There are many large sets of curated, carefully-crafted text-to-image prompts that will generally produce high quality images compared to a simplified version of the same prompt.
By using an LLM to take a large curated set of prompts and simplify each one to only its core elements, we could easily create a dataset of simple<->crafted prompt-pairs.
For example, the crafted prompt "Keanu Reeves as spiderman, muscle extremely detailed, fantastic details full face, mouth, trending on artstation, pixiv, cgsociety, hyperdetailed Unreal Engine 4k 8k ultra HD, WLOP" would be stripped down to the much more natural "Keanu Reeves as Spiderman".
We could then train a language model to reverse this process (effectively “unsimplifying” prompts).
Further improvements could be made possible by iterating based on real-world feedback.
Could we train a language model to more closely align the language of the image model with the language of those writing prompts, even those already well-experienced with writing prompts for text-to-image models?
Could such a model even learn from patterns such as the ordering of miscellaneous keywords in order to produce more desirable results?