This was created by using 200+ curated images generated using LexiVision AIO-VNSFW SDXL checkpoint, captioned using Qwen3-VL-32B LLM, and trained for 725 steps (10 epochs) as an experiment in refining the base Z-Image-turbo model's output.
While 0.85 seems to do ok, I'd recommend trying weights as high as 1.1 when using this LoRA.