You may have noticed that the Z-Image-Turbo model doesn't produce a lot of variety when varying seeds. This native workflow is an attempt to generate more variety by starting encoding with an SD 1.5 model. Z-Image still does most of the work, and at the end, so I'm calling my use of SD 1.5 a "prefiner".

The prefiner adds variety to everything, but the un-described background seems most affected. While SD 1.5 generates a 512x512 image, there appears to be no problem with resizing the result to most shapes for final generation.

A few notes after 24 hours:

• My favorite scheduler for Z-Image is now sgm-uniform, rather than the workflow default.

• Turning up the CFG on SD 1.5 can reduce, but not remove, the patterns all over the place.

• Occasionally, a relatively high step break of 2 or 3 has the effect of turning the whole image cartoony. It's a nice effect but I can't do it reliably yet.