Dataset quality matters more than settings for character LoRA training

For character LoRA training, using ~60 images with ~70% face shots yields better results than tweaking settings. Going beyond 100 images often leads to overfitting, restricting poses to those in the dataset.