> I still cannot fathom how it is that image gen is still so far ahead of LLMs in regards to prompting (still can’t weigh prompts for chatGPT but I can on any i

> I still cannot fathom how it is that image gen is still so far ahead of LLMs in regards to prompting (still can’t weigh prompts for chatGPT but I can on any image generator). Is that something you'd really have use for? In case of images it's easier to immediately judge the effect of a prompt change and its direction. LLMs just blurt out a blob of text that you have to read and analyze every time. Even when rerolling the output for the same prompt, you get a very different-looking result. I find comparing a large number of prompt-output pairs very exhausting. Meanwhile, scrolling through a set of outputted images is a breeze, and the output quality is comparable at first glance. LLM prompts can also be much longer, consisting of multiple sets/layers of instructions, etc. You just slap a new sentence at the end to modify the behavior. Meanwhile, for image generators it feels like every word counts, and can sometimes take the output in an unintended direction, hence the need for surgical control over word weights.