ruDALL-e
Experiments with fine-tuning and sampling from ruDALL-e
Bored Ape Yacht Club
- Initial weights: Malevich - Learning rate: 5e-6 or 1e-6 - Training amount: 1 epoch - Dataset: The 10_000 apes from their website, paired with their description.
As those stupid apes are fairly similar, getting the right learning rate is important at 1e-5 the model basically collapses on the mean monkey (no accessoiraces, brown background). 1e-6 gives fairly unconstrained results (less monkey-y), 5e-6 contrains strongly on the monkey shape but allows stuff to happen.
At 5e-6
lofi
Face like an alien, realistic rendering
Monet Claude painting
anime
hippie free love sunshine
cute crazy monkey
collage
fish eye lens effect
Gorillas
embroidery
Photoshop
Art Nouveau graphics
van gogh painting
flat colors
ink drawing
watercolor
shaggy leather
cyberpunk
cinema poster
vector graphic
detailed oil painting
line art
orange punk psychedelic painting
pixel art
Soviet agitation poster
three-dimensional realistic rendering
graphic novel
Beksinski Zdzislaw picture
electric colors
medieval parchment
psychedelic musical cover
communist socialist youth fighter
At 1e-6
lofi
Failure
anime
collage
fisheye lens effect
embroidery
photoshop
flat colors
ink drawing
watercolor
detailed oil painting
line art
pixel art
Soviet propaganda poster
graphic novel
electric colors
medieval parchment
photorealistic
Asterix & Obelix
I generate a dataset from the Russian versions of the first eight Asterix comic books. See the tutorial on how to automatically split a comic book page into its panels see.
Sickle
























Army
























Gallia
























Obelix
























Romans
























Romans are crazy
























Gladiator
























Magic potion
























звездочка
























борьба
























Naruto
Same as with the Asterix comics. I generated a dataset from around a dozen Russian issues of the Naruto comic books (all black & white). Sadly this fine-tuning did not seem to work very well, so below are just some random outputs.
Naruto









Emojich
The Emojich is only trained on Emojis. Below are some sampling examples. The Asterix one is again (really shortly) fine-tuned on the Asterix panels. The others are just sampled from the initial pre-trained Emojich weights.
Asterix panels
Amsterdam
Napoleon