testing 16 dim 8 alpha lora trained for 2'000 steps (20 images, 10 repetitions, 10 epochs)