraits want tall frames, establishing shots want wide ones, token art wants square.
Consistency features. Seeds, style references, character references, and trainable personal styles.
Resolution and upscaling. Can you get a clean 2000-pixel-wide image without a separate upscale step?
Hosted or local. Hosted tools are instant and simple. Local runs on your own hardware, gives the most control, and never queues behind other users.
Motion output. Some tools animate a still portrait into a few seconds of movement — great for a session intro.
Terms of use. If you ever want to publish, print, or stream your campaign, read the license before you generate, not after.

Quick profiles

Midjourney remains the favorite for painterly fantasy aesthetics and moody lighting. Stable Diffusion, especially through a node-based interface, is the control champion — you can dictate pose, depth, and composition with structural references, and train a personal style on your own images. DALL·E follows long, literal prompts unusually well and is friendlier for props with words. Ideogram handles typography better than most, which matters for posters and handouts. Adobe Firefly markets itself on a commercially safer training position, which some creators value for published work. Leonardo and Krea sit in between, offering real-time sketching and strong control panels without a full technical setup. For animated intros, Runway and similar video tools can push a still image into motion.

A decision shortcut

If you need forty consistent NPC portraits, pick whatever offers style references plus batch generation. If you need exact geometry, go local with structural controls and draw the walls yourself. If you need a poster with readable text, choose on typography first and everything else second.

A Repeatable Workflow for Campaign Art

Random prompting produces random results. A short, boring process produces a campaign that looks intentional. Here is the loop that works at an actual table.

Step 1 — Write a one-page visual bible

Keep it short enough to actually reread. One page containing: the era and technology level, a five-color palette, three descriptive adjectives that define the whole campaign, a list of forbidden elements (no modern zippers, no neon), a preferred art style, and a lighting rule such as low-key torchlight with cold highlights. This page is the single most valuable thing you will write, because every prompt inherits from it.

Step 2 — Lock a prompt template

Stop writing prompts from scratch. Build one template and swap the subject each time:

[subject + role] + [pose or action] + [wardrobe and gear] + [environment] + [lighting] + [art style] + [camera and framing] + [quality notes]

Filled in for a cleric: human cleric in her fifties, holding a cracked censer, worn ivory vestments with silver thread, standing in a flooded cathedral nave, shafts of cold dawn light, oil painting with visible brushwork, three-quarter portrait, 85mm lens, shallow depth of field, muted palette

The template does two jobs: it speeds up writing, and it keeps the style language identical across every asset, which is what makes a folder of images feel like one world.

Step 3 — Generate in batches, triage fast

Ask for twelve to twenty variations in one go. Judge at thumbnail size, where composition and silhouette are visible and small errors are not distracting. Keep three. Note which seed or reference produced the best one. Do not fall in love with the first acceptable image — the third attempt is usually better and always cheaper to reach than endless nitpicking.

Step 4 — Run an edit pass

This is the step everyone skips, and it is the difference between AI art and good AI art. Repair hands and props with inpainting. Extend the frame with outpainting if you need a banner or a wide handout. Upscale. Then push every image through a light color grade — the same slight contrast curve and saturation shift — so the whole set shares a palette even if it came from different tools or different days.

Step 5 — Package and archive

Adopt a naming convention immediately: campaign_assettype_name_version. Keep an approved folder and a maybe folder. Export ready-to-use formats: square PNGs with transparent backgrounds for tokens, wide JPEGs for handouts, tall images for character sheets. When a session is three hours away, you will not want to hunt through a folder of two hundred untitled files.

Prompt Craft for Fantasy Subjects

Prompting is not magic words, it is specification. The more precisely you describe a scene the way a camera would see it, the less the model has to guess. These starting points cover most of what a campaign needs.

Character portraits

Describe the person, then the framing, then the light. Race and class alone produce generic results; age, build, weather exposure, and one distinctive detail produce memorable ones. Example: grizzled dwarven scout, sixties, braided grey beard with iron rings, leather jerkin stained with road dust, seated on a broken stone wall, overcast morning light, ink and watercolor with loose linework, waist-up portrait, soft background blur.

Two rules worth following. First, describe wardrobe across every character in the party using shared vocabulary — if one member wears a wool traveling cloak, use that phrase for all of them so the clothing language matches. Second, avoid prompting in the style of a named living artist. It is both an ethical problem and a legal one, and describing the technique instead — visible brushwork, flat ink shading, gouache texture — works just as well.

Monsters and creatures

Models default to familiar anatomy, so unusual creatures benefit from function-first descriptions. Instead of naming a beast, describe how it moves and what it does: a quadrupedal ambush predator with an elongated skull, translucent skin over bioluminescent organs, low crouching stance in a peat bog, cold blue rim light, creature concept art, full body side view. Silhouette prompts — asking for a black shape against a pale sky — are an excellent way to test whether a design reads clearly before you invest in a detailed render.

Environments and establishing shots

Build scenes in three layers: a foreground element to frame the shot, a midground subject, and a background landmark. Add weather and time of day. Example: ruined watchtower on a cliff edge, crumbling stone with ivy, storm clouds gathering over a frozen valley beyond, wet slate and moss in the foreground, late afternoon grey light, matte painting, wide establishing shot. If you need a specific scene for a specific plot point, generate several angles of the same location and reuse them as the party travels through.

