Support me at https://ko-fi.com/elderagent
Anima LoRa of Meng from Avatar the last Airbender, higher res dataset
Prompts:
mengatla, black hair, brown eyes, hair horns, twin braids, pink hair ties, blunt bangs, pink kimono, purple obi
Optional 'tooth gap'
Description
Comments (8)
Can I ask how you train your Anima loras on the site? I've tried several times and Anima just doesn't do as good of a job at capturing the style and design of a character as Illustrious for me. I've trained on both 1500 steps and 3000+ steps
I think it depends more on the negatives on the prompt and the checkpoint you use, and even then they're still not that consistent
@elderagent
But you can't really set negatives during on-site training for Anima
@yohasakura4110 We can't control the training, but we do control the generation
@elderagent yeah I know, but as far as training settings, what do you use for Anima? For me I've only used the current defaults. And only set the clip skip to 2 and keep tokens at 1
@yohasakura4110 Yeah same, the difference really is on the generation side, otherwise we would have to re-train loras every time there was a small change in the checkpoints
@elderagent Also, how many steps/epochs do you usually use for your Anima loras?
@yohasakura4110 I leave those on default, same as checkpoints or epochs