This is an attempt to train a new concept using SDXL. A 3000-step lora was trained at rank 2 and alpha 0.5 in an attempt to minimise style transfer, yet the model requires very high stregth to work (>0.9), taking a hit in quality Model was trained on a single caption, "a person is blowing a large balloon". Further experimentation using other captioning methods and different rank/alpha values are underway.

The model cannot reproduce the concept accurately Due to high lora strength, even at rank 2, style transfer occurs.

Model should be loaded at clip skip 1 and stregth above 0.8 on most checkpoints.