- A Surprise for Me
- A Sample of My Midjourney Portfolio
- PromptCraft
- This is your Brain on Cyberdrugs
- The Rabbit-Hole
- Further Reading
A Surprise for Me
Recently, I discovered that I’ve been grossly underestimating just how far AI image generation has come – in terms of how powerful the models are, and how accessible they’ve become for the lay user.
Turns out, anyone can interact with many of these models via user-friendly web interfaces (Craiyon, Midjourney, Stable Diffusion), less-user-friendly web interfaces (Disco Diffusion), or by running on one’s own hardware (many, many others).
Every image you’ll see on this page was generated using Midjourney, by me, and each took less than 3 minutes to output. Continue reading this post after the gallery.
(Also, see my entire collection here)
A Sample of My Midjourney Portfolio
There’s this idea in the A.I. Art community of the “Latent Space” – a dimension of unrealized, raw potentiality – from which A.I. models pull and mold raw materials into a unique image via the text prompt they’re fed (by a human).
PromptCraft
Creating text prompts is one of the more interesting parts of generating images. The AI model is a mediator between you and the latent space, and it uses your text prompt to make art.
People call this Prompt Crafting, and there are entire websites, databases, and Discord channels devoted to both casual/unstructured discussion and exhaustive experimentation and iteration.
Shit, there’s even a marketplace for prompts.
This whole model of “feed AI words, get image output” gets kinda freaky for me. Prompt crafting (whether on not you’re aware that’s what you’re doing) is in that uncanny valley between “regular computer-touching to make it do what I want, like make a math spreadsheet” and “seance with an alien intelligence to understand how it interprets language by reverse engineering what it paints on the wall after I tell it something”.
This is your Brain on Cyberdrugs
Interested? Here’s my recommended entry-point into the scene, via Midjourney.
- Join the Midjourney Discord.
- Enter any of the “newbies-###” chat rooms.
- Type /imagine prompt:<Literally any combination of words, descriptors, names, etc. that you can come up with>
- Press Enter, and wait 60 seconds or less for your image candidates.
- Upscale (U#), or make Variations of (V#) any combination of the four candidates using the interaction buttons beneath your image. (Or any other image you see in the channel! The newbie channels are hoppin’ busy, and half the fun for me is playing off others’ generations.)
- Read the Documentation, and go nuts
- Quickly run out of free image generations
- (Optional) Pay for near-unlimited, private DM-with-bot image generation
- Fall down the rabbit-hole of AI image generation
The Rabbit-Hole
Disco Diffusion
Not beginner-friendly, with longer generation times around 10-15 minutes.
Open source.
- Discord server
- Access via Google Colab – Start here!
- Zippy’s Disco Diffusion Cheatsheet
- A Traveler’s Guide to the Latent Space – Deep dive/guide into DD.
- Disco Diffusion Artist Studies – Experiments using specific artists’ names
- Offhand’s Disco Diffusion Prompt Keywords – Prompts and their effects
- Contains a nice list of links, too, in cell E4
Craiyon (link)
(formerly DALL-E mini)
Very beginner friendly, but very little customization. Closed source.
Stable Diffusion
I haven’t used this yet. Open source via Github
DALL-E 2 (link)
Closed beta waitlist
Further Reading
- Midjourney Image Prompting by @shambibble
- CLIP Prompt Engineering by Matthew McAteer (Below the reams of prompt-image relationship tables of mushrooms, dragons, castles, etc., there are indeed many more interesting paragraphs, and MORE LINKS).
- Softology’s Blog: Text-to-Image Summary – Part 1
- My Midjourney profile page
matthew-field.ca 