Train with a dataset of 260 images. It is recommended to use a weight range of 0.7 to 0.9.

The "slight" version is the one that reduces the learning rate, thereby lowering the degree of fitting. When used in combination with LORA, it performs better with a weight of 1.