Z-Image Base prompting experiment: composition and environment
#13
by MaximUA - opened
I ran a small controlled experiment with Z-Image Base in ComfyUI to better understand how explicit prompt structure affects image generation.
I tested composition and environment separately, using the same character and workflow wherever possible.
The full experiment, including the generated images, settings, methodology and observations, is documented her[e:
Reddit link]
Main setup: Z-Image Base INT8, Qwen3 4B, 768×1368, 50 steps, CFG 4.
I'd be interested to hear whether others observe similar behavior with different subjects, seeds or prompting approaches.