A little unstable, but don't let that discourage you: Very specific held objects are the kind of thing SD based models struggle with (and Illustrious is way better at them than Pony was!), rather than necessarily being your fault.

I do think 18 images is a bit small a dataset though (I recommend 20 as a minimum for LoRAs, and for something relatively complex like this I'd suggest at least mid 30s). If you didn't already know, you can actually use photos as training data if you tag them "realistic, photo (medium)", which should get you a lot more data to use (you can actually use a lot of "weird" mediums if they can be pinned to a recognized tag. "Sketch", and "3D" are the other two big ones to take advantage of, but I've even gotten "pixel art" to work.).