I can’t remember the first time I made this lora costume. Some friends in the group found some pictures on ins and wanted me to reproduce lora. The technical level at that time was completely insufficient, and the learning rate was even worse. I don’t understand it at all, so I have been training with the learning rate recommended by the trainer. I only adjusted rep and ope for fun. After these days of study and communication, I have a little understanding of the learning rate. I plan to take out this training set and try it. Overall, it is better than the version I made for the first time. a lot of.
However, the relationship between 1.5 and extremely small fishing nets still cannot be better reflected. At least the overall reproduction of the costumes has come a long way.
Here, share it with you, thank youDescription
I have always used weight 1 for testing before, and everything was normal. However, when I used a certain 2.5D realistic checkpoint, it turned out to be overfitting. Because when adjusting the training set, the facial features are generally covered with white circles, and the appearance of the white circles means fitting. When the weight is 0.9, everything is normal. You can try to jump repeatedly between 0.9 and 1. Ha ha
FAQ
Comments (5)
烧~🌶🌶🌶
Your loras are so beautiful, I hope you make SDXL version of these T_T
chilly :))