If I said a quality waifu i would be lying. I''s is my least favorite Masakazu Katsura's Work. I was searching for images either from AI Amano(Videogirl AI) or Karin Aoi (DNA2) rather unsuccessfully until I found a very nice I''s doujin and what can i say I am a weak man.
Thus we got this... Iori is rather plain and the artwork is scarce and varies a lot. The anime style is almost completely different from Katsura's also.
V2
Another very old LORA retrain, as i mentioned in the old comment above, the anime screens were awful and the style didn't match in the slightest so i purged most of it. Other than that i did some more cleaning and re tagged everything.
Seems to work fine with: <lora:yoshizuki_ioriV2:0.7> yoshizuki_iori,
Was only able to isolate a single outfit:
School_uniform_Purple_shirt_Blue_skirt_White_neckerchief_Black_thighhighs: Her school uniform.
There was some more stuff in the anime but the style is extremely generic and doesn't reminds me of Katsura's work. Honestly Seishinja, Kuusou Zikken or redlight(Don't search for them if you have any type of sensibilities) make a much better job that the actual official anime.
Anyway as most of the Anime was purged form the dataset I will add it with the tag summary so people can experiment.
V1
Regardless... Effective wight seems to be between .4 and .7.
Got stable results with: <lora:Yoshizuki_Iori:0.55>, (Yoshizuki_Iori), big_breasts, bangs
*Edit: adding blush to the prompt might help, I was checking the tags and it was rather high. (<lora:Yoshizuki_Iori:0.55>, (Yoshizuki_Iori), big_breasts, bangs, blush)
At .45 it is a coin toss if the lora will accurately represent Katsura's style(got 1 out of 5 I could say "that's Iori" with confidence.). That's the problem with Iori being rather plain. At .65 it is more consistent but the doujin source might start to make things look granulated. Therefore I got best results at .55.
Surprisingly all training data was SFW even if it came from doujins, I just put this at NSFW just to be safe.
Description
FAQ
Comments (3)
this is that one style i love the most. not gonna lie.
can you make an artstyle-lora out of this one and others from the same artist? just asking...
Unlikely as i would need to clean a lot of doujins and manga. The anime are mostly a no go as they don't respect Katsura's style. And well cleaning up sfx and speech bubbles is a pain. I think this one and the Kazami Yuuka from touhou are my most doujin/manga intensive ones and oh jeez were they a pain. They also look stupidly better as 2/3 line art for shape + 1/3 anime for coloring in the dataset does wonders.
@knxo okay. no worries from my side. i really respect it. sounds like you went through a lot of trouble to come this far. i still am thankful for this. and no worries, i was just asking out of couriosity and not because "you have to do that for anon".
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.