It really can’t. It can take your original prompt and fluff it out to obnoxiously long text. It can take your visual concept and sometimes render roughly the concept you describe (unless you hit an odd gap in the training data, there’s a video of image generation being incapable of generating a full wine glass of wine).

A pattern I’ve seen is some quick joke that might have been funny as a quick comment, but the poster asks an LLM to make a “skit” of it and posts a long text that just utterly wears out the concept. The LLM is mixing text content in a way consistent with the prompt, but it’s not mixing in any creatively constructed comment, only able to drag bits represented in the training data.

Now for image generation, this can be fine. The picture can be nice enough in a way analogous to meme text on well known pictures is adequate. Your concept can only ever generate a picture, and a picture doesn’t waste the readers time like a wall of text does. However if you come at an LLM with specific artistic intent, then it will frustrate as it won’t do precisely what you want, and it’s easier to just do it yourself at some point