Props, handouts, and in-world documents

Letters, bounty posters, potion labels, and treasure maps are prop work, not illustration work. Generate the paper texture, wax seal, and ink staining, then place your own readable text with a layout tool. The result looks intentional instead of garbled. Keep a shared template for parchment so every handout in the campaign feels like it came from the same archive.

Battle Maps, Tokens, and Virtual Tabletops

Tactical assets obey geometry in a way portraits never do, so treat them differently.

The standard assumption is a five-foot square on a one-inch grid. For print at 300 DPI, that means roughly 300 pixels per square. A modest 24 by 36 inch map therefore wants a file around 7200 by 10800 pixels, which is beyond what most generators produce cleanly. The practical answer is to design the layout yourself — a simple grid tool is enough — and use AI art only for texture, decoration, and viewpoint illustrations.

Tokens are the friendlier target. A one-inch circular token wants about 300 pixels of diameter at print resolution, exported as a transparent PNG. Generate a tight head-and-shoulders crop with consistent lighting, cut it to a circle, and add a thin border ring in a color coded by faction. A uniform border radius and outline thickness does more for visual cohesion than any amount of prompt tuning.

For virtual tabletops, consistency beats resolution. Export tokens as square PNGs with transparent backgrounds, keep file sizes reasonable for slow connections, and name files the way your table software sorts them — usually by faction, then name. If you generate isometric scene art, use it as a backdrop layer, not as the grid itself.

Keeping Visual Consistency Across a Long Campaign

Continuity is the hardest part of AI art and the part that most improves the result. Five techniques do most of the work.

Reuse reference images. Feed the approved portrait of a character back into the tool whenever you need a new pose. Reference beats description every time.
Reuse seeds. If your tool exposes seeds, keep the seed for any character you will draw again.
Freeze a style block. Copy the same style, lighting, and camera phrases into every prompt for a given campaign arc.
Lock wardrobe and props. Write down what each character wears and never change it. Consistency of clothing reads as consistency of person.
Grade everything at the end. One shared color adjustment across a folder hides a surprising amount of stylistic drift.

Accept some imperfection. Minor variation reads as different lighting or a different day, not a different person, as long as the silhouette, palette, and outfit stay stable.

Ethics, Attribution, and Table Etiquette

This subject deserves a straight answer rather than a slogan. Most image models learned from enormous collections of human artwork, and that fact has real consequences for working artists.

Practical positions that most tables can live with:

Check your tool's terms before you publish, sell, or stream anything you generate.
Do not imitate a named living artist. Describe technique instead.
Disclose when you use it. A one-line note in your session recap is enough, and honesty prevents awkward conversations later.
Keep AI art for home games. If you are selling a product, commission or license human art for the cover and key illustrations.
Do not claim authorship of an image you prompted. Say you directed it, which is accurate.
Raise it at session zero. Some players have strong feelings, and a two-minute conversation avoids a mid-campaign argument.

The broader principle is simple: use these tools to fill the visual gaps in your home game, and keep paying human artists for the work that represents your campaign to the outside world.

Common Mistakes and How to Avoid Them

One enormous prompt. Long prompts are not automatically better. Split description into subject, environment, light, and style, and cut anything the model ignores.
No style template. Without a repeated style block, every asset looks like it came from a different campaign.
Judging at full size. Triage at thumbnail size, then inspect the survivors.
Using a hero shot as a battle map. Beautiful, unusable. Different job, different tool.
Ignoring hands until print day. Fix them at the edit pass, not after you have printed forty tokens.
Mixing aspect ratios. Collect portraits tall and scenes wide from the start so layouts do not fight you later.
Endless generation. Set a cap: twenty candidates, keep three, move on. Prep time is finite.
Replacing description with pictures. Show the image, then still describe the smell of the room and the way the innkeeper avoids eye contact.

FAQ: AI Art for D&D

Do I need an expensive computer? No, if you use a hosted tool. Local generation gives more control and privacy but wants a capable graphics card and some setup patience.

How do I keep a character looking the same across many images? Use a reference image of the approved portrait, reuse the seed if available, and repeat an identical style and wardrobe block in every prompt. Consistency comes from repetition, not luck.

What is the best tool for battle maps? Honestly, a dedicated map tool or a simple grid drawing program. Use image generation for textures, props, and viewpoint art, then place them on a grid you control.

How do I fix mangled hands and weapons? Inpaint just that region with a tight prompt describing the hand or the hilt. Cropping and reframing is often faster than repair.

Can I animate character portraits? Yes. Several video tools will turn a still image into a few seconds of subtle motion — blinking, drifting smoke, a slow push-in. Excellent for a session-opening title card.

Will my players mind? Most groups accept it warmly once they see consistent, campaign-specific art. Be transparent about it, and ask at session zero if anyone objects.

Can I sell a campaign that uses generated art? Only if your tool's license permits commercial use, and even then, consider it a placeholder for commissioned work on anything customer-facing.

How many images do I actually need to start? Fewer than you think. One party portrait set, one villain, three location establishing shots, and a token sheet will carry a dozen sessions.

The technology will keep changing, but the workflow will not. Write a visual bible, lock a prompt template, generate in batches, edit ruthlessly, grade everything together, and be honest about how the images were made. Do that and your table stops asking what the villain looks like — because they already remember the face.