Using the same dataset as applied to the BigLiminal chroma set, which is a specific plain English oriented dataset meant to introduce the liminal prompting to the core model without requiring any specific prompting. It simply makes normal things liminal. If you wish to introduce even more liminal, you can use the "liminal, " tag with two spaces after.

I suggest 4 steps 1 cfg. It doesn't need negatives by default. Increase the steps to around 8 for huge structures like huge cityscapes, large suburban environments, complex interiors, and so on. It'll work, mostly. This one is less controllable than Chroma with a fair control of the images still.

D8 is fairly good at immediate deviations. It'll keep Z-Image mostly intact and still derive some useful and fun structural changes.

D16 derives a bit, still keeping much of Z-Image intact while introducing some very interesting deviances. Expect higher-differentiation and more difficult to conform images, while simultaneously introducing very interesting new elements.

D32 derives a lot for larger-scale imagery. The overall structure and composition of images will yield a higher fidelity output for much larger-scale differentiation for things like massive structures, huge interiors, massive cityscapes, and so on.

Small, medium, huge.

Have fun my friends. This is a very powerful outcome and it produces very interesting images. The training data is advanced, the output is very interesting, and the prompts are built in a way that makes plain English prompting highly effective.