Stable Diffusion — Technique and AI Art Discussions
Just a quick follow-up to the previous two posts:
- Starting Out with Stable Diffusion | by Eric Richards | Sep, 2022 | Medium, and
- Stable Diffusion + MidJourney + DALL-E Have Rung the AI-Art Bell | by Eric Richards | Sep, 2022 | Medium
I’m going to cover a small learning on technique with new references and then general discussions.
Technique and Resources Updates
The most interesting technique I’m experimenting with right now is Prompt Editing where you can provide a transition from starting at one thing and then blending to another. Within your prompt it takes the form of [from:to:when]. Like creating a hybrid creature:
Portrait photo of a [jack rabbit:antelope:0.75], prairie, golden hour, professional photo
In the above case, Stable Diffusion was asked to transition from a jack rabbit to an antelope 75% of the way through the steps. More info on using it within the Stable Diffusion WebUI under “Prompt Editing” — super great image matrix here:
Resources — Styling Your Prompt
The kind of shared information I really like right now consists of folks who are poking the model to determine what works and what doesn’t work (some might be trying to stab it with their steely knives. But. They. Just. Can’t. Kill. The. Beast. [Greg]). Hard working folks are creating excellent visual indexes showing what modifiers and which artists produce interesting results via Stable Diffusion. This includes:
- The Ai Art — AI, Art, and everything in between (the-ai-art.com) — I like this one — it has a drop down to get specific info like artists or styles. It has examples for MidJourney and Stable Diffusion (and, yeah, I’d say MidJourney has a certain visual richness that Stable Diffusion doesn’t).
- Stable Diffusion Artist Style Studies | Gallery View (notion.site) — more artist examples generated via Stable Diffusion. Well. This is actually pretty vast. I haven’t gotten past the “C”s yet. Dang.
- Stable Diffusion Modifier Studies | Gallery View (notion.site) — what do the various modifiers modify? E.g., cyanotype, HQ, infrared, linocut, tilt-shift, etc etc. Very useful. I have “glitch art” on my shortlist.
- Comparison of all Artists in Stable Diffusion (urania.ai) — generated artists examples, ordered by their frequency of their representation (?).
- Still my favorite: list of artists for SD v1.4 A-I/J-Z (rentry.org) — this has gotten so big it’s more than one page now.
When it comes to artists, I’ve discovered some have a very strong style that gets applied and others seem weaker. Even ignorable. If the prompt’s subject source material has its own consistent strong style, then some artist’s styles just won’t have impact, and others hit like a hammer. The other night, I wanted to create various artist styles of Death from The Sandman. It ended up fixed on “deathrock” images and SD was not conforming to half the artists I tried until I relaxed the prompt some to disengage from “deathrock” images.
My poking of the stick has me conjecturing that, within the Stable Diffusion model and training, some artists matter more than others, perhaps in congruence with the subject matter of the prompt. And if an artist does nothing but portraits of ladies and you just ask for a landscape, well, plop, there’s probably going to be a lady in it.
Talking About Some Stable Diffusion / AI Art
Some of the more interesting follow-up discussions I’ve seen recently:
- AI Art Is Here and the World Is Already Different (nymag.com) — interesting round up. “This flood of machine-generated media has already altered the discourse around AI for the better, probably, though it couldn’t have been much worse.”
- Moar worms: AI-generated images open multiple cans of worms (axios.com)
- University of Oxford: Reflective Equilibrium in a Turbulent Lake: AI Generated Art and The Future of Artists | Practical Ethics (ox.ac.uk)
- This US political site has started using AI to illustrate its articles — The Verge — interesting to see how it spreads (and how it doesn’t).
- ‘A.I. Should Exclude Living Artists From Its Database,’ Says One Painter Whose Works Were Used to Fuel Image Generators | Artnet News — screaming “Greg Rutkowski!” is the new Leeroy Jenkins.
- Daniel Eckler 👨🚀 on Twitter: “Stable Diffusion is only 30 days old… a MEGA THREAD 🧵 on its rapid rise.” / Twitter — interesting Twitter thread compiling stuff done in the last 30 days. I especially liked the “Age of Illusion” video: DIE ANTWOORD — AGE OF ILLUSION (Ai VIDEO)
- “No Irish — I mean — No AI” — Getty Images bans AI-generated content over fears of legal challenges — The Verge
- Damn, there are a bunch of visual models out there — Model Database — Upscale Wiki. — buuut how much do you trust running a random model on your machine?
- Bias — yeah, that hate meat is already fueling ripping DALL-E 2 a new one. Suddenly it appeared DALL-E 2 was now more inclusive in its results. Was it just a front-end hack to append inclusivity to a user’s prompt? Folks are trying to figure that out.
- Da-whoops! Artist finds private medical record photos in popular AI training data set | Ars Technica — if it can be crawled it might be in an AI model. Oy. Inadvertent privacy breaches now in one convenient location.
