With the recent improvements that were rolled out in ChatGPT's image creation models (which became incredibly popular as the entire world seemed to Ghiblify itself), one thing that I got pretty excited about were people showing off how you could use it to create fairly sophisticated comic book strips. People showed dialogue in their images as well as character consistency, so I thought I'd try it out. Note that all of the images produced in this post are directly done with ChatGPT 4o.
A Ghiblified Avina, maintaining her signature determined look and posture
First, let's go through an unedited "story" where I gave as little guidance as possible to ChatGPT, so you can see what it looks like when the AI is given loose guidelines. For context, I initially wanted to try to reproduce the story I'd created about Oshuur, but I kept running into "policy violation" replies from ChatGPT. I'll go into these more later, but for now let's look at a story, produced by ChatGPT, about Avina opening up a Coffeeshop... 😅
Alright so your first question might be "why the heck is this story about Avina opening up a coffeeshop??" As I said earlier, I initially wanted to replicate my Oshuur comic book so I provided those images as guidelines and gave the story to ChatGPT to reproduce. It replied that it couldn't do it due to policy or content violations, but wasn't very clear on what the issue was. I tried to remove the stylistic references to Todd Macfarlane, thinking that might be the issue, but it didn't work. I tried to remove the "violence", but that also didn't work. I tried asking ChatGPT to remove any reference that didn't fit its own policies, but nothing worked. As a result I had to start fresh, and instead of providing it with a storyline I decide to give it just an image (the original artwork of Avina) and the request to write a non-violent story about Avina in a fantasy medieval realm. I suggested "such as having her open a coffeeshop" (reminiscent of Legends & Lattes) and, well, ChatGPT did exactly that.
Let's evaluate the above comic strip:
(1) The artistic style is quite good. It looks and feels like a comic, and although it's not super premium I am sure I could alter the prompts to change the look and feel. Overall the quality is solid, though there are certainly some errors (e.g. look at Avina's face in the last frame of page 1...)
(2) Character consistency is solid. Avina looks mostly the same throughout, though the style did change noticeably in the first 3 pages. She even wears the same armor and cloak throughout nearly all the frames, though there's a change in a few of the last frames.
(3) The story makes little sense. I mostly just asked ChatGPT to "go to the next page", so I offered very little help there. I'll get to what can be done a bit later, but in general I think ChatGPT currently isn't completely capable of creating a story on the fly and analyzing its own outputs to ensure storyline integrity.
(4) Secondary character consistency was a bit rough. If you look at the orc he changes very noticeably throughout the frames.
(5) Text bubbles work well. They're visually well executed and MOST of the text makes sense, though there are some weird (and hilarious) ones (Who's reapy for some frel-brewed coffie!??)
Alright so next up I wanted to take a slightly more structured approach. I gave ChatGPT an image of Fizbo and asked it first to create a script for an 8 page comic book. It did so and provided me with each page. In a nutshell, Fizbo shows up in a town and does a bunch of tricks to an amazed crowd, then gets challenged to a duel by a pompous wizard, prepares his magical trinkets pre-duel, duels the wizard and seems to lose but then tricks the wizard with some powerful magic coming out of his cards.
The first page looked great.
Fizbo looks solid and the setup clearly follows the script. Only the last frame above didn't make much sense, as Fizbo looks confused about the magic in his card (as if he's not in on the trick).
I proceeded to page 2:
Here ChatGPT totally lost its own plot. There are suddenly 2 Fizbos, the environment changed from the Town Square to a tavern, and instead of a wizard challenging Fizbo he somehow is setting up for the show, insisting (twice) that we shouldn't "look so glum". I was glum.
I proceeded to remind ChatGPT of its script, and asked it to try again. To that, it produced something hilarious:
I kind of love this one. He steals gold! Then steals a chicken! Then an orc steals the chicken but it's the chicken exclaiming "My ale!". Then Fizbo, who has proceeded to growing a beard in the meantime, uses a floating hand to steal a mug of ale from a dwarf. Epic.
Alright, I retried, this time spelling out the second page of the script to ChatGPT, and after a couple of tries got this:
So now we're sort of back on track, but several things are wrong. The conversation's weird ("You! Challenge you to a duel!"). The reactions are all off. And the final line should be said by Fizbo, not the wizard.
I clarified each of the mistakes to ChatGPT, and it then produced this:
Well, we've got a new wizard now, ha! But it's getting closer. I wanted to try a surgical change - swapping Fizbo for the wizard and the wizard for Fizbo in the middle two frames, so I asked it todo that but change nothing else. Here's what I got:
Another new wizard! It DID however manage to swap the characters as I asked it to do in the middle frames, BUT it changed the bottom frame and didn't quite get that one right.
Ok so my conclusion is fairly close to how I feel about most AI-generated products today. Whether it's writing code, writing articles, doing translations, creating images, or creating videos, today's AI tools are mostly limited by the directions provided to them. Furthermore, they struggled with high precision tasks, so even if the human guide provides highly detailed requirements, the AI will eventually jumble things up and make errors. And because it aims to please, it'll fill in the blanks when it's not sure and create, in many cases, more problems as a result.
The answer to the question "can ChatGPT create a comic book" however is a resounding YES. With patience and precise instructions, as well as some post-production editing to mix and match frames and fix small issues around text or other cosmetic errors, it's entirely possible to create a reasonably good comic book entirely with ChatGPT. In particular I think the character consistency (which is rather painstaking to do in Midjourney) is a massive improvement that actually enables multi-frame and multi-page outputs such as comic books to be possible.
While the outputs above are obviously not great, I'm sure it's possible to create great outputs. It would be quite time-consuming however as each image output takes several minutes and there are many issues to fix each time (and each attempt creates new issues).
Will Avina's lattes be burned and will Fizbo's chickens be furious? Only ChatGPT knows. And now, I leave you with a Ghiblified Fizbo for good measure.
