Every DM knows the moment. The party walks into the tavern, you describe the scarred half-orc behind the bar, and someone asks “do you have a picture?” You do not. You have a browser tab of stock fantasy art where nobody matches your description, a map you sketched on graph paper at 1am, and four players waiting.
Session prep has always been bottlenecked by art. You cannot commission a portrait for every NPC the party might talk to, and you certainly cannot commission forty of them in a style that matches. This is the one creative problem image models are unambiguously good at solving, because a campaign’s visual needs are repetitive, reference-driven, and consumed on a phone screen for ninety seconds.
What follows is the practical version: which assets are worth generating, the art-direction decisions that make forty images look like one campaign, how battle maps differ from everything else, and the four things you should not generate at your table. The Nano Banana family suits this work because it holds a face and a style across repeated generations, and Nano Banana Pro is the one to use for anything you will print as a handout or project on a table screen, where resolution actually shows.
The Five Asset Types, and What Each One Needs
| Asset | Where players see it | Shape | Style rule | Worth generating? |
| NPC portrait | Phone, Discord, VTT token | Square or 4:5, head-and-shoulders | Consistent medium across the whole roster | Yes — the highest-value asset by far |
| Location establishing shot | Shown once when the party arrives | 16:9 wide | Same palette and light as the campaign bible | Yes |
| Battle map | VTT, tablet, printed | Top-down, gridless export | Flat lighting, no dramatic shadows | Partly — see below |
| Handout / prop | Passed around, or printed | Portrait page | Real type set by you, never generated | Image yes, text no |
| Item / loot card | Inventory, Discord pin | Square | Isolated object, plain background | Yes |
The portrait row is where almost all the value sits. A recurring NPC with a face your players recognize does more for immersion than a beautiful establishing shot they glance at once.
Build a Campaign Style Bible Before You Generate NPC Forty
The failure mode is not ugly images. It is forty images that each look like a different game — one oil painting, one anime bust, one photoreal render, one storybook sketch. The party stops reading them as a world.
Fix it by deciding three things once, writing them in your campaign notes, and pasting them into every prompt:
- Medium. “Muted digital oil painting, visible brush texture, no line art.” Or “inked woodcut with limited color.” Pick something and never drift.
- Palette. Three to five colors. “Cold slate blue, oxblood, tarnished brass, bone white.”
- Light. “Single low warm source, heavy shadow, faces half in darkness.”
Then generate two or three portraits that nail all three, and use those as your style references for everything afterward. This is the single highest-leverage habit in the whole workflow. A campaign where the shopkeeper, the duke and the lich were clearly painted by the same hand feels made; one where they were not feels scraped.
NPC Portraits: The Recurring-Character Problem
One-off NPCs are easy. Recurring ones are where it gets interesting, because your players will absolutely notice when the quest-giver’s scar switches cheeks between sessions.
For anyone the party will meet more than twice, build a small reference set the first time: one frontal, one three-quarter. Then for every later appearance — wounded, older, in court dress, possessed — feed that reference and describe only what changed. Keep the face structure, eye color, scars and hair in the “do not change” half of the prompt every single time, even when the new prompt seems unrelated.
A trick worth stealing: generate the whole recurring cast in a single session rather than one per week. Consistency is far easier to hold when the style references are fresh and you are comparing outputs side by side.
Battle Maps Are a Different Job Entirely
Do not use your portrait style here. Battle maps have functional requirements that pretty art actively fights:
- Top-down, flat lighting. Dramatic side-lighting throws shadows that hide half the floor. Ask for even overhead light.
- Generate gridless. Image models are bad at drawing consistent square grids — they warp, skip, and drift. Generate a clean gridless map and let your VTT overlay its own grid, which is also the only way the squares will line up with token movement. Check your platform’s expected pixels-per-square before exporting.
- Readable terrain, not realistic terrain. Players need to tell floor from pit from difficult terrain at a glance on a tablet. Say “clear boundaries between surface types, high contrast between walkable floor and obstacles.”
- No text. No room labels, no compass rose with lettering, no legend. Add those yourself if you want them.
Honest caveat: for complex tactical encounters with specific chokepoints, purpose-built map tools still beat prompting, because you need control over exact geometry. Generated maps shine for organic spaces — a forest clearing, a flooded cellar, a marketplace — where “roughly this shape” is fine.
Handouts: Generate the Paper, Set the Type Yourself
The torn letter, the wanted poster, the cultist’s ledger page. These are the props players physically hold, and they are the ones people most often ruin by asking the model to write the text.
Generated lettering falls apart at readable size, and a handout is defined by being readable. Generate the surface — aged parchment with a burned edge, a bloodstain, a wax seal, a blank poster with a portrait space — and set the actual words in any design tool with a real font. Your players get something legible, and you get to control exactly what clue the text contains.
Weak Prompt vs Working Prompt
| Weak | Working |
| “Fantasy tavern keeper, half-orc, D&D character art” | “Head-and-shoulders portrait of a female half-orc innkeeper, mid-forties, broken left tusk, burn scar across the right jaw, grey-streaked black braid, stained linen apron. Muted digital oil painting, visible brush texture, no line art. Cold slate blue and oxblood palette. Single low warm lantern light from the left, face half in shadow. Plain dark background. Square crop. No text.” |
| “A creepy dungeon map” | “Top-down battle map of a flooded stone cellar, roughly square. Even overhead lighting, no cast shadows. Clear contrast between dry stone floor, shallow water, and rubble. Four support pillars, one collapsed stairwell at the north edge. No grid lines, no text, no labels.” |
The pattern in both: one specific physical detail per clause, the style bible restated, and the exclusions named. “No text” belongs in essentially every prompt you will ever write for a table.
When an NPC comes out almost right — good face, wrong cloak color — edit that one element in Nano Banana AI rather than regenerating. A reroll gives you a new face; a local edit keeps the one your players already met.
Four Things Not to Generate
- A player’s character, unprompted. PC art belongs to the player. Some love it, some have a specific mental image they do not want overwritten, and some commission real artists. Ask first, every time.
- “In the style of” a living artist by name. Describe the medium and technique instead. Naming working illustrators in prompts is the fastest way to make your table uncomfortable, and the technique description gets you a better result anyway.
- Anything presented as commissioned or hand-drawn. If you publish your campaign, share maps online, or post the NPC gallery, say what it is. Nobody minds AI prep art. People mind being told it was something else.
- Assets for a product without checking the rules. Selling a module on DriveThruRPG, Itch, or a Kickstarter brings platform disclosure policies and marketplace rules into play, and they have changed more than once. Home table is one thing; commercial release is a different conversation.
Conclusion
The DM’s art problem was never about quality. It was about volume, consistency and a Tuesday deadline. Write a three-line style bible and restate it in every prompt, build reference sets for recurring NPCs, keep battle maps gridless and flat-lit, generate the parchment but set the words yourself, and put “no text” in every prompt by default. Do that and you walk into the session with a roster of faces that look like they come from the same world — which, for the ninety seconds your players spend looking at them, is exactly as much art direction as the game needs.
