Makes things in the style of the 90s visual novel Divi-dead, by C's Ware.

V1 Illustrious version notes:

Re-tagged most images to be more stable. Understands the character names as combined words (may need a little extra help with hair, but as a starting point):

harukahojo azusayuuki yukoanekohji mayumihimuro sachikoseino mikiseino mamiokui akihyuga kimihitosatake yutaito shotaroshijo sano takashinishizaki

The scenery is now included so that background art feels more correct, and each scene except some interiors where it was vague was tagged with one of:

• daytime (most common)

• sunset, red clouds (evening scenes)

• night, glowing lights (night scenes with lights on, may also use lit windows)

• night, dark (night scenes with no lighting)

• night, moonlight (a few scenes near the end of the game IIRC)

Depending on the scene's art. As a result, it does a pretty okay job of getting the time-of-day mood. I couldn't get it to understand the lit windows, but hey, lit windows seems to work in combination with night, glowing lights so it may be an intuitive tag for Illustrious. It usually tries to add people to scenery,no humans may be used to stop that.

Fun discovery: When I was tagging characters, I made an intermediate version but it over-obsessed on curled hair to the point where every character was cosplaying as Mayumi Himuro. Instead, I found that excluding things like long hair or blonde hair etc. from the portrait images actually made it learn what hairstyles each character had. In a programming sense, it's like destructuring what you don't care about, then the custom term you set will do a union spread of the remainder across related images. Interesting! As a result, it understands that eg. Azusa Yuuki has long reddish hair, or Sachiko Seino has short bob cut black/blue hair. By not specifying those hairstyles, it associated those as character traits. It may still need a bit of a hint from time to time, but it gets the gist more often than not.

Oh, if you're using any of the male character tags you may need to add 1boy due to the inherent bias of Illustrious (and humanity in general).

Clothing is a bit trickier. It seems obsessed with Aki's cleavage, so you might need to work a bit to get it to behave. Azusa's dress is best described as orange sundress, white straps even though it was tagged as an orange dress otherwise it'll look more like the uniform. It'll probably add a bow regardless.

I also removed a few of the omake images that I felt might have been messing with the training, but those are pretty inconsequential. This seems more like a good v1 now. It could still use a bit of work, but what doesn't? There's a few more images and probably some tagging learning I could add to it if I did another version.

Alpha version notes:

It feels a bit over-baked so you may want to back the strength off a little. It doesn't understand the characters names as I haven't tagged them, but you can give it the gist if you describe them and it'll probably get close.

The source art was really small and hard to work with, but I wanted to try something interesting. If I have time I'll come back and tidy it up and see if I can get it to be a little more flexible.