This workflow can simply generate Text and Image 2 Image or it can go crazy: I added 4 image loader that get captioned and used for your prompt generation. Plus a random image picker function is available too.

I noticed that Z Image is kind of rigid when it comes to different seeds, so I had the idea for this and it helps: Since Z Image uses Qwen 3 the text encoding is out of this world, you can put a lot into the prompt and it will show up in the image. But you have to describe a lot... with this approach you can circumvent all that.

You should be able to just start without changing any of the parameters. The last image loader in line is also used for the I2I generation. Play around with the denoise value in the ksampler to your liking the rest is explained in note boxes.

If something is broken let me know. This is my first public workflow.

Have fun.