The last month I have figured out how to put a merged uncensored mixtral 8x7B model into the character of being a master at writing stable diffusion prompts and recommending an art style and artist to use for the prompt. These prompts are very effective using SDXL models. You can find these models on huggingface and you can find a version of llama.cpp that can run them on github.
I just use a single lora for gross control of the images. I choose the lora based on the topic of the image. Just with this level of control I am able to effectively replicate a lot of other images I find on the internet, even when those images use specialized loras that I don't have.
The first thing you need to do is tell the chat ai program that it is a stable diffusion text prompt master that rewrites descriptions into highly detailed, descriptive stable diffusion prompts that describe the art style to use and the best artist to recommend for the subject and the art style.
Then you say, Rewrite this prompt: follow by a list of keywords or a text description. You can feed it a prompt alone with a brief description of what the image should be. You can even just give it a brief description of what you want.
If you read what it says and want some changes, you can ask it to rewrite the prompt giving it the recommended changes you want to see. Or you can just copy and past it into the stable diffusion text box and edit it to make the changes you want.
Then hit generate and make changes to the positive prompt to add things or reinforce what you want more of, and list what you don't want to see in the negative prompt. Once you start getting what you like, make it generate a bunch of them. :D
Once I started doing this it became rare to get bad results. I had to struggle to get enough bad results to even enter 5 bad images in the last contest.
This has been a real game changer for me. I hope it helps you generate images too. I want to note that these large language models are very good at translation and can probably let you work on the chat ai side in your native language, while generate an english text to use in stable diffusion.