Unfinished Kris

‘Cuz I’m a nerd, I decided to take a prompt across multiple generative AI tools to see how the results might vary. I wanted to see how quickly I could prompt acceptable results, and just for funsies, in the process, how many bad results I might prompt.

‘Cuz I’m cheap, I only used free AI options. I used ChatGPT, Grok, Gemini, Claude, Copilot, OpenArt, Leonardo, and WordPress’ integrated “magic featured image generator”, which is Cloudflare Workers AI, so says Google.

I’ve learned over the last few years that using AI to generate images is not all that different than commissioning art from a human artist. The result will depend greatly on the details you provide. The same artist can appear brilliant or idiotic based on the gap between what is in the commissioner’s head and what is in the project spec.

Of course, all artists are different in how much they can intuit from a poor spec. So, I hypothesized that the same is likely true for different AI models.

The Prompt

Before we get to the results, I’ll let you see my prompt without a big preamble:

Create a comic strip with two characters.

Style: pop art meets sunday funnies

Layout: 3 panels – 2 small panels above one large panel

Character descriptions: 

1. Tooth – a full tooth with root is the body. It has a face, and arms and legs, and a little tuft of hair at the top. Tooth is straight-laced and tidy, a little nerdy.

2. Guy – a human guy who is kind of a lazy stoner type. He’s scruffy and his hair is unbrushed and he needs a haircut.

Background: Tooth and Guy are roommates. They are very much an odd couple situation.

The idea:

Guy is brushing his teeth with a toothbrush, but it turns out to be his roommate, Tooth’s, hairbrush.

Panels:

1 (top left): Tooth is walking into the bathroom. He’s in the middle of getting ready for work. He is saying, “I can’t find my hair brush. Have you seen it?”

2 (top right): Guy, looking a bit hungover, very tired, still in sloppy pajamas, is brushing his teeth and he’s is the middle of brushing his tongue.

3 (main bottom): We see Tooth and Guy. Tooth is standing in the bathroom doorway looking at Guy. His arms are crossed and his lip is curled in disgust. He looks mad. Guy is still holding his tongue out like he was when he was brushing it, but he is now looking at the toothbrush with a look of “oh shit” on his face because he was accidentally brushing his teeth with Tooth’s hairbrush. Tooth is saying, “Bro. Again?”

Okay, now a little post-amble…

I don’t consider this to be a “poor” spec. It’s decent. I took some time to write it up, and I think most human artists who got this spec would get the joke and create something fun that expressed the expected sentiment. If I had specific colors in my head about the bathroom, the pajamas, the hair, the toothbrush, etc., I would have wanted to include them – but I didn’t, so I didn’t include them.

The Results (Round 1)

The Results (Round 2)

You’ll notice fewer results in round two. The reason for this is the tight limit on the number of images you can generate in one sitting on some of the token-based tools. I definitely didn’t mind not going for a 2nd round with Leonardo, but I did want to give the Nano Banana model on OpenArt AI another chance. It created my favorite first round result. So, I did a few tricks to earn enough tokens to try again.

I tried a few more times, a few different ways to get Grok, Cloudflare Workers, and Gemini there, but things didn’t get better. I’ve worked with artists like that before. Artists that gave me output like the Cloudflare Workers AI results didn’t stick around very often – too Avant-garde for my brain. Maybe I’ll get it one day, but today is not that day.

Sometimes a bad result is just a case of asking the right artist for the wrong result. Grok didn’t do well in this challenge, but I bet’cha it would come out with amazing results in a video creation comparison. I’ve had that exact experience with an amazing video artist in the past.

I have a feeling that Claude is in the same boat as Grok in this challenge. I’ve heard lots of praise from different people about Claude, so I may have been asking a fish to compete in a climbing contest. This was my first time trying it.

This was a fun little experiment. Some takeaways:

1. My hypothesis was correct. Some tools were better at intuiting through ambiguity than others – just like some artists are.
2. Nano Banana creates some lovely graphics, but I got vastly different results when it was used in Copilot Vs. Openart.
3. Not all great tools will be great for all applications.
4. No matter how much time you put into your prompt, the instant you get a result, you’ll realize where you went wrong.
5. The instant you realize where you went wrong, in most cases, you can just try again!
6. Unsurprisingly, because I wasn’t specific, Guy was a white guy in every result.
7. Guy would be a terrible roommate.