Graphic Novels and Comic Books

By antares

This is a short appendix to my previous blog post, that you can read here : https://budgetpixel.com/blog/an-interesting-comparison-for-graphic-novels I wanted to see if it was possible to number each page/sheet of a comic and insert a title in the opening panel of the story or, if you prefer, on each page of the story. Additionally, I needed more control over the accuracy of the lettering (captions and dialogues), especially when using SeeDream 5.0 Pro. For the page format, I chose 3:2 horizontal, because it fits better on the monitor. So, I wrote the following prompt for GPT-Image-2, using the same reference image as in my previous post: " Using the two characters and the same environment and background of the reference image, create a graphic novel page with 9 panels with black borders, about the woman entering a new chamber in the dungeon. They look attentively at the mysterious figures on the walls, then the woman turns to her companion asking if he can decipher their hidden meaning. The man looks around rising his torch to better illuminate the room, and tells the lady, calling her Diana, that he thinks they have reached a dead end. Diana seems worried but continues looking around in search for a clue. Make all the scenes and the lighting very consistent, starting from reference image. Put a small yellow circle in the bottom right angle of the image, with the number "1" clearly written in black inside it. On the opening vignette (top left) write the title of the story "TOMB RIDERS" in comic style capital letters. Pay particular attention to the correct and elegant lettering of the captions . " Here is the result I got: Nothing to say: the lettering is perfect, without errors, the page number is present, and the title is in its place, clearly visible . With SeeDream I knew I had to be more careful, because its lettering is less precise and not infrequently it makes mistakes. So I slightly modified the prompt like this: " Using the two characters and the same environment and background of the reference image, create a graphic novel page with 10 vignettes on a black background without margins , about the woman entering a new chamber in the dungeon. they look attentively at the mysterious figures on the walls, then the woman turns to her companion asking if he can decipher their hidden meaning. The man looks around rising his torch to better illuminate the room, and tells the lady, calling her Diana, that he thinks they have reached a dead end. Diana seems worried but continues looking around in search for a clue. Make all the scenes and the lighting very consistent with the reference image. Put a small yellow circle in the bottom right angle of the image, with the number "1" clearly written in black inside it. On the opening vignette (top left) write the title of the story "TOMB RIDERS" in comic style capital letters. Pay particular attention to the correct and elegant lettering of the captions, writing them in comic style capital letters . " And this is the resulting image: There are no errors in the dialogues, and the captions are correctly written. Only, the font of the dialogues is not the one I want, and they're not written in capital letters. But I think that with some further refinement of the prompt I will be able to achieve the same good level of GPT lettering also with SeeDream. I hope this post can be useful to everyone interested in creating comic books and graphic novels.

Tags: comic book, graphic novel, lettering, ai image prompts, title, page number

Comments (9)

antares

I see. Since I'm interested in this exchange of info, I can tell you that I mostly work with up to 9 reference images. In this regard, I find SeeDream more accurate than GPT in preserving the content of a reference image. SeeDream is also rather accurate in speech balloons, but fails miserably in captions. So far I have not been able to pinpoint the cause of this problem and resolve it properly.

antares

Based on the tests I've done so far, I've found that both models allow for very long prompts, up to 5000 characters, and that the longer, more precise and complete the prompt, the better the results will be. It takes time, but with if you apply the right method, you can often get a good image on the first try with both models!

panos

I niticed the 5000 words - prompt. I can't do GPT here so I don't remember if it has the same limit. CHat GPT does not, since I did one there, over 5000, yesterday and I copy/pasted the prompt here on SD 5 to see how it would be done and I had to delet the last parts of the promt, as they were over the 5000 words linit.

antares

Well, it's not 5000 words, it's 5000 letters (space and punctuation included). However, how long are your prompts? I barely reach 3000 characters with prompts that seem very long and detailed to me, describing each vignette, with every caption and speech balloon clearly specified ๐Ÿค”

panos

That is indeed so. A combination of the 2, taking into account SD's weaknesses is something I'm considering for speed up of production.

antares

Yes, although overall GPT remains more reliable ๐Ÿ˜Ž

panos

So, it worked!!!! