Karin Aoi - DNA²

I present you Karin Aoi from DNA².

Ignore the following rant:

*Foams from the mouth and then from every other orifice* Well... stuff happened. There's almost zero fanart of Karin. This is trained on the original manga and the anime. Oh god the anime... The blue-ray release seems to either be using a laserdisc or tape master, distortions and all. If was as if someone put a vhs on a dumb upscaler. I had to do a screencapping run, denoise and then img2img the captures to be somewhat useful. The manga in some ways was worse. All copies in manga sites have scanning artifacts, literally you can see where the pages were folded and most of them are 378x600(minuscule). Had to scourge the net until i found a chinese translated copy with a slightly higher resolution and better quality. Kudos to that Chinese guy who seemingly sacrificed a Tankobon to the scanner gods. May you rest in peace innocent tankobon. Still the images were still crap though not as horrible as the manga sites(Makes me weep for my own physical copy that is in the hands of a buddy but has not expected return date.) So I had to experiment for a while for a good jpg denoiser and manga upscaler. So while everything in the dataset is canon I don't think i have a single original image left. Btw the results of my sweat and tears have been codified into my guide so check if you need to upscale manga.

Seems to work fine-ish with: <lora:aoi_karinV1:.9> aoi_karin

It works fine at lower weights but as i mentioned her outfit is slightly undetrained.

For outfits i only isolated 1 and a variant:

• Purple_bodysuit_White_Pauldrons_White_breastplate_Asymmetrical_legwear_White_thighhighs_hood_down: her normal outfit.

• Purple_bodysuit_White_Pauldrons_White_breastplate_Asymmetrical_legwear_White_thighhighs_hood_down_Purple_Cape: her normal outfit with her cape

There are not enough images of her fully armored self but it is in the dataset tagged power_armor. Her hoodie/helmet up version is even more scarce so i didn't even bother.

The LORA came a bit undercooked outfitwise, but it is serviceable. I might retrain it in the future pruning the dataset a bit as it ended up rather large(324 images) due to me refusing to discard any the damn images i basically made.

I would upload the dataset but even if cleaned up, upscaled and img2img-ed, the images are still very close to scans and screencaps, so i will just add the tag summary.