This lora is to produce oblong/thin/narrow face shape. The dataset is fairly diverse, but most faces of this type have a fitting type of nose, so there’s a bias. Also, I do various edits to hide face features I don’t want to see, these are often unique enough that I don’t usually see them in the results. In the case of this specific lora, you can see those weird shapes appearing when the prompt is not specific enough, so be wary of that. If I ever fell like I should do a version 2.0, it will be trained on unique faces without theses edits.