AI and Me

This past month, I decided it was time to refresh the panels as it had been over 4 years since they were last updated. I first reached out to the graphic artist I’ve worked with for years, but while he’s brilliant with icons, large panels aren’t his specialty. That’s when I decided to give AI a try.

After consulting with ChatGPT, I decided to use InvokeAI, because its free, runs locally, and doesn't require a cloud service. It was very easy to setup but it took several hours of experimenting with models and prompts before I could start producing useable results.. When I finally settled on models and was comfortable-ish with the prompts, I started work on the new panels

ChatGPT would generate the initial prompt and then we would refine it over multiple iterations until I got something close to what I wanted. For the prompts, not only do the words you use matter but also the specific order, phrasing, and negative prompts! Simply changing a single word, "warrior" to "fighter" for example, could completely alter the generated image. A few of the generated images were nightmare fuel (mouths in weird places or extra limbs). And the models really struggle with specifics. For example, if I told it I wanted "one statue" I might end up with one. Or five, depending on the other words in the prompt, the order of the words, etc.

I used the "generate" functionality in Invoke, which uses a text prompt, rather than reference images. In my experiments the models just didn't adhere closely to the reference image provided. There are many settings within Invoke. My final settings were: Euler as the scheduler, 100 steps, and a CFG Scale of 10. I also fixed the seed so all my images would  have a consistent style. I used the Architecture and Juggernaut models exclusively.

Sometimes I had to compromise between my original vision ("a sphinx in a hidden temple") and what the AI could reliably produce ("a statue of a winged woman holding a candle"). Being willing to compromise was a major hurdle for me initially. It was rare that I got what I wanted on the first try. But sometimes the AI would surprise me with something I hadn't considered and it turned out to be perfect.

Here are a few before and after pictures. First up is "ancient ruins", used for searching for an artifact.

Here is the one used for land travel.

The best advice I can give: patience. It took hours to create and refine the prompts, slowly iterating until the image was something I was happy with. The process can be frustrating, but if you stick with it, the results are worth it. 

​Until next time!