#anythingv3

I lucked into the initial generation and thought I could make a Raiden x Sara fanart out of it. Looking back at it now, I see issues I would hopefully fix if I were working on this today, such as the shading and the stiffness of the faces. I think I went a bit too heavy on the harsh high contrast shading for this one, and it was an issue for me for a lot of pictures.

Typing in "Miss Piggy" got me Miss Piggy, exactly how you think a hentai bot would draw her. No, no pig features, just boobs for days. Sometimes she'd be surrounded by partly-melted Miss Piggy puppets, usually not.

But Kermit. Oh boy. Kermit. I'm not even going to comment, just look at this:

Wow.

Also Fozzy is an adorable sad anime girl now:

"Toriel" just gets you sad anime tiddy mommies which, ok, AnythingV3, fair. But other waifus of the Undertale universe generate some good weird:

Look at this Alphyne right here. Like I feel like this is them from some weird JoJo's Bizarre Adventure crossover. Bananas.

Not as weird as what happens when I type metaton:

(Yes, I watched Interspecies Reviewers, and I liked it because I'm a filthy pervert. This blog isn't me making fun of horny people as a non-horny person, this is me making fun of horny people *as a horny person* just so we're clear.)

But these results are. Not what I expected.

Dogs and cats, living together, mass hysteria.

On the left we have what it usually draws, and on the right we have what it skins and dresses her in.

Well. "Dresses" is an overstatement, what it really does is put her in skin-hugging leotards and underwear with fringes and boas made out of this honestly fairly decent anthro fox thing.

Now, full disclosure, I only know "renamon" from the furry fandom so I don't really know what's going on when I type it into SD. Same goes for Gatomon here:

The Nightmare Abacus

Interestingly, you can kind of see that it knows patterns that should be on an abacus, but it gets kinda chaotic in there. Actually it gets really wobbly because from here out, it's only getting crazier.

Prompts like this reveal a specific kind of brittleness in the model: it hasn't been trained on a lot of actual pictures of actual abacuses, especially not the subtraining for anime aesthetic, so you get a lot of the vibes of what an abacus is, but not like, actual pictures of abacuses.

On a hunch I added "4girls" to the prompt and stable diffusion immediately threw out the abacuses entirely in favor of the girls, but it did remember that they were supposed to be the math club.