v1 - was too overcooked and destructive
v2 - This lora trained on the same dataset but with rank 32, low noise bias and 1.5K steps only.
Resulting in much better preservation of z image inherent capabilities
在大量我喜欢的二次元动漫图像上进行训练。
排名 128
步数 15000
很可能过头了。