Use a language model to outline your own campaign, from the villain and the key locations to the encounters and clues, then turn each scene into an image your players can see. The story stays yours, and the heavy lifting of drafting and picture-making gets shared with the tools.
Most dungeon masters do not lack ideas, they lack Tuesday night. Between the villain's motive, the three towns, the secret that links session two to session nine and the handout you promised your players, prep eats the evenings you meant to spend playing. This page splits the work in two: a language model drafts the adventure with you, then image models illustrate each scene, so the table sees the place instead of only hearing about it.
A campaign is a lot of small writing jobs stacked together: a premise, a cast, a set of places, a pile of encounters, and a trail of clues that makes the whole thing feel planned. A language model is good at exactly this kind of structured first draft. You give it the tone, the party level and the three things your players keep asking about, and it returns options you can react to, which is far easier than staring at a blank page.
The second half is the picture. Players remember a scene they were shown for ten seconds longer than a paragraph read aloud. Text-to-image models turn a line like a flooded chapel lit by green lanterns into an image you can put on the table, and the same toolkit also runs the chat model, so planning and illustrating happen in one place. Picasso IA lists 502 models in its full catalog, 75 of them language models and 212 of them text-to-image, and new accounts get free credits to try the whole loop before paying for anything.
Keep one rule in front of you: this is your campaign. The model proposes, you decide. Ask it for original material built from your ideas, and do not paste in the text of a published module and ask for a rewrite. Published adventures belong to their authors, and a campaign you built yourself is also more fun to run, because you know where every secret is buried.
Start in the toolkit with a chat model. Claude Opus 4.7, Gemini 3.1 Pro and GPT-5.4 are all in the catalog, and they differ enough in voice that it is worth trying your premise on two of them. Claude Opus 4.7 accepts a system prompt and holds a long conversation without losing earlier details, which suits a campaign bible that grows over many messages.
Open with a system prompt that sets the rules of the collaboration: you are the dungeon master, the model is a co-writer, it must ask before inventing major lore, and it should keep a running list of named characters and places. Then describe your table: how many players, their level, the tone (grim, heroic, comic), and the one theme you want. A short opening request works better than a long one, because you can steer each answer.
Plan in layers. First the premise and the villain, then the three or four locations the party will actually visit, then the encounters inside each one. Resist asking for a whole campaign in a single message; the result is usually a generic arc with names you will forget by Friday.
| Layer | Ask the model for | Check by hand before the table sees it |
|---|---|---|
| Premise | Three one-paragraph hooks with different tones | Does at least one hook connect to a character backstory? |
| Villain | A motive, a plan, and what they do if the party ignores them | Is the motive understandable, and can the party actually reach them? |
| Locations | Four places, each with a mood, a secret and an exit | Do the distances and travel days make sense for your world? |
| Encounters | Five options per location, mixing combat, social and exploration | Are the difficulty and the rules right for your party and system? |
| Clues | A web where each secret can be found in at least two places | Can the party still progress if they miss one clue? |
Tip: ask the model to write the villain's plan as a timeline of what happens each week if the party does nothing. A villain who moves on their own makes the world feel alive, and you get a ready-made list of consequences to reveal.
This is the loop most dungeon masters settle into after the first adventure. It assumes you have a rough idea and an evening, nothing more.
The hard part of illustrating a whole campaign is consistency. Ten scenes from ten prompts tend to look like ten different games. The fix is boring and effective: keep the same style sentence in every prompt, keep the same model for a given arc, and describe light and weather explicitly. A scene prompt that names the place, the time of day, the one striking object and the camera angle gives the model enough to work with.
FLUX 2 Pro is a sensible default for atmospheric landscapes and interiors, and Seedream 5 Pro is worth a try for painterly, detailed fantasy art. Nano Banana Pro accepts up to 14 reference images alongside the prompt, so you can hand it an earlier scene and ask for the same setting from a new angle, which is a practical way to keep a recurring tavern or fortress recognizable.
For characters, the same trick applies. Generate your villain once, keep the image you like, and feed it back as a reference whenever that character appears. The technique is covered in more depth on consistent AI characters, and the party heroes have their own page on AI D&D character art. For a more cinematic feel, concept art for games and film shows how key scenes get iterated, and AI storyboards is useful if you want to lay out a dramatic set piece panel by panel.
The most common failure is accepting the first draft. Models drift toward the familiar: a dark lord, a lost artifact, a tavern where the quest begins. Push back by asking for what the story should avoid, and by giving the model your own odd details, such as the town that runs on a single ferry or the villain who is secretly right about one thing.
The second mistake is illustrating before deciding. If you generate ten images and then write the scenes around them, the pictures drive the plot. Write the scene first, then picture it, so the art serves the story.
A co-writer that never gets tired is still not a game master. These are the places where you will feel the limits quickly:
None of this is a reason to skip the tools. It is the reason the final pass, the one that makes it your campaign, stays with you.
It can draft one, but a draft is not a campaign. A language model will happily produce a premise, a villain, a handful of locations and a pile of encounters in an evening, and that is a real head start. What it cannot do is know your players, balance the rules for your party or remember session three when you reach session twenty. Treat the output as raw material, rewrite what does not fit your table and run the game yourself.
There is no single winner, and the catalog lets you compare. Claude Opus 4.7, Gemini 3.1 Pro and GPT-5.4 are all available, and they differ in voice: some lean toward vivid description, others toward tidy structure. A sensible test is to give the same premise to two of them and keep the one whose villain you would enjoy running.
Reuse three things: one style sentence at the start of every prompt, the same image model for a given arc, and reference images for anything that recurs. Nano Banana Pro takes up to 14 reference images, so you can feed it the tavern from session one when you need it again in session six. Perfect sameness is not realistic and small details will drift, but a shared style line and a reference image keep a recognizable look across a whole story.
Not reliably. Image models can paint a gorgeous cavern or a ruined keep as a scene, but they rarely produce a top-down map with a correct scale, a grid and walls that line up with your encounters. Use generated images as mood art and handouts, and keep drawing or building your tactical maps with the tool you already trust. If you want a rough layout idea, ask the language model to describe the rooms and exits in words, then sketch from that.
It is better not to. Published adventures are the work of their authors, and pasting one into a model to reword it serves neither your table nor the people who wrote it. Use the tools for original material instead: describe your world, your villain and your themes, and let the model riff. If you plan to share or sell what you make, check the license of your game system and the publisher rules that apply.
For your own table and handouts, generally yes, but rules differ by model and by country, and the law on AI-generated images is still settling. If you plan to publish, sell or stream with the art, check the terms that apply to the model you used and to the platform you post on, and do not generate real, identifiable people. Fictional characters, creatures and places are the safe and the fun territory here.
Fewer than you think. Pick the moments that deserve a reveal: the opening hook, each new location, the villain, and the big set piece. Four to eight strong images for a session carry more weight than thirty forgettable ones, and they cost fewer credits.
Yes, and it is one of the quickest wins. Describe the character's look, clothing and mood in a sentence, add your style line, and generate a few options. Keep the one you like as a reference for later appearances. Monsters work the same way, but unusual anatomy can come out strange, so generate extra variants.
Yes. The planning side is just structured writing, so you can ask for a campaign in any tabletop system, or in none at all. State the system and edition at the start so the model uses the right vocabulary, and still check any rules text against your books. The image side does not care about the system, it only needs a description of the scene.
Generations are paid in credits, and the cost depends on the model and the settings you choose, so a number written here would go out of date. Single images generally cost less than video. New accounts get free credits to try the loop, and current plans are listed on the pricing page, which is the place to trust for figures.
Your next session does not need another weekend of prep. Open the Picasso IA toolkit, brief a chat model on your party and have the first scene on screen before your players arrive.
Choose your language
