Create Custom D&D Characters With AI Image Generators

Why AI Image Generation Changed Character Prep

Few things stall a new campaign faster than the gap between the hero in your head and the blank circle on the battle map. For decades, that gap was closed with either artistic skill or money — a commission queue measured in weeks, a sketch you redrew six times, or a hastily cropped piece of art borrowed from somewhere on the internet that never quite matched your concept.

AI image generators collapsed that timeline. A player who can describe a character in two sentences can now see a finished portrait in under a minute, iterate on it a dozen times, and arrive at session zero with art that actually matches the backstory. Dungeon masters can populate an entire tavern with distinct faces before the first die is rolled. The bottleneck moved from rendering to deciding — and that is a much better problem to have.

The catch is that the tool alone does not produce a usable character sheet asset. A random fantasy render with beautiful lighting is not the same as a portrait that reads clearly at 200 pixels wide, holds a consistent face across five scenes, and fits the visual tone of your table. This guide walks through the workflow that turns a prompt into art you will actually use.

What Makes a Character Portrait Work at the Table

Before touching a prompt box, decide what the image has to do. Character art serves several different jobs, and each one has different requirements.

A token lives inside a 1-inch square. It needs a strong silhouette, a face that reads at small size, and a background that either disappears or frames the head. Fine embroidery is wasted here; a distinct helmet shape is not.

A character sheet portrait is viewed at medium size, often printed. It can carry more detail — scars, holy symbols, the specific shade of a cloak — and it benefits from a clear light source so the form does not flatten out in grayscale printing.

A handout or reveal image is the dramatic one. This is the full-body illustration you show when the villain steps out of the shadows, and it can afford atmosphere, dramatic lighting, and environmental storytelling.

A reference set for consistency is the least glamorous and most valuable. Four to eight images of the same character in different poses, angles, and outfits give you a visual bible you can hand to a human artist later, use for AI-generated scene art, or simply keep straight in your own head when describing your character at the table.

Write those requirements down. Most disappointing AI character art comes from generating a handout when you needed a token, then trying to crop it into something it was never designed to be.

The Prompt Formula for Fantasy Characters

Text-to-image prompts for fantasy subjects work best when they are structured rather than poetic. A reliable formula is: subject → silhouette → gear and detail → pose and framing → lighting → style and medium → technical constraints.

Subject and silhouette first

Start with who the character is in plain terms: "a half-orc paladin in her forties, broad-shouldered, shaved head with a braided crest." Silhouette words — broad-shouldered, tall, wiry, hooded, hunched — do more for recognizability than any adjective about personality. If your character is meant to be instantly identifiable across a table, the outline is what carries that.

Gear, materials, and cultural cues

Name materials instead of vague qualities. "Weathered steel plate with riveted pauldrons and a cracked leather belt" gives the model something to render, while "epic armor" gives it nothing. Include two or three signature items and stop — a character with a lantern, a broken shield, and a raven familiar reads better than one carrying twelve props.

Pose, framing, and camera

Specify the shot. "Three-quarter bust, facing slightly left, eye contact with the viewer" produces a portrait you can crop into a token. "Full body, dynamic action pose, low angle" produces a handout. Adding a lens or framing hint — "50mm portrait," "wide shot," "head and shoulders" — nudges composition in a predictable direction.

Light, mood, and medium

Lighting determines whether an image feels like a painting or a photograph. "Warm rim light from a campfire, cool moonlight fill" is specific enough to be reproducible. Then name the medium: oil painting, digital concept art, ink and watercolor, dark fantasy graphic novel. Medium words affect texture and edge treatment far more than style buzzwords do.

Technical constraints

Finish with the practical stuff: aspect ratio, background treatment, and negative prompts. "Neutral dark background," "no text," "no watermark," "centered composition" all prevent rework. Aspect ratio is not cosmetic — a 1:1 square suits tokens and sheet portraits, while 3:2 or 16:9 suits handouts and stream overlays.

