For the 2B model, it's trained on more images at a higher resolution. I've had my best luck with the LoRA strength on 1.01 and cfg set to 6Trained from art by Ke-Ta on various booru sites, due to the questionable content in some of the training data using "nsfw, nude" in the negatives may be desirable.Trained on google collab (thanks to @HollowStrawberry)I've had good results on abyss orange v2 with strength set to 0.9, but you'd do well to experimentNow it's a LoRA, much better than V1 at that, might be more flexible in the models it'll run well on too.

Files with SHA256: cf7eecb6f820a5bebab63f2f6264ecb7921934e026f926f49dad79264fed436c