This may not be the most scientific analogy, but when I interact with large language models like GPT, Stable Diffusion, or MidJourney, I imagine I'm playing a high-stakes game of pinball. My prompt is the ball, and I aim to achieve a high score by hitting high-value neuron clusters. If I submit a superficial query, the ball bounces back with a shallow response. To extract deeper insights, I need to identify and target complex ideas (I'm mixing metaphors here). When I encounter a new topic, my initial question usually is, "Who are the experts on X, and what do they write about?" Subsequently, I bring in mentions of experts and key topics. In the case of art generators, this involves obtaining the appropriate keywords from GPT initially.
My objective is to make my text or image generator connect with all the essential nodes pertaining to a topic to extract the most conceptual or visual detail. This is an iterative process – I scrutinize the response for details that are unclear and request further explanation on those points. To generate a single image or social media post, I may have to make 10 queries – for images, it could require 50 or more iterations. With large language models, you can leverage the prior context you've established as a foundation to delve deeper. With image generators, you can modify an image by adjusting keywords, and even produce derivative images to correct errors or introduce new elements.
In short, if you limit yourself to a single prompt or image, you're not maximizing the full potential of these tools.