I’ve been wanting to do this for a while, and I finally got around to doing a Google image search for every name on their list of artist mods and making a brief description of their work. It might come of some use here.
Summary
Alex Hirsch: Gravity Falls
Alphonse Mucha: medieval-ish illustrations
Amanda Sage: fantasy illustrations covered in threads
Ben Bocquelet: Amazing World of Gumball
Bernie Wrightson: b&w comics
Canaletto: realistic paintings of buildings
Caspar David Friedrich: portraits, landscapes
Dan Mumford: bright modern graphic art
Dan Witz: realistic tight crowd paintings
Edward Hopper: 20th century Americana
Ferdinand Knab: buildings, landscapes with moody skies
Gerald Brom: sf/fantasy graphic art
Greg Rutkowski: fantasy art
Guido Borelli: bright colorful buildings
Gustav Klimt: colorful detailed abstracts
Gustave Doré: greyscale fantasy
H.R. Giger: organic horror
J. G. Quintel: cartoonist
James Gurney: dinosaur riders
Jean Tinguely: found object sculptures
Jim Burns: sf graphic art
Josephine Wall: colorful fantasy art
Julia Pott: cartoony animal drawings
Kandinsky: abstract geometric art
Kelly Freas: sf and f illustrations
Leonid Afremov: street scenes with bright colorful lights
Max Ernst: weird paintings
Moebius: sf and f drawings
Pablo Picasso: abstract art
Pendleton Ward: Adventure Time
Pino Daeni: paintings of women and girls
Rafael Santi: classical paintings
Rebecca Sugar: Steven Universe
Roger Dean: colorful stylized fantasy art
Steven Belledin: fantasy and horror illustrations
Steven Hillenberg: Spongebob Squarepants
Studio Ghibli: anime studio
Thomas Kinkade: whatever you call it
Tim Burton: movies
Van Gogh: Van Gogh
Wadim Kashin: colorful abstract sf-ish scenes
Wes Anderson: movies
Zdzisław Beksiński: weird horror
Meanwhile, I tried one of the names from the list I posted, Dan Witz, who does creepy near-photorealistic illustrations of tightly-pack crowds. I didn’t have much hope for it (his work seemed too tightly packed for the AI to grok) but I was blown away by the results. If you ever try images with groups of people/creatures, try sticking his name in there. (Examples later after I finish digesting them.)
Okay, samples. The mushroom photo that I’ve used and shown here a few times? This time it bacame some sort of mushroom god rising like the great pumpkin before a crowd.
Speaking of noise files, I recently tried this comic panel. The details are lost, but it makes images with a pleasing aspect ratio and a nice blue tone. The trees on the two sides frame the images, or become a desk or such visible under the ends of a tattered image. Works well with landscapes.
Tried pasting this into the Github Disco Diffusion (default low/medium settings):
“Evolved” version:
Plausibly it’s using the same, or very similar, code. Some variant of diffusion, anyway. But then is there a way to access the myriad individual settings on Nightcafe…
Okay, I was in a mood for landscapes, but I wanted to try a location that none of had posted here yet. I tried Painted Desert National Park, and it indeed knew to create something painted and deserty. But that isn’t my big discovery of the day. My big discovery was wondering about making a night scene, and simply tried “night sky” and “milky way” as modifiers. It works. (Click the album for more, night and day.)
I tried the new Google Imagensite (promises a more sophisticated parser model than the one Dall-E, diffusion, etc. have been using, and therefore able to understand more complicated prompts, and also it does not work in exactly the same way—first it creates a small image and then upscales), but only cherry-picked examples and no public demo, so same as Dall-E from where we are sitting. Maybe someone will spend the $$$ to train at least a public mini version, but that is the obstacle right now for people who just want to play with these models to see how it works. You can’t run it at home (I can barely run Disco Diffusion without lowering the quality, hacking the code a bit, and waiting half an hour for each image)
Interesting to think we are probably still at the beginning of what these systems can be trained to do, and eventually using the current systems will look like being fascinated with Pong.
