This update focuses on improving the overall image quality and enhancing recognition of Videl’s hairstyle variations (DBZ / DBGT). Unfortunately, I lost the original dataset after formatting my PC, so I didn’t rewrite all the DBZ / DBGT tags for every image. From now on, simply describing the hairstyle in the prompt should be enough to get the desired version. I’m currently experimenting with adding an absorption concept to the LoRA . However, due to the limited amount of high-quality training data, the results remain quite unstable and inconsistent.