Reference Fusion and Visual Consistency

The hardest problem in AI character art is not quality. It is consistency. The same prompt rarely returns the same face twice, and a character whose nose changes between every image never feels like a real person.

Reference image fusion solves most of this. Instead of describing a face in words, you supply source images — a portrait you liked, a photograph of a specific face shape, an armor reference — and let the tool blend their features with your text prompt. Because the model is working from pixels rather than adjectives, the resulting face is far more repeatable.

A practical setup looks like this:

• Anchor portrait: one image you consider canonical. Every future image references it.
• Detail references: separate images for armor, weapons, and color palette, so you can swap outfits without changing the face.
• Weighting: keep the anchor weighted heavily for facial features and lightly for clothing, so costume changes do not drag the face along with them.
• Seed stability: where a tool exposes a seed value, reuse it. Changing the seed reshuffles everything, including the parts you liked.

Build a small consistency kit for each player character: the anchor portrait, two alternate angles, one full-body shot, and one expression study. That kit is worth more than fifty one-off images, because it lets you generate new scenes later — the character in a throne room, in a rainstorm, at a funeral — without losing who they are.

A Step-by-Step Workflow: From Concept Note to Finished Portrait

Here is the process that consistently produces usable art without endless rerolling.

1. Write a one-paragraph character brief. Name, ancestry, class, age, build, three signature visual traits, and the emotional tone you want. This is your source of truth and it prevents prompt drift.

2. Generate a wide first pass. Run eight to twelve variations with the same core prompt, changing only light and medium. Do not chase perfection here. You are hunting for a structure you like — a face shape, a pose, a color scheme.

3. Pick one and lock it as the anchor. Save it. Note the prompt and any seed value. Everything downstream references this image.

4. Refine in small steps. Change one variable per pass: swap the armor, adjust the lighting, tighten the framing. Multi-variable changes make it impossible to know what improved the image.

5. Produce the derivatives. From the anchor, generate a token crop, a full-body handout, and two or three scene images. Crop the token manually rather than asking the model for a square close-up — you will get a tighter, better-composed result.

6. Upscale and clean up. Upscale before you print. Fix obvious artifacts — extra fingers, melted straps, floating jewelry — with a quick pass in an image editor. A five-minute cleanup separates hobby art from something worth showing the table.

7. Archive the whole set. Keep the character brief, prompts, anchor, and derivatives together in one folder. When you return six months later, you will be able to rebuild the character's look instead of starting over.

Building a Cast Quickly: NPCs and Monsters

For dungeon masters, the value of image generation is volume. A campaign needs dozens of faces, and most of them need to be memorable for exactly one scene.

Work in batches with a shared visual grammar. If your campaign is set in a rain-soaked trade city, give every NPC the same palette — wet stone grays, lantern yellows, faded indigo — and vary only the subject. That single constraint makes a cast of thirty characters feel like they belong to one world instead of thirty different art styles.

For monsters, describe function before appearance. "A quadruped ambush predator built for low ceilings, no eyes, long forelimbs" generates more interesting creatures than "scary cave monster," because it gives the model anatomy to solve. Then iterate on the result: keep the silhouette, change the surface — chitin, scales, fungal growth, rusted plating.

Reuse is your friend. Generate a base creature, then create three variants with small prompt changes to represent a leader, a juvenile, and a wounded specimen. The visual family resemblance tells your players something about the encounter before you say a word.

Tokens, Handouts, and Virtual Tabletop Assets

Virtual tabletops have specific technical needs, and ignoring them produces art that looks great in a gallery and terrible in play.

Tokens want a centered head or upper body, strong contrast between subject and background, and a readable silhouette. Generate at a high resolution, then crop to a square with the eyes roughly one-third from the top. A subtle vignette or dark outer ring helps the token separate from a busy map.

Handouts want a 3:2 or 4:3 frame, room for text if you plan to caption them, and a slightly wider composition. Leave negative space on one side — a face pressed against the frame edge looks cramped when you drop it into a handout layout.

