My workflow for Z Image Base with merged clip (Qwen3 4b & Qwen3 4b Thinking). Using ClownsharKsampler and a fixed seed. I find the optimal result is 30 steps and cfg 5.5 with res2m/beta or res2s/bongtangent and to generate at 1440x1440 (or equivalent rations). Negative prompt based on a Qwen negative that I found on civitai ages ago.
You can find that Thinking clip here: https://civitai.com/models/2277954/qwen34bthinking2507