PROMPT: A woman kneeling in prayer by a flowing pool, underground grotto, dripping gray stone stalactites, glowing green fungus on the walls, moody lighting, rule of three quarters

The perfect prompt is part poem, part shopping list. A haiku 77 tokens long. A scavenger hunt for an AI that some days seems to read my mind and some days just can’t quite understand that I want *all* of the character’s skin to be blue. Yes, the face as well. Yes, of course the skin around the eyes should be blue. Yes, I know that the skin around the eyes is usually some shade of brown, but for this picture it should be blue.

CUSTOM STYLE: Ooze
STYLE PROMPT: made of curvaceous transparent blue plasticien, dripping wet transparent blue plasticien skin and hair, dripping bioluminescent blue slime, bioluminescent, extremely pretty wet transparent blue plasticien face, extremely big transparent blue plasticien rendered eyes by Pixar, slimy, glossy, happy, melting transparent blue plasticien clothes, dynamic composition, dynamic lighting, soft shadows, sharp focus, $pink, pink skin, natural colored skin, low quality, worst quality, blurry, cropped, poorly drawn hands, poorly drawn face, poorly drawn eyes, cross-eyed, duplicate, cloned face, extra limbs, missing limbs, fused fingers, extra fingers, text, error$ Style: Buliojourney V2

These images were created using Dream by Wombo. When Stability AI released its open-source text-to-image generator Stable Diffusion last August, it could only run on powerful computers with seriously beefy graphics cards. Canadian app developer Wombo took Stable Diffusion, simplified it, and wrapped it in an app that lets you create supercomputer-level pictures on your phone.

The Dream app is based around “styles,” which are sort of like Instagram filters for your art. Want to make a Pulitzer Prize-worthy photo of a nun punching a clown? Use the Realistic V2 style. How about Sailor Moon punching Totoro? Try Anime V2. Subterranean fungal growths? Flora V2. (The V2 means that the style uses the latest version of the AI, Stable Diffusion 2.)

On Wombo’s Discord server, paid members can use the aptly named Wombot to generate pictures using styles that aren’t in the app, and even create styles of their own using Wombo’s styles as a base. There are thousands of user-created styles, some surpassing WOMBO’s in terms of realism and beauty.

CUSTOM STYLE: 80s-synthwave:         
STYLE PROMPT: synthwave grid in the style of outrun wallpaper on the ground and objects, dark sky, synthwave colors, neon vector graphics, palm trees Style: Anime V2 

The styles I make do not rival anything for realism. I go a different direction, seeking neon synthwave streets and bioluminescent women made of slime. Dinosaurs with anachronistic secondary sex organs. New art by Patrick Nagel, still alive in his 70s and creator of the new smash hit slice-of-life anime series, Handsome Man with Stubble and His Redhead Minion Are Just Friends, Okay?

CUSTOM STYLE: Nagel SFW                 
STYLE PROMPT: by Patrick Nagel, beautiful eyes and face by Patrick Nagel $bad anatomy, extra arms, extra legs, missing arms, missing legs, deformed, mutants, multiples, ugly, ugly face, ugly eyes, small eyes, poorly drawn eyes, messy eyes, hidrocor lenses,  matte eyes, cross-eyed, low quality, hard shadows, bad composition, lost in the background, out of frame, blurry lines, complex shading, 3d shading, photorealistic, username, watermark, signature$

Most of the things Wombot and I make have never before been seen, so my styles often require many, many tokens. But so do the base styles they are built upon. And I still need to leave some tokens for the prompt. Each object, description, or action I ask for in a picture uses up a certain amount of tokens, whether it be in the prompt, the custom style, or the base style that the custom one is based on. If a prompt goes over Stable Diffusion’s 77 token limit, bad things start to happen. Often, the styles just stop working, and the picture gets rendered with no style at all. So custom styles can’t get too complicated. But as far as I can tell, Wombo’s styles are not created using negative prompts, special prompts that tell the AI what NOT to put in the picture. Wombot makes anything you put between two dollar signs at the end of your style prompt a negative prompt, and these get their own allotment of 77 tokens. So a cunning prompter can get what they want by putting its opposite in the list of negative prompts. I wanted my characters’ eyes to have a ring around the iris to keep it from bleeding into the rest of the eye. The technical term for this is hidrocharme lenses. Eyes are hard for Stable Diffusion, so anything that describes them tends to use up a lot of tokens. Adding “hidrocharme lenses” to this style prompt left me no tokens for the regular prompt, so I tried putting its opposite, “hidrocor lenses,” in the negative prompts. Now my irises stay were they are supposed to.

PROMPT: close up portrait, 30 year old woman, pixie cut hair, beautiful face, beautiful piercing eyes by Ilya Kuvshinov

But even in the simplest of prompts, details about one part of the image can sometimes have unexpected side effects on the rest of the picture. Why does this 30 year old woman look 15? Find out in our next installment, The Curse of Ilya Kuvshinov’s Eyes.