More Fun with A.I. Art

A few weeks ago, I wrote a post about my adventures with A.I. art. This is a followup with more images and more lessons learned. I've mostly had more fun with A.I. art, but I'm also discovering its limitations. And my limitations, too.

Prompts: The Key to Success

Prompts are the string of keywords you provide to the AI from which to generate an image. They can be quite short (e.g., “leaves on the current”), leaving the AI a huge latitude in creating and image, or extremely detailed and full of technical instructions, e.g.,

mine excavation on an asteroid in deep space, mining, geodesic domes, smelting, tron cycle, Chronicles of Riddick, concept art, landing zone, starship, dramatic lighting, ultrarealistic, cgsociety, cinematic lighting, --ar 9:16 --v 3 --no noise --no laser

(Scroll to the bottom to see what MidJourney did with this prompt.)

Crafting prompts is part experience, part inspiration, and part luck. I don't have the time or patience to tinker with a prompt through 500 iterations to get what I'm envisioning. I'm more inclined to try a group and go down a temporary rabbit hole. It's fun to enjoy the comparatively rare successes, and mostly shake my head in perplexed wonder at the failures. My author buddy S.J. Pajonas is creating one image per day to post on her blog and Facebook page, so she re-rolls the variations to get an image she likes. She's better organized than I am. 😸

Tip: Look for Reddit threads and Discord servers where users are kind enough to share images and prompts. They'll give you ideas for prompts to try for yourself, and those users have often spent hours and days in refining their prompts to get pleasing images.

Variations

What are variations, you ask? Good question; it depends on the service. “Variations” is my term for options and variables the service lets you select that aren't (technically) part of the prompt. For example,

Seed: 1- to 12-digit number that can radically affect the outcome in unpredictable (to me) ways. The same everything else but a different seed produces quite different results.
Some services allow you to set the output resolution. This can save credits (that cost money) if you're dubious that the A.I. will produce a usable image.
Aspect ratio tells the AI how to size the generated image. Some offer standardized settings, and some produce better results for square images regardless of what options they offer.
“Prompt weight” tells the A.I. how much to pay attention to your prompt vs. freewheeling it and going its own unpredictable way.
Starter image: Some services allow you to upload an image for which you have the rights(!) to tell the A.I. what you're looking for. Do not be uploading that favorite urban fantasy cover image by Dan Dos Santos unless you licensed it from him. Otherwise, it's thievery, pure and simple, because you're uploading an image you don't own to a public server that you don't own.

A.I. Art Fun

The services I've played the most with lately are MidJourney, NightCafe, and Dall-E. I've signed up for various subscriptions so I have room to play. I can always cancel the ones I don't use. They are all addictive to varying degree, catering to the hope that the next combination of prompt and variation you try will give you an image that will blow your socks off.

NOTE: All the AI art services I've played with are fairly terrible, right out of the gate, with realistic animals and human hands. I'm sure the services will eventually improve, but don't say I didn't warn you.

NightCafe failure with dogs - AI art fail

NightCafe's demented idea of dogs (click to enlarge)

MidJourney's equally demented idea of hands (click to enlarge)

MidJourney

MidJourney (MJ) is still in beta-test mode, so all the prompting is done via sending commands to the MJ bot connected to a Discord server. The interface and organization are confusing for beginners and requires some crashing around to find instructions. Prompts allow a lot of control, provided you learn the arcane magic to create them.

Click to enlarge. Prompt: “copper brick road leading to the ocean shore, green and white splashing waves, seagulls in the far distance, sunrise, light fog, magical realism, HD, cinematic lighting, –ar 5:3”

Click to enlarge. Popular advice says that the first 5 words of a prompt are the most important in MidJourney. Someday when I'm in a stubborn mood, I'll test it. Prompt: “movie poster, futuristic starship freighter in deep space, blue halfrear lighting, in the style of Jeff Brown and Stephan Martiniere and Chronicles of Riddick and Guardians of the Galaxy and Dune, planet background, red halfrear lighting, distant deep space universe, concept art, matte painting, trending on artstation, trending on deviant art :: defocus::-0.5 –ar 9:16”

NightCafe

For reasons I don't recall, I periodically get free credits for NightCafé. I'd probably be better with this service if I used a methodical approach to test prompts against the various options. NC recently integrated the DALL-E 2 engine as an available algorithm, but I haven't played with it much.

Click to enlarge. This is the raw image; I used graphics tools to clean up the mistakes and used it as a background for another purpose. Prompt: “ethereal music notes, cyan waves of digital sound, planets, stars, romance, magical realism”

Click to enlarge. Prompt: “futuristic architecture of Abu Dhabi in the style of Chronicles of Riddick movie, cyberpunk, hyperdetailed, glowing with energy, concept art, masterpiece”

DALL-E

DALL-E is powerful. It's also frustrating for me in that I can't figure out how to create predictable results, unless you count predictable failures. Nonetheless, I keep trying because I've seen stunning results produced by others.

Click to enlarge. Prompt: “travel poster, foreground massive spaceship intricate detail, planet in the background, distant universe, red halfrear lighting 360 panorama, matte painting, concept art”

glowing cat illustration by DALL-E - fun with AI art

Click to enlarge. Note that DALL-E simply ignored the pawprints instruction altogether. Prompt: “A magical glowing cat sitting on top of a hill looking at the northern lights, neon glowing pawprints leading up the hill, high detail, digital art”

Bonus Fun with A.I. Art

Here are the four images MidJourney came up with from the example prompt above. MJ will let me choose any of the 4 to upscale or use as a seed image to roll variations. If I don't like any of these well enough to mess with them, I can start with a new, tweaked prompt to get more what I'm looking for.