its a mixture of a couple of shortfalls
this type of thing possibly already exists within the latent space. but I have noticed the way prompts have been written and how it forces you to write them that effects the way images are created with Z-image. specifically how anything that uses natural language.
it has not been my strong point. too short of a prompt and it doesn't follow, too long and it gets confused and just defaults to the first few tokens.
but if you don't overdo it with the amount of words in your prompt, it starts to change the way it works.

my training settings are mostly default. with the difference being for styles I use shift instead of weighted.
llms also give generic answers which lead to nowhere 90% of the time. or maybe Im not asking the right questions. but I found in the past, trial and error always fared better, as I have a strange way of learning on my own.

please, I really appreciate your feedback and I hope you do not think that my reply is dismissive. you're literally one of the greats on here, I'm just doing my best with the resources I have😅​

thanks 🥺​ Im gonna try and see if I can make it better