This is ComfyUI experimental workflow.

My impression is that Z-Image doesn't react to seed changes as much as earlier models. Side effect of having high adherence perhaps. For sake of experimentation, I'm looking to introduce a bit of chaos into generation process.

If you experiment with CFG setting you may find out that lowering it below 0.5, result in pretty much random image - unrelated to prompt. What if we used a very low CFG but only for the first generation step?

I made this workflow for easy adjustments of CFG, prompt and lora for each step of generation. So I could for example, apply a lora but skipping first two steps. Or use different prompt for given steps - start with one concept, establish image structure and then use it for something entirely different (which has similar effect as using low CFG). There's also preview for each generation step, so you can monitor what is happening where (if you do you may find that actually whole composition is set up in first step).

Just an idea if you're keen to experiment. It requires rgthree nodes installed.

BETA notes:

I fixed steps 2-9 having randomized seed for no reason, as result it would always re-render from beginning, even if repeating last render. I also removed my (completely wrong) notes about CFG.

It looks like CFG affects negative prompt the following way: at 1.0 negative prompt has no effect at all. More than 1.0 and it has impact, but at cost of increased generation time. Less than 1.0 and negative turns to positive with increased generation time again. At 0.0 CFG, negative acts as positive and positive is ignored. Interestingly, if you set CFG to 0.5 you get a blend of positive and negative prompt, as if both are treated as positive and results blend together. Something to abuse creatively.

Also added latent transformation nodes which could be applied after step 1.