This LoRA seems to have overmemorized from the training data, as you can see from the attached XY plot. Note how each generation gravitates towards the same particular memorized girl/pose despite starting out very differently.

Ideas for a potential V2: Use a much smaller DIM (2 or 4) to reduce the likelyhood of memorization and rebalance the dataset to have more variety (same number total is fine, but cull and replace over-represented angles/poses)