I really enjoyed the style of images ChatGPT would produce for a fictional film, so I trained a Z-Image Turbo LoRa on some ChatGPT generated images with prompts provided by ChatGPT itself. It did pretty much exactly what I wanted it to do.

I use it around 0.5-0.9 strength and I'd also play with samplers and schedulers as they have a fairly significant impact on final product. I think dpmpp and beta has worked the best for me, but it's also much slower than other samplers.