LoRA

You only describe what you want removed in a way and leave what you want associated with the token you create. If I'm training a style then caption everything but not the style leaving that with a term like s7yler. The model will know all the other words you say and associate the existing tokens to them but will be left with all this left over information which it will call s7yler. After many images of various examples of s7yler the bettter it can learn. Flux is very fast to train so it really your goal of what you are attempting. If you want to do a face usually 20 good diverse images are good enough, if you want the face and body type then 60-100 is better. The important thing that will mess you up training 1000 images or 10 images is 1-2 images have the same composition. Think of it as if I layer what I want trained on a transparency how much of the same information would I see overlap? You want none. That is what causes over training. So work on varied angles.