Similar to my last released ???????????? model, this one should be less damaging to the overall style while also making it easier to achieve large sizes.
Also I improved the accuracy of the tags, and doubled the dataset size to ~400 images.
Increasing LoRA strength is actually useful for achieving larger sizes in combination with the right size tags.
Training Details:
~400 images
160 epocs
learning rate 2e-4
text encoder LR 1e-4
base model Av3
clip skip 2
random flip
tag drop chance 0.15
network dropout 0.25
bucketing at 768
dim 32
alpha 16
225 tokens
cosine with restarts
training with tags, tags attached next to model download
use the new weighted captions + dropout in Kohya that way more important tags were trained at a higher weight (weight of:2).
Kohya LoRA LoCon, does not require any additional extensions to use
