Daily Generations: More Bard Troubles

I wasn’t happy with the output from my last Bard Troubles post, so decided to give it another go. See here.

Rather than give it my own original work to start from, this time I went with only the description. I think it actually performed better with this method.
We ran into similar challenges as last time regarding representation in training data. I’ve written a couple of times on this issue recently, so for this post I’ll highlight a related but different challenge when using generative imaging.

If you’re trying to create something that has not been created many times before by a human hand it will likely take many more iterations for AI to land on the vision you’re trying to achieve. Today’s case study of a modestly dressed, dark-skinned, female elf with red hair is not exactly groundbreaking. It does quickly become clear that the source material in the training data did not contain many examples with a similar mix of traits.]

Creating D&D characters may not be when you first run into this issue, but if you are working on something truly novel or innovative you will likely find the robot is just not “getting it”. Generative AI is the best mimic ever created, but it is not yet truly creative.

There is always talk about future AI that will possess the ability for true inception, but as I understand it this will require a fundamental shift in the way the technology works on a basic level. I won’t rule out this possibility, but for now and the foreseeable future we will continue to need human designers, artists, writers, and all manner of creative professionals to continue to innovate.