“Don’t tell Kiwi-chan!”

“Best side-quest ever!”

“Come on, at least buy me dinner first!”

“Let’s hit the goblin cave next!”

“Note to self: chatgpt is a lousy spell checker.”

“Can you believe we almost skipped this room?”

“It was definitely an N word.”

“I can’t believe we get paid for this!”

“I thought you could Detect Traps!”

“The spell finally worked!”

“Really could use a Pokeball right now.”

“Can you believe the Baron thinks this is punishment?”

“The warnings come after the spells. I see that now.”

“We’re getting so much XP today!”

“Why does this keep happening to us?!?”

“Ohmygod-don’t-stop-it-tickles!”

“Yeah, shoulda swiped left.”

“What happens in the dungeon…”

“I’m gonna be so late for class.”

“I’ve done this quest six times!”

“Oh, god, I’m never getting my bike back.”

“Worst. Homework. Ever.”

“I love this dungeon!”

Just dicking around…

“Um, ‘Trans Magia’ never did that before!”

Robo-futa raising kittens

“So, what is AI good for?”

Bonus round! Picture-in-picture

This took me 2.5 hours to generate. Not that I didn’t enjoy most of the intermediate results, but there were a few tricky bits to resolve:

1. Getting her shorts off: her outfit is iconic, and as a result it’s baked into the training data to the point that just adding “denim shorts around ankles” gave her a second pair. I had to include “((naked))” for a clean reset, add her standard costume, then use “((bare ass))” after to reduce (but not eliminate) the chance that she’d still have shorts on. I also had to leave out the suspenders when dressing her, because they expected to attach to something, creating a small denim belt around her waist. Oh, and I had to add “legs together” to keep the shorts from being distorted or split in half.

2. Getting a picture of mini-Misty onto the wall, wearing her little red bikini from Beauty And The Beach: this started out with a happy accident when I first had her bent over a doctor’s desk, and the engine generated a framed photo of her in a different pose. My mental picture of the scene had her being (coughcough) examined by her doctor, who had been after her since she was a disturbingly-sexy little girl. TL/DR, there are two ways to reliably generate multiple people with SwarmUI: specify number of subjects and suggest positions for them in the scene, or use the new segment feature shown below. Both are very finicky and often require setting a different aspect ratio to make room.

3. And, yes, I had to do a naked reset for mini-Misty as well, even though I didn’t want anything showing. It kept trying to composite the bikini on top of her usual outfit, even though it wasn’t mentioned and the segment prompt is nominally independent.

I tried varying the two parameters to segment, but they did not make the results more predictable, so I ended up going with the defaults and hitting the button a lot.

The incantation method I used for most of the multi-girl pictures was:

((six girls)), girl on left and girl in middle and girl on right and girl above and girl below and girl in front. pose from behind or on all fours or bent over or facing away or kneeling or standing or dynamic. Teresa Holmes and Brandy Frost and Ann Fields and Misty and Dawn.

The number of girls actually generated, their positions, and their styles made it clear that the precise instructions above are only mild hints. It was good at picking from the suggested poses, though.