Bad LoRAs, trained on less than 20 image dataset