This is the attempt of a version of the Lora for z-image turbo. Unfortunately the training process failed halfways (at epoch 28 of 60), so it trained only about 1500 steps instead of 3000. I hope it works roughly as intended....