Starting a bunch of high photo count trained models that are as flexible as possible in creations, while keeping the look of the character as close to the source as possible. If you have feedback or recommendations for improvements, I'm more than happy to receive your feedback. Also happy for private messages if you don't want to post publicly, I just want to make sure the community is happy with these models.
If you really like a model but there's something you dislike, I'm happy to re-train it from a lower step starting point into something you'd specifically like to see. On Civitai I've noticed many characters model LORAs are very rigid or have poor stability between images where they never look like the same person twice, and I'm hoping to dedicate enough time and energy into creating qualitative character LORAs that actually look good.
Don't be freaked out by the amount of training steps, as I ensure the models will never be rigid, and take painstaking effort to dynamically train them to increase creativity and prompt adherence. I constantly re-work Learning Rates, decay, mask/unmaked rates, etc, I do this at every few convergence epochs to make sure they never stray from the source material. All of my models are trained painstakingly at Batch Size = 1, full 1024 res training, and every other option to slow the training while increasing the quality... Because I love you guys, and love higher electricity bills even more! Try them for yourself. I test each model with as many LORAs as possible to ensure it maintains character dominance in genning, as can be seen by the included images. I've seen too many character models where if you set ANY other LORA to slightly over 0.000000001, then the woman turns into a monster.
The first batch of models are all Asian models, and they all share the same trigger prompt;
Asian woman
*You can use these models in whatever way you want, as I hold no liability for their usage and am merely creating these for fun.
*Model is completely fictional, and in no way representative of a real person. Any similarities are purely coincidence.
Description
R16 lower file size model added

















