The data set was only 5 image at 1024x1024. The other forms the player can take all seem to be palette swaps - which would do more harm than good. So naturally this LoRA is bit stiff. Adding weight often fixes any resistance you encounter while prompting.
Trained at 768x768 at 500 steps.
Weights of 0.5 - 1.0 depending on the use case. 1.0 will make a three quarters view analog of a "gargoyle". 0.5 to .75 seems best for free prompting.