Daniel Solisさんのイラストまとめ

Ponder these orbs, my friends.

Occasionally the outputs look like a wanted poster searching for evidence of a cryptid, like #4 here.

#1 is another intersection of tripodal body plan, armor plates, and long nose. I like to think the tripodal animals all come from the same planet. Some evolutionary pressure favored three legs and long prehensile snouts.

Perhaps 1 in every 50 outputs were completely alien, like #1. I think the preponderance of legs, trunks, and tails in the training data confused the AI, so it just made these tentacled horrors. Occasionally it was just a long snoot, like #3.

Most of the training images were side views, hence the recurrence of that same composition so often. However, I did have some bats, which are illustrated face-front, wings spread. That silhouette seems to have influenced a few of the outputs.

I did notice a variant of the "Giraffe Problem." In my case, it was an Elephant Problem. A lot of the training data were elephants, rhinos, and hippos. So the AI just kept making elephants.

But... I must admit, the source data is really skewing how charismatic the outputs can be. Most of the time, it's trying to draw "realistic" versions of medieval illuminated marginalia. I'm surprised I didn't see a goat farting fire.

One thing the Mammal data set has over the Bird data set is the chance for "cute." Very rarely, there is a fuzzy little smiling friend. Most of the time, it's just a random quadraped though.

Mammals just don't have a lot of color variety, at least in the source data. There are so many voles, shrews, pachyderms, etc. It tends to make the outputs pretty bland in their coloring. The variety is more in where the head goes and how big it is.

🧵 These Mammals Do Not Exist

I trained AI on 200+ public domain scientific illustrations.

CREDITS
Source: Biodiversity Heritage Library https://t.co/CyR22DbQbi

AI: LookingGlassAI v.1.1 https://t.co/nRHfga0NJ9

Previously: These Birds Do Not Exist https://t.co/NuQFy9dfWG