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 <a target="_blank" rel="ugc" href="https://github.com/erosDiffusion/ComfyUI-EulerDiscreteScheduler">FlowMatch Euler Discrete Scheduler</a> 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.<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/7b3f4aa2-87e4-4cfb-9879-044edb1a971b/width=525/7b3f4aa2-87e4-4cfb-9879-044edb1a971b.jpeg" />

Windtunnel Face

Files with SHA256: 89ad907fdf40e3fed9ff4546f9f0ad5b2dd89da0c34dba19a1b638785da0e486