I retrained the Sprites LoRA with captions preferred by flux and had to reduce the number of training images. I kept the tagging from previous data sets and think it adds good versatility in prompting. Check it out and see what you think.