Basically OP says that overfitting is a common pitfall, you don't want to overtrain the model, because then everything will look like your training data, and vice versa with not enough training steps. So it's a bit of a balance. If you search for "dreambooth" on the SD subreddit, you will see a lot of examples of dreambooth results and also some that show overfitted and underfitted results. https://www.reddit.com/r/StableDiffusion/search?q=dreambooth...