LORA trigger word is thylane
45 images. 100 epochs. All sample images used it at .8
I had released this as a LYCORIS previously, but I'm learning to use OneTrainer, and this was my first successful test run, so I thought I'd share it. 5e-5 learning rate. Used the same learning rate but 30 epochs for text encoder, and same lr for unet with 100 epochs. adamw8bit, cosine. OneTrainer was FAST. Blew through this in about an hour where Kohya_ss would have taken 3 hours. I think this came out better than the previous LYCORIS, seems to hold the model's likeness a little better, but needs to be used at about .8, lower and it starts to lose it.