LHC-anon here, just coming off a 24 hour training run that ended up not going anywhere. LR too high, which I should have seen coming to be fair. Not that surprising desu. Lowered LR a bit and increased Batchsize and gonna try again. I have passed the 100 chuubas, added two artists that are a bit weak on noob and spent some time writing scripts for a bit more efficient dataset preparation. All jpegs have been deartifacted as well, so while not png quality, they are at least less compressed.
Decided to do a small lora to get some quick dopamine and did a small beribug lora. Let it train for 1500 steps, though it was already overcooked by step 750 imo, which is crazy fast. Even the 250 step one is already almost there.
Lora isn't particularily flexible, I must say, but still surprised that it is getting most of the small details already. 4 arms is a bit of a struggle, mostly because it tends to mess up hands in some way, see:
https://files.catbox.moe/bcjmm5.pngLora:
https://files.catbox.moe/gjeffi.safetensorsreally only "beri \(beribug\)" is needed, but putting "4 arms" in positive and "6 arms" in negative can help get them more consistent. Later epochs are really good at this, but also suffer from 1 pose syndrome.
Lora is trained on vpred 0.6.
Hopefully the next run is more successful.
PS: Haven't had the chance to play with IP adapter yet, how good is at "transfering" chuubas, so to speak? So character knowledge?
>>90805359Really comfy until the last one. Pretty impressive how well it got it though.
>>90801237Hefty jugs. Nice work
>>90800898>>90801055Cute!
>>90809267Catbox allows registered users to create galleries and is usually the goto of most anons as far as I know.
