Large Language Models Continue Being a Mixed Bag

I don’t have a lot to say tonight but I am showing up and trusting the process.

I used Microsoft Copilot to generate some images for use in my D&D game today. Mostly they turned out OK but there’s always small things that I wish I could change or tweak or combine. I always count the number of fingers on people the AI generates, for example, but there are also often smaller things. Copilot will generate four images for every prompt, and often, I’d love the ability to say “give me the third image but with the beard of the fourth image.”

Four images generated by Microsoft Copilot from the prompt "Obadiah Stane from Iron Man, dressed in medieval robes and armor, with dark red skin and long black ibex-style horns, holding a golden goblet in front of a roaring fireplace" — Only one of these looks sufficiently like Jeff Bridges in “Iron Man” for me to use (top right.)

Let me give you an example. I’m designing a dungeon and I can see it in my head but I wanted a picture of it to help me visualize it, and maybe if the picture was good enough I could show it to my players if and when they ever get there.I really wish I had a fun human artist friend I could give these prompts to, and that they would modify the images for me. Maybe that would be too much effort for random requests from me, though. I might be presuming too much about this theoretical friendship.

Image generated by Microsoft Copilot from the prompt "Looking out over a large shaft carved out of the earth going down deep. On the far wall is carved a bas relief of a dragon whose head is at eye level, and whose body extends down into the depths, its wings and arms extended out to ether side wrapping around the inside of the stone shaft. The observer, a dwarf carrying a torch, stands on a balcony opposite the dragon carving's head." — *Wrong: the platform is to the left of the bas-relief, the body should go all the way down, the wings should wrap around the inside of the well… ugh, this was all wrong.*

So the images get about 75-80% of what I ask of it. But asking for stories and creativity from them in writing is still very disappointing. As mentioned yesterday, I’m coming up with scenarios for an upcoming session of my campaign. I thought I’d ask ChatGPT if it could weave together my players and their character’s goals and backstories, and use them as plot hooks for one scenario. Maybe I’m bad at writing prompts, but all it did was give me separate scenes for each character, and it (in my opinion) chose only the most trite, cliche options for all of them. It wasn’t worth the effort to type in all the information I gave it. I punted on all the “ideas” it gave me and am starting over from scratch.