This is my second trained model. I haven't figured out how to train LORA yet. Basically everything is using the default preset. So I'm still a novice. I found that if using the 1.5 base model for training, the resulting LORA doesn't seem to match any CHECKPOINT, especially the commonly used CHILLOUTMIX. The characters become very childish. So I trained this LORA with CHILLOUTMIX. The effect of using it to generate images is better. Personally, I prefer more realistic characters instead of cosmetic faces. 

It may be because the pictures I used for training were 512*512. When generating images with 512*512, the images will be slightly distorted. Using 768*768 will improve a lot. To generate a half-length image, at least 768*1024 is required to have a better effect