Because we only have the distilled turbo version of Z-Image loras can be unpredictable, especially when combined, but the good news is in a lot of cases you can get the style you want just by prompting.
Like SDXL, Z-Image is capable of a huge range of styles just by prompting. In fact you can do use the style prompts originally created for SDXL and have most of them work just fine: [twri's sdxl_prompt_styler](https://github.com/twri/sdxl_prompt_styler/tree/main) is an easy way to do this; a lot of the prompts in these examples are from the SDXL list or TWRI's list. None of the artist-Like prompt use the actual artist name, just descriptive terms.
Prompt for the sample images:
{style prefix}
On the left side of the image is a man walking to the right with a dog on a leash.
On the right side of the image is a woman walking to the left carrying a bag of
shopping. They are waving at each other. They are on a path in park. In the
background are some statues and a river.
rectangular text box at the top of the image, text "^^"
{style suffix}Generated with Z-Image-Turbo-fp8-e43fn and Qwen3-4B-Q8_0 clip, at 1680x944 (1.5 megapixels) halves when combined into a grid, using the same seed even when it produced odd half-backward people.
Full listing of the prompts used in this images. Negative prompt was set to a generic "blurry ugly bad" for all images since negative prompts seem to do nothing at cfg 1.0.
Workflow: euler/simple/cfg 1.0, four steps at half resolution then upscale and over-sharpened followed by another 4 steps (10 steps w/ 40% denoise). I find this gives both more detail and a big speed boost compared to just running 9 steps at full size.
Full workflow is here for anyone who wants it, but be warned it is setup in a way that works for me and will not make sense to anyone who didn't build it up piece by piece. It also uses some very purpose specific personal nodes, available on github if you want to laugh at my ugly python skills.