/vt/ - Virtual Youtubers

>>73881789
https://twitter.com/eus_ing/
>>73885636
https://pastebin.com/kMG546nb
Besides the scale_weight_norms option, which was added after, I used the first config for my previous Mako lora (which was said to be overfit) and the Moruru lora (1200 steps) that should still be in my mega folder. I used a version of it without min_snr_gamma, noise_offset, or adaptive_noise_scale for an unreleased Tatara Kogasa lora that gets her eyes right more or less 100% of the time (unless turned around or from side is prompted which lowers that).
The second config is my current character lora config that I've been using for every other character lora. It has redundant/later overwritten lines because I copypasted some of CCC anon's config into it. I also use a version with 0.0002/0.0001 unet/te lr that works significantly better for non-limited datasets and characters with more complex details like Mako.
https://mega.nz/folder/dfBQDDBJ#3RLMrU3gZmO6uj167o-YZg/folder/0bp3CLzB
At the moment I'm also using a version of the second config with 16/8 dim/alpha and 0.0002/0.0001 unet/te lr to train style loras. I have those running for 3000 steps which may be completely unnecessary.