I am not sure the reason; I guess it's because they pruned some data from SDXL 0.9 and then released SDXL 1.0, I used the same dataset, same network Rank, same network alpha to train the same character Lora with SDXL 0.9 and SDXL 1.0, and the result is that the SDXL 1.0 lora generated more bad ??????? images. Here is another file:

https://civitai.com/models/117324/character-ganyu-genshin-impactsdxl10

(the sample images are picked images with less problems)