A bit late to the party but here's a realism style lora for Z-Image-Turbo(ZIT).

I have been training this lora for quite a while but after several iterations i arrived at this lora which i believe is capable of producing realistic composition, textures and people instead of the default model perfect looking people.
Note that ZIT by default is very capable of producing very realistic images and this lora might hinder some of the fine textures but for the most part i believe the final image is much more in align with what a normal real life image would look like.

But anyway TLDR; at a strength between 0.8 - 0.9 for most prompts this lora should produce better realistic images than default ZIT.

I highly recommend using Euler sampler with ays scheduler if you are using comfyui for best results.