<h3 id="makes-things-in-the-style-of-the-90s-visual-novel-divi-dead-by-c's-ware.-5l95h9091">Makes things in the style of the 90s visual novel Divi-dead, by C's Ware.</h3>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):<code>harukahojo azusayuuki yukoanekohji mayumihimuro sachikoseino mikiseino mamiokui akihyuga kimihitosatake yutaito shotaroshijo sano takashinishizaki</code>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:<ul><li><code>daytime</code> (most common)</li><li><code>sunset, red clouds</code> (evening scenes)</li><li><code>night, glowing lights</code> (night scenes with lights on, may also use <code>lit windows</code>)</li><li><code>night, dark</code> (night scenes with no lighting)</li><li><code>night, moonlight</code> (a few scenes near the end of the game IIRC)</li></ul>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, <code>lit windows</code> seems to work in combination with <code>night, glowing lights</code> so it may be an intuitive tag for Illustrious. It usually tries to add people to scenery,<code>no humans</code> 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 <code>long hair</code> or <code>blonde hair</code> 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 <code>1boy</code> 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 <code>orange sundress, white straps</code> even though it was tagged as an <code>orange dress</code> 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.<hr />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.

Divi-dead

Minor tweak. Added a few missing character pictures, added <code>lit windows</code> to some night scenes although it seems to have minimal effect.

Files with SHA256: 513133cd5554aa4931e2e153a939cb38fd41684fa0e1accb3c22b9a94ab8eda7