Scene and background art for maps benefits from a consistent camera angle. If your battle maps are top-down, generate top-down or near-top-down environments so the perspective does not fight the grid.

Player portraits in a VTT sidebar usually render around 256 pixels. Test at that size before committing. Detail that vanishes at thumbnail scale is detail you paid prompting time for and cannot use.

Motion: Animated Tokens and Recap Clips

Static art is the baseline, but a small amount of motion goes a long way at the table. Animated tokens — a slow idle breathing loop, drifting embers around a fire genasi, a cloak that shifts in the wind — make a virtual tabletop feel alive without becoming a distraction.

The efficient approach is to animate from a finished still rather than generating video from scratch. Start with your cleaned-up portrait or token, define two or three keyframes, and interpolate between them. Short loops of two to four seconds work best: long enough to read as intentional motion, short enough that nobody notices the seam.

For recap videos and session recaps, the same principle scales up. Assemble a sequence of your strongest handout images, add slow pans or parallax, layer in music and captions, and you have a recap your players will actually watch. Consistency matters more than motion quality here — if every image in the sequence features a recognizably identical party, the recap feels like a story rather than a slideshow.

Mistakes That Ruin AI Character Art

Chasing perfection on the first prompt. Rerolling a single prompt fifty times is slower than generating twelve variations and choosing. Volume beats stubbornness.

Overloading the prompt. Twelve adjectives about personality do nothing. Five concrete visual facts do everything. Personality belongs in your roleplay, not your prompt.

Ignoring silhouette. Beautiful rendering with a generic outline produces a forgettable character. Test your design by mentally filling it in black — if you cannot tell who it is, the design is not finished.

Losing the anchor. Generating without a reference image guarantees the face will drift. Every new image of an established character should start from the canonical portrait.

Forgetting the final crop and cleanup. AI output is a draft, not a deliverable. Cropping, color correction, and artifact cleanup are part of the workflow, not optional polish.

Skipping table etiquette. Be clear with your group about how images were made, avoid prompting in the recognizable style of a living artist, and do not use generated art to imitate a specific published character wholesale. Generated art is a personal aid for your home game, and treating it that way keeps the table comfortable for everyone.

FAQ: AI Character Art for Tabletop Games

How many images do I need per character?

Four to eight is the sweet spot: one anchor portrait, one full-body shot, one token crop, one expression study, and a couple of scene images. Beyond that, returns drop fast unless you are producing art for a published product.

Can I keep the same face across an entire campaign?

Yes, but only with reference images. Lock an anchor portrait, reuse the same seed where available, and change one variable per generation. Text prompts alone will not hold a face over dozens of images.

What aspect ratio should I use?

Square for tokens and character sheet portraits, 3:2 or 4:3 for handouts, and 16:9 for stream overlays and recap video frames. Decide before you generate, because cropping a carefully composed 16:9 image into a square usually destroys the composition.

Is it better to generate a token directly or crop one?

Crop. Generate at high resolution with a wider composition, then crop to a tight square in an editor. Direct square prompts tend to produce awkward head placement and dead space.

How do I make NPCs feel like they belong to the same world?

Fix a palette and a light source for the whole cast, then vary only the subject and costume. Shared color temperature does more for world cohesion than any amount of shared lore text.

What should I do about artifacts like extra fingers?

Treat them as normal. Upscale first, then fix them in an image editor, or generate two or three variants of the same prompt and composite the clean parts. At token scale, most minor artifacts disappear entirely — check at final display size before spending time on repairs.

Do I need different prompts for different ancestries and classes?

You need different visual nouns, not different templates. The formula stays the same — subject, silhouette, gear, pose, light, medium — but the concrete details change. A dwarven cleric and an elven rogue use the same structure with completely different nouns filling it.

How long does a full character set take?

With a locked workflow: fifteen to twenty minutes for the first character, including brief writing and cleanup. Subsequent characters in the same campaign go faster because your palette, medium, and framing decisions are already made.