Why text based UIs can be so difficult. Here is my brief journey with chatgpt to get an image. The first pass wasn't bad. The effect was overall quite good. The characters were jibberish but I could live with that. The problem was the image I wanted was basically 'drawn in'. But overall, I was impressed.
@Cappyjax Thanks for trying that! I don't doubt this is a hard prompt. My goal wasn't to 'trick it' but to see if I could increment my way to a solution. I feel that just isn't possible with the current chat + output variation of the model
@scottjenson in my experience the generative image systems can get directionally correct but once you need to tweak specific details, it all falls apart.
Last try. At this point I just gave up as the variability was so great I couldn't 'hone in' on an image I'd like. This shows both the inability to talk to ChatGPT in a way that preserves progress, but how it's internal randomization is just makes incremental refinement impossible.
This illustrates (no pun intended) the failings of AI to make intuitive leaps — leaps which the human mind can discern instinctively. I don’t know what ephemeral alchemy exists within our minds that makes communication possible, but I do know that humans can communicate without words. In contrast, AI’s hamfisted attempts to understand & reflect instructions rely on an exact threading of the linguistic needle in order to accurately represent what is asked of it.