I present to you "Red Stars and Stripes / Capitalist Comrades," a comic written by ChatGPT, illustrated by Midjourney, and assembled by me. It took me about six hours to put together.
ChatGPT script, for those who are curious: https://chat.openai.com/share/a3237b50-6a10-4fce-b58a-9ed434.... If you scroll down, you'll see that ChatGPT also provided descriptions of the illustrations. I used these descriptions as starting points for prompts that I fed into Midjourney. At the very least, I had to append "In a comic book style" to each description, and usually the modifications were more extensive than that.
Ironically, I often ended up going back to ChatGPT's original description for each scene verbatim (except for adding "in a comic book style") because Midjourney just wouldn't do what I needed it to do. It's almost as if ChatGPT knew Midjourney's limitations and crafted prompts that led to unsatisfying but accurate results. That was a better result than what I got with my more creative prompts, which Midjourney couldn't handle at all.
The main limitation I ran into with Midjourney was that it doesn't seem to be good at mixing "real world" contexts that have social meaning or semantic value. For example, I unsuccessfully tried for 2 hours to generate a scene involving a blue collar/factory worker-type person arguing with a diplomat at the United Nations. It could only generate a scene either with all factory workers in a factory or all diplomats at the United Nations.
Another curious example: I couldn't get Midjourney to generate any kind of symbol, flag, or emblem associated with the Soviet Union whatsoever. The best I could do is get the Vietnamese flag to appear when prompting for "communist symbols." But in those cases, Midjourney would only generate Asian people in those scenes. I couldn't get another type of person in any scene with a prominent communist symbol, ever. Same thing when I prompted for "American worker in Mao-style suit." The worker and every other person in the scene were also Asian, every single time. Seems like the model is overfitting in some strange way.
No comments yet.
