Adeline is the 40th release in a series (RLY Thot Shots) of female character LoRAs.
Adeline is a German librarian currently living in Berlin. Her co-workers think she's a quiet, shy super nerd. They're only partially correct. She does indeed claim supreme nerdity--she has an MS in library and information science and has essentially memorized The Silmarillion. Her supernerd-ness is unquestionable, but she is not shy. Her girlfriends affectionately refer to her as the "Reaper of Berlin." She straight up snatches souls (figuratively). On a night out, she attracts all the attention and pulls whomever she wants, only to drop them the next morning. Seriously, fear The Reaper.
Getting started (basic) workflow here:
To share your own art and discuss all things (sfw and nsfw) realistic AI image/music/video generation, join myself and other model creators and enthusiasts on Unstable AI.sland: https://discord.gg/B6BSCbdAJX
Description
Re-trained on Z-Image. Strength ~1.0 for ZiB, ~1.4 (1.35 best) for ZiT.
FAQ
Comments (6)
First, thank you very much for uploading the ZIB/ZIT version! Adeline looks pretty good and I'm really looking forward to ZIB/ZIT versions of your other new ladies, too! One thing I need to ask about Adeline: Her face looks different compared to the Krea2 version. Her distinct facial features from the Krea2 version got a bit lost in the ZIB/ZIT version (or at least it seems so). For instance, her overall face and her nose are significantly rounder in the ZIB/ZIT version and also her eye area is slightly different when I compare your images of both versions. What's the reason for that and how can this be solved to make her look like the Krea2 version? Does it depend on the used checkpoint or on the lora weight or both?
Yeah, it's frustrating, but that's natural. Each base model interprets input differently, and despite the same training set, learns details differently and will have a slightly different output than other models. There isn't really much you can do about it. One option, which I haven't tried yet, is taking some of the output from one base (e.g. Krea 2) and adding it to the training set for another base (e.g. Z-Image). This might help resolve some of the macro detail issues.
From your side, try playing with the strength of the LoRA or specifically prompting what you're looking for in terms of detail (e.g. "narrow face" etc.)
@RaymondLuxuryYacht Yes, maybe this is an option to try. Including some of the Krea2 output images in the training of the Z-Image model. I'm very excited if that would help resolve the issue. You could try it with Adeline or with the other new ladies that don't have a Z-Image version yet. Let me know if it's successful. In the meantime, I will try to play around with strength and certain prompts to get her face closer to the Krea2 version.
@RaymondLuxuryYacht No matter what I try, I cannot reproduce her gamine, elfish face from the Krea2 version. No matter what checkpoint and no matter which prompts I use. The LoRA is too strong. But when I reduce the strength too much I completely lose her likeness. So, the only remaining option is to directly use or include her Krea2 output images in the training of the Z-Image version.
@terror_from_beyond122 Added some images from the Krea 2 version and re-trained. It's a bit closer now. Still not perfect though. I'll upload it in a bit.
That thot is really hot!