Ymir from Queen's Blade Rebellion Version. The Iron Strategist
Notable tags: aqua eyes, blonde, twintails, pointy ears, drill tails, black dress, hat, suspenders.
Notable traits: due to the nature of the costume the ai trends to show up her chest so make sure to add "nsfw" and "nipples" to negative prompts if undersired.
Description
FAQ
Comments (5)
I really hope this remains up; but I saw the Ophis (HSDXD) LoRA get taken down with a similar level of exposure.
If you don't mind, could you retrain this LORA with the same method as the one you used for Sialeeds? That worked way better for the checkpoint I'm using than this Ymir LORA, and I really want to reproduce this with that model.
Thanks for your efforts either way.
I use exactly the same method for all of them since I'm not really that adept at making loras. I'm aware the performance depending on the model may change, but honestly Im not on a point yet where I could guess whats happening on certain models and lot less how to solve.
All I can tell you I used the same process than with Sialeeds, if there is any difference in performance in a particular model, I have not the slightest idea why. I wouldn't be able to train it in a different way because I only know one way.
May be in the future if I learn more and understand more. but right now...
@neclordx Then I'm guessing the reason for why Sialeeds responds better to the checkpoint I'm using is that it has a better dataset than this Ymir one.
I completely understand if you don't want to, but would you mind providing the dataset you used for making this Lora? In exchange, I can give you the Kohya settings I'll use to try to make a more flexible lora. Again, I completely understand if you don't want to share since gathering the dataset is the most work in making a Lora.
Also, since you seem to be focusing on making loras of characters without many images out there, here's a Lora training guide that deals specifically with that. https://rentry.org/59xed3
It even has a section on training a lora using a single reference image. I hope this helps you at least.
@neclordx For transparency's sake, these are the settings I was going to use for my attempt. https://pastebin.com/hjCZvSJ4. In the other loras I trained, the best result wasn't usually the full 10 epochs but one in the middle, in one case, even epoch 1.
I believe you can copy that to a .json text file and load the settings in kohya_ss, if that is the lora training software you're using.