侠客行水墨风There is no distinction between versions. If you insists, I think dim512 is better. I used the same way to train the model. You can test it yourself. I am still refining it smaller, and the effect is until I feel that I can’t stand it anymore.you can use this config in lora block weight plugin to Extract stylePALETTE:1,0,0,0,0,0,0,0,0,0,0,0.8,1,1,1,1,1This is one of the few loras that I have trained that can maintain a good rendering effect from 512 to 32.It occurred to me that I could let newbies in lora training take a look at the differences between different dim values.You are welcome to test my lora and commit your commentThe loss of these loras are all around 0.11

ink_xiakexing

ink_xiakexing_v3_64.safetensors

Files with SHA256: 1028d3a680bd329787fd7fcbccf995820b055701f74b19c44766fd6c4b3158ab