tiny experimental run on MOSS-Image Base. Low learning rate (2e-5) is much better than higher lr. Never use 1e-4.Maybe I need better caption model to improve the training dataset but idk.The dataset doesnt contains kemono bara things but you can try it. It can not identify artist's name tag right now so just use it as a normal style lora. <s>I'm sure it knows artist styles but you can not use tag to control its style output.</s>

Files with SHA256: 813f3e04c7e3d6cfa2f81548d5665a564fee9adc6fad02b29f68ff36fde0aee0