The new MOSS Image version is my first LoRA trained on that model. I used the default settings in AI Toolkit and it seems to have worked ok. Thanks to the advice of @EnragedAntelope I'm using dpmpp_2m_sde and beta for these and it's looking pretty ok so far.

NOTE: The Z-Image Turbo version doesn't seem to play nice with character LoRAs. It can distort the likeness and introduce errors/deformations in the scene. Try using more (e.g., 18) steps and increasing the model shift (e.g., 10). Also, for less noisy/blotchy images use the FlowMatch Euler Discrete Scheduler along with Euler_ancestral.

If you can't get a closeup photo, make sure you aren't describing things like pants, shoes, excessive background details, etc.