Playing AI Dungeon
Intro to Image Generation
Here are some basics to get you started generating incredible-looking images in AI Dungeon using Stable Diffusion.
Generating Images through AI Dungeon using Stable Diffusion requires you to spend credits to create images.
In adventure settings, you will find a new tab labeled
Art next to
Story. You must change the image source to
Stable Diffusion by clicking the right side drop down arrow to utilize this new feature in your stories.
Once chosen, go back to the main input bar and click on the
Do/Say/Story button until the new
See mode is selected. Then, like you would with Do, Say, or Story, type a word or two of something you’d like to now see in your adventure and press enter.
The AI will think for a moment and produce an image from your word prompt. There are a few things you can do with that picture.
Click on the generated image and your prompt will appear over the picture. You may notice extra words were added to the end; this is on purpose. Appending these terms helps the AI generate a higher quality image, so we have automatically added them to the end of your own prompt in an effort to generate the best possible output.
You will also see three dots
••• appear in the upper right corner of your image. Clicking on those dots brings up multiple options:
Love the generated picture?
Share Image or
Download Image lets you show or save the image outside of AI Dungeon.
New Image if you are dissatisfied with the generated picture and want the AI to try again. This allows you to edit your original prompt before requesting another image.
Remove Image if you are dissatisfied with the generated picture and do not want the AI to try again. This essentially deletes that output and will not generate the exact picture again.
Undo to Here work as they would with text to completely undo inputs and go back to previous parts of your story. Using the
Redo input command immediately after undo actions will bring back the exact same picture.
Be sure to
Report images you feel are not up to our community guidelines.
Practice generating story images for a while with just a few words. Prompt text allows anything in Unicode, including emojis, slang, misspellings, and foreign languages (though none work as well as English). Capitalization is ignored by the prompt, but punctuation may help with modifiers in a series (i.e. “blue ball, red brick” vs “blue ball red brick”).
Basics of Prompt Engineering
The phrases and terms below are important to learn as you start building longer prompts. At first, integrate one or two with your simple word prompts to see how they affect your pictures. Eventually, you’ll combine multiple phrases to generate a higher quality image, but that’s skipping a bit ahead.
Medium: a watercolor painting of…; a photo of…; a portrait of…; an illustration of…
Camera Angle: close up of…; overhead view of…; view from the river of…
Adjectival Modifiers: dramatic; red; stone
Foreground: ship, castle, dragon
Background: in the forest; on a beach; buildings in a city
Style: art nouveau; orientaliste; minimalist; dark and foreboding; Pixar style
Quality HD: beautiful; sharp; intricate; award winning
Artist: by van Gogh; by Donato; by Tyler Eldin
When constructing your prompt, start with the kind of image you want: a watercolor painting, a close up photo, a pencil sketch, etc.
Then, list adjectival modifiers with what they are modifying: red flag, stone wall, dramatic lighting, et.
Now add the main subject of your picture, what you want seen in the foreground: ship, castle, dragon, et.
Describe the background: in a forest, on the beach, above the clouds, et.
List specific art style(s) you’re after: art nouveau, cyberpunk, pop art, et.
Include terms about quality: sharp, beautiful, intricate, et.
Finally, put the artist(s) you’d like the AI to mimic: by Van Gogh, by Donato, by Michelangelo, et.
Research artists who make the kind of art you like. Mix and match their styles to create something original. Listing multiple artists in your prompt allows the AI room to create new and interesting designs and often produces a better outcome too, because it may not have enough context about your desired subject to generate an image in a single style.
When you feel prepared, see the next section to take your AI Dungeon pictures to the next level.
Create Complex Images
Now that you’ve generated some pictures with simple one or two word prompts, you can dive in to the world of image generation with these more advanced directions.
Structure the Written Prompt
The most important aspect of AI-powered image generation, your prompt is a set of instructions the AI uses to create the picture you request. This prompt should contain 400 characters or less and describe the subject, style, and specifics of the image you are after. Short, simple prompts can create more realistic outputs, while long, detailed prompts can create more intricate outputs.
[Medium] of [Subject] in [Style] with [Details] by [Artist(s)]
Implement Different Designs
The AI needs general artistic direction to create your desired image. Including specific style terminology is the best way to influence the look of your artwork in the correct way.
In the style of…
The ability to mimic renowned artists’ works is an incredible aspect of AI-created art. Referencing an individual artist in your prompt can help the AI generate a more coherent, relevant image. Referencing a popular style, like an animated show or movie, or particular production group, like Pixar or Studio Ghibli, will also guide the AI to create with those characteristics.
If the model repeatedly falls back on something generic, it doesn't have enough reference examples of that particular artist+topic. To avoid this, list multiple artists and let their styles combine to make something new and less derivative.
Include Additional Details
After defining your subject and style, descriptive modifiers bring your artwork to the next level. Describe the picture’s
mood (light, dynamic, bleak, dark);
energy (passionate, subdued, forbidding);
size (monumental, ornate);
structure (geometric, spontaneous); or
lighting (well-lit, dramatically lit).
Describe a situation, rather than an expression. If you want someone to look worried, describe a situation that would make them worried, rather than saying "worried expression".
Adjectives often get mixed up between two objects, so keep things fairly simple. For example, the model doesn’t track differences between "a blue ball on a red plate" and "a red ball on a blue plate" very well.
When generating a person, include a reference name to improve portrait quality. For example, use "looks like a young Albert Einstein" instead of just "man".
Use age to control appearance. For example, if you want your character to be a man with wrinkly skin and a grey beard, you could simply ask for an 80-year-old man.
Tip: the model is terrible at drawing hands. Try to set up your prompt so hands don’t appear in your final image.
Fun fact: the model is also pretty bad at generating dinosaurs, but dragons always turn out well.
Influence the Final Output
Try different ways of formulating what you are asking for. Repeating terms in various ways can emphasize them more, so a prompt including the words ‘mechanical, robot, and android’ is more likely to generate an android than a prompt that just says ‘android’.
Adding descriptive elements toward the end of your prompt helps the AI understand and generate distinct conditions in the output. Your work could benefit from words and phrases like: highly detailed, surrealism, triadic color scheme, smooth, sharp focus, matte, elegant, the most beautiful image ever seen, illustration, digital paint, dark, gloomy, octane render, 8k, 4k, washed colors, sharp, dramatic lighting, beautiful, post-processing, picture of the day, ambient lighting, epic composition, etc.
Going bigger can make things look more epic, but sticking with 512 x 512 gives the best composition for most subjects. Going above this size increases chances for poorer quality output. For portraits, stick to 512 x 640. For landscapes, 1024 x 512 is often okay because the AI easily fills the background.
For more in-depth tips and tricks, check out this guide.
© Latitude 2023