AI Toolkit for Illustrious LoRAs: It just doesn't work!
Submitted two V2s made in AI Toolkit. Had to wait 48 hours to get them because moderation hold. When I finally get them back, they're just not good.
One is just a straight up failure. After Epoch 3 it somehow learned "print pajamas" as "corrupt the entire image" (cover image related). How is that even possible?! I submitted a V2 with some tweaks to see if AI Toolkit is really just super random at producing anything not garbage (I've heard from another creator AI Toolkit is very random in results), but I'm going to be waiting for days to actually get it back.
The other model at least passes the bare minimum functionality test (yay?), but suffers from low face accuracy, and random hallucinations pasted on everything, while just flat out not learning some things clearly present in data. I put 7 clear pictures of the character's winter coat and hat in, and for some reason it's actually less accurate than just calling them by tags alone would be: The coat not only looks completely different and generic, but it's consistently generated as a jacket (doesn't cover pelvis) despite ample full body pictures that show its length and using the tag "coat"! AI Toolkit has somehow made a LoRA less accurate than tags alone!
Both also suffer from random dopplegangers, afterimage, and siamese twins (none of which are in the dataset) in "1girl, solo" generations and randomly having completely mangled, sub-SD1.5 hands (both these models used 480p broadcast tape as source, but the V1s made with even lower quality data on Kohya did not suffer this). I've seen these issues consistently pop up in AI Toolkit LoRAs. These issues were present in the 1 to 1 comparisons I made when Kohya was still working, so its not my dataset that's the issue.
Is there some magic setting for making Illustrious models in AI Toolkit that makes it not horrible? We need Kohya back, because not having a training option that actually works with the site's most popular models (Creator Studio>Analytics>Base Models lists "Civitai-wide base-model usage" and total generations by base model for the day) is just unacceptable.