Nathan, you handsome devil. I made you a character LoRA. Figured I'd publish it so I have all the information for it handy.
Taking a hint from others, I just trained it for 500 steps on the minimum number of headshots (2, in this case). Seems to do a decent job.
I don't expect anyone to use this, but you can if you want! I originally got his face from an I2I generation with Creamy Ranch Mix, (don't worry, the face in no way resembles the cowboy I got it from, it was just the hat and composition of the original image that carried over) so it was impossible to reproduce his face from T2I because that strong jawline and fucked up hairstyle just always looked different in T2I, naturally.
Nathan is my alter-ego, and he likes all kinds of women. You can depict him with whatever situation you want, but if it's something truly out of character for him, I might apply a sticker of disapproval. But yeah have at it if you want.
Prompting Guide:
This model gets the face right but you have to prompt a man, just using the trained tags:
n4th4n, c0wb0y_h4t, l34th3r_j4ck3t, pl41dwill have potentially unexpected results, especially with multiple characters.
I do plan on refining this LoRA when I have time, or making it for another model, but for now I just need Krea 2 to reliable produce Nathan and his signature outfit.
Description
Comments (2)
looking at the gallery, specifically the prompting and the image composition, your approach was to take a previous SDXL/Pony/Illust dataset, as is, and retrain with that using a Krea 2 trainer, correct?
.. ah, looks like you provided a dataset. Answers a lot of my previous question but it only has two images - is that the entirety of images used to train this character?





