I wanted to generate Kirie on Cottonoob v2 but none of the existing loras seemed to do her justice, so I attempted my own lora. I am not an expert, and this was more of a proof of concept for myself than anything else.

V2:
So, I tried to improve the model from v1 by redoing the entire dataset using controlnet reference, and adding "orange hair, auburn hair" in the negatives. It did seem to do quite a bit of good towards making it look less genetic anime character, and more like Kirie, though i still am struggling with her skin tone. I've added a number of new tags (though i messed up one, dang it). I've added angle tags ("from above, from below, from front, from back") which seems to give the model more flexibility, but without the tags makes slightly more boring images. Lastly, due to the new dataset having backgrounds, "white background" is insanely harder to achieve, sorry. All in all, a step forward, but im unsure if I'm entirely happy with it yet.

Process/Credits:
I used the illustrious Lora from JudgeNaegi to create a high quality 60 (not a specific number, just what i ended up with) image dataset of Kirie with multiple angles, and clothing. I then used the civitAI Character Lora maker to autotag the images. (i used her name in Japanese order as the Trigger Word, and slightly finetuned a few tags to better reflect the content. Something I saw the autotag didnt do and im wondering if i should retrain is "from above", "from below", "from front". I don't know if these angles matter in the end for the tagging dataset. Comment if you think they do.) ~~Lastly, I asked grok to deepsearch the most optimal parameters for training a character lora on civit. I noticed it pulled up a number of guides created by lora makers, so thank you for your write-ups~~!

Funny thing, while I was writing this up, I thought this lora was trained for NoobAI and not illustrious, but I somehow mistaken Cottonoob v2 for a Noob checkpoint. Turns out its trained on Noob, but it's still illustrious. This sort of complicates things, since my parameters were tailored for Noob not illustrious, yet somehow this lora works extremely well. I'm not sure what to make out of all this, but I'm pleased with the results on Cottonoob at least. (You might also understand the excitement i had thinking i made a cool Noob lora, only to realize it totally wasn't. Ah, oh well. Maybe next time.) Now, does this work on Noob? Yeah, it's usable, but to much lesser accuracy.

Some tidbits:

Hair: The model isn't perfect, unfortunately when "facing side" or "looking to the side" she faces from the right. The left side isn't suppose to have her hairclip, but there isn't any useful images I could generate from that angle to make it work, sadly.
Clothing: I included my best attempt at her cat outfit, but it's just not working as I intended it. It's usable, but not perfect. Also, I couldn't get the graphic design on her black shirt, so I left it as solid black instead. (this took a while, since the checkpoint WANTED to give her some sort of inaccurate design) Her black shirt tends to be tucked in more often than i want it to be. And lastly, sometimes the sweater around her waist is the wrong color. Really difficult to get this one right.
Overall there are some oddities, and i obviously cherrypicked my best generations, but it works with some patience.

Clothing tags (adjust weights as needed):

school uniform, serafuku, red neckerchief, long sleeves, red neckchief with blue tip (v2 only)
school uniform, white collared shirt, red necktie, grey skirt, white socks)), shirt tucked in, light blue clothes around waist, light sweater (v2 only)
naked towel, green towel, covered navel
black t-shirt, denim skirt, loose shirt, shirt over skirt
one-piece swimsuit, school swimsuit
orange bikini, navel
light purple kimono, purple obi,
pink bra, pink panties
neck bell, hairband, red fake animal ears, bell, cat ears, red leotard, gloves, covered navel, red gloves, collar, pink fur trim, covered navel, (cleavage:0.8), fur trim leotard (v2 only)

Weight: 0.8

Upscaler isn't recommended.

I'm considering making a version 2 in the future, but I'm out of buzz for now, XD. so this will have to do for now. Enjoy~

Search AI Models & Addons

very few results