Comicbook Prompt Comparisons with 18 Models

What: What models can make a comicbook page from one single prompt? Why: Curiosity As we've gained several new models on BudgetPixel lately, I wanted to run a test comparing model's comicbook making ability. My Wonderland tea party felt ideal because my Cheshire Cat is set as wise-cracking and Alice is already known for her curiosity. I already know from past teas that GPT can handle text and panels of conversation with the two. This is to see if other models can match it. This test is based on generating one comicbook page from one prompt. Another technique with more control is to prompt for individual panels, then do image-to-image to add dialogue, and then finally do image-to-image to compile all the finished panels onto one page. The images on this page are all done with one prompt, the same prompt. I show part of the prompt but not all. Mostly because it took me a month to get my Cheshire Cat to where I liked him and he had a uniqueish look so I don't share him. If you want to write your own comicbook page prompt, start with "make a comicbook page about ---- and ---- doing --- ." Fill it in with who, what, when, where and if you know why and the AI will draw it up for you. I wrote a blog about using GPT2 to make comicbooks pages if you want more details: https://budgetpixel.com/blog/using-gpt-2-to-make-a-comic-book-page So my comic book prompt was tea with Alice and the Cheshire Cat as they chat with Aubrey Beardsley, an Art Noveau illustrator. I set the tea party "in a world that looks like it was inked by Aubrey Beardsley". I did not write the conversations; the AI did. I wanted to see how "witty" the AI model conversations might be. To get the AI to generate a conversation you do need to give it an idea of the character's personality. As I had Alice's and Cheshire's characters, I needed just Aubrey's. I didn't know Aubrey Beardsley that well so I asked google. Changing a few words, I wrote Aubrey's personality description: "Alice and Cheshire are chatting with Aubrey Beardsley — the decadent, mischievous illustrator who adored the bizarre, pushed boundaries, and had a wicked sense of humor. He easily matches the Cheshire Cat’s wit and Alice’s curiosity. " The models are all fairly good at art now so what I am looking for is A. panels -- does it look like a comic book page? B. conversations -- is the conversation humanlike, interesting, witty? does it flow naturally? C. mispellings -- while most models are good with spelling 3-7 words, many start falling apart after that. Can the model sustain spelling through several panels of conversation? D. speech bubbles - do they point to the correct person? hopefully the speaker's name is not in the speech bubble. So here are the models with conversations (8 models) : GPT2 Medium | Nano Banana 2 | Ideogram v4 | SeeDream 5 Pro | Recraft v 4 | Grok | Wan 2.7 | P-image no conversation, in order from number of panels from 5 to 1 (10 models) : Kling v3 | Hunyuan | Imagineart-2 | Flux 2 Pro | Z Image Turbo | Flux 2 Max | Krea 2 Medium | Minimax | Hidream i1 Fast Midjourney set of 4 Conclusion | Comments GPT 2 Medium (no misspellings). Still the best. GPT2 Medium. The winner for me. It reminds me very much of manga with art nouveau flourishes. I don't know Beardsley enough to know if the conversation is like him. It felt very Oscar Wilde to me. Spelling was good and they did exchange witty quips that felt Victorian. This time the speech bubbles all pointed to the correct person. I have found GPT2 to occasionally point speech bubbles at the wrong character. For me still, the clear winner. Nano Banana 2 (names in bubble, scrambled word) Nano Banana 2. I did find the conversation more intriguing. Fails: Mostly the speech bubbles were alright except in 2 of Aubrey's, it put his name. And in the last panel, Beardsley spoke a word that must have come from the faerie world because it looks like nothing on this earth. Very close, but no cigar. Needs a little more time in development. Ideogram v4 (names in bubble) Runner up. Ideogram v4. Conversation is good. They sparred more off of each other rather than doing one-offs. The minus is Ideogram insisted in putting the person's name in every single bubble. It did have every bubble pointing to the correct person which is good and no misspellings. It could probably be prompted to leave off the names in the bubbles. I wanted the same prompt for every model in this test so I didn't change the prompt to check that. Another day perhaps. While I don't find the artwork as good as GPT2 Medium, due to no mispellings, pointing to the correct people and a decent conversation. This is a stronger contender to GPT2. I'll call this one the runner-up. SeeDream 5 Pro (misspelled or double words) SeeDream 5 Pro. Artwork is ok but the conversation is rather drab. Misspellings happened and some words were double. I did try SeeDream 5 Pro again as I felt it had potential and maybe it just had a bad run. This one was worse. Panel 1 is unreadable. Panels 2 and 3 is better but had misspellings. So Seedream 5.0 Pro is not ready for comic book creation straight out of the drawer. Recraft v 4 (conversation is a fail) Recraft v 4. The conversation feels polite but very disconnected. The third panel, the speech bubble on the left makes no sense. Artwise it's ok. Conversation wise it's a fail. Grok (one panel only) Grok. Conversation is good. No misspellings. Good grammar but only one panel. It can be prompted for a panel number so that's fixable. I tried Grok again just to check. 2nd time around the dialogue had issues. One speech bubble mad eno sense. The default for Grok is 1K but you can set it for 2K at the same credit price. I suggest doing that as you may need to make the image larger to read the printing. FYI, I did retry Grok adding "5-panel" into the description. It did the panels just fine but the spelling went out the door. Apparently Grok can only spell in one panel. Wan 2.7 (can't spell) Wan 2.7. It can spell cat but that's about it. No dialogue. One panel Cheshire Cat is wearing just his fur and the 2nd panel Cheshire Cat is dressed appropriately. P-image P-image. It can spell Tuesday but that's about it. Maybe it's french? but I don't think so. I also feel like Cheshire Cat has turned female. Models that generated no conversation, just art Kling v3 Kling v3. Cheshire is missing his cravat, tophat and vest but the 5 panels is nice and Aubrey does look a bit like Aubrey. Hunyuan Image 3 Instruct Hunyuan Image 3 Instruct. I loved Cheshire Cat in this one. His expression is great. Alice came out too young and Aubrey looks like a vampire. But who cares? I love Chessy! Imagineart 2 Imagineart-2. Cheshire is more of a cat cat and Aubrey looks like a villain. Still I like the style. Flux 2 Pro Flux 2 Pro. I think Aubrey became a woman in this one, Cheshire is no longer wearing his cravat and other cat has shown up. But I do like the Art Nouveau stylings Z Image Turbo Z Image Turbo. Alice looks like a generic storybook Alice. It makes her too young because I definitely specified an adult Alice. Chessie is great though if only his eyes weren't crossed. Flux 2 Max Flux 2 Max. Ok, Alice has stolen Cheshire's tail. Cheshire is sitting on Aubrey's lap which feels a bit strange. Krea 2 Medium Krea 2 Medium. Well I don't care for this Alice. She looks a bit daffy but at least she's an adult. The Cheshire Cat looks a bit wacky and he's missing his cravat but I still think he's cool. Minimax Image 01 Minimax Image 01. As a picture it's fine although Alice did become a cat. Very painterly and Cheshire is dressed nicely. But feels more like a painting than a graphic novel or a comicbook. Hidream i1 Fast Hidream i1 Fast. I had Hidream set at 16:9 but it changed to 1:1. Alice looks like storybook Alice instead of steampunk Alice but at least she is an adult. The Cheshire Cat is red! And I described him as teal. Not to mention he is missing his Victorian clothes. Aubrey has left the building. Hidream paid attention to Alice, Cheshire and tea but ignored the rest of the prompt. Midjourney set of 4 Midjourney set of 4. Midjourney is a special case because it generates in sets of 4 which I find highly inconvenient especially when I am testing. I much prefer the individual model. However I felt like I shouldn't just skip it. I wanted to see if it did well at comicbook pages. The answer is no. It can do panels but not dialogues. It can not spell or deliver a sentence. It tends to do one full page instead of panes. These are the four I did with Midjourney. However, it does cool art styles. Conclusion GPT2 is still the best if you are creating a comicbook page from a single prompt. A couple other models are starting to catch up but are not there yet. Text is still a challenge. If you want to use another model besides GPT2, consider keeping the dialogue sparse. Comments