Seraphina Nightshade is about to set this party on fire!
The name "Seraphina" has a Hebrew origin and loosely translated to "fiery one"
Please checkout my other OCs: https://civarchive.com/tag/citron%20oc
Plot updates:
Unfortunately Sera has been partying a little too hard recently... she might have burned down a building... or two!
She is in debt and needs to fallback on her old career. Luckily she still has contacts in the underworld and plenty of bullets.
ZIT/IL version:
What if I told you I carefully selected my dataset to make sure her butt had the right shape?
The Illustrious version has more real photos in it than previous versions so it very nicely and consistently makes realistic images when on a realistic checkpoint. The Lora output face looks nothing like the training data. This character is entirely unlike anyone other than my OC Seraphina Nightshade.
Old stuff
This was an experiment to see if I could use random Halloween cosplay pictures and AI generated anime pictures to create an OC.
I think in retrospect, I would have giving her a different outfit. Her outfit wasnt this extreme in my AI generated training data.
I let ChatGPT name her so I might change her name later.
Things I learned in this project
Yes you can mix real pictures with anime to make an OC
A hand looks very contorted when its on fire lol
Fishnets are hard to get good training data for.
Training an Lora to make good Green Fire is easy!
70% of the images were real photos and these seems to have resulted in a baked in art style. This was not very good at working with art style Loras. -- Needs more research
Description
Seraphina, long hair, green eyes, black hair, bangs,
fishnets, long sleeves, black jacket, green bikini, navel, thighhighs, short shorts, open clothes,
<lora:SeraphinaNightshade_ZIT:1.0>,
FAQ
Comments (3)
I just really really like this look. Dunno what it is, but it speaks to me. Maybe I'm weird.
(btw, your showcase images shows me ZIT doesn't know bondage. At all. Interesting -- I thought it's a also-nsfw model, right? so isn't bondage knowledge basic, like it was for ILL and Pony? curiouser and curiouser)
LOL SAME. I guess its the edgy/goth appeal. Unless there is something specific about Sera that you like? Maybe its the green and black combo?
ZIT is technically uncensored but if you try to generate something NSFW it doesn't know what NSFW details look like. FLUX will not create nipples, ZIT will make nipples but it doesn't know what they are so they end up weird.
Ironically its so uncensored that if you say "large breasts" sometimes it will try to create nipples even if you don't want them. lol For example I had to re-prompt this several times to hide her nipples https://civitai.com/images/115250006. I could be wrong but my understanding is that ZIT doesn't listen to the negative prompt so there is no way to "remove" things from a generation.
Anyway for bondage and other things like it, my guess is that ZIT basically didn't have lewd content in its dataset. It has uncensored data but nothing that focused on sexy stuff so it never learned lewd things.
Disclaimer: these are just my observations. I'm not an expert on ZIT and I haven't read any of the documentation so I could be wrong.
@CitronLegacy Heh, good to know I'm not alone in my Sera-stanning.
Yeah, I do love goths (Marnie!) so that's obviously the main thing going on, but I think smth about the specific gothiness of her look is my cup of tea. Prob the combo of outfit and blackhaired blacklipped vixenness (maybe also the green fire? I love mages, but I dunno). You just hit the "perfect sexy gothiness" nail on the head, hard to know how.
Thx so much for these details about ZIT!! -- that's a major key to figuring out prompting it, and i hadn't realized yet. (even tho u say u don't know for sure, it does sound correct).
P.S.: speaking of ZIT's knowledge, I experimented with ZIT's Sera a bit and posted my gen of a bank robbery for 2 reasons:
she's a badass super-sexy roberrette ^_^ (thiefette?), and also:
I recently realized that NO ai models can do "aiming gun at ___". Danbooru knows "at viewer/ at self", but AI is so atrocious as spatial recognition (AFAICT?) even the big-corp massive LLM-based models like GPT (and flux) can't do a simple "aiming at person". At all!
And now my gen shows me ZIT has same issue.
So I wanted to ask if u have any clue if it's physically possible to train a lora for that? my guess is no, but if you think yes I'll add it to top of pile of loras I plan on creating one day (soon? maybe! new year resolutions ^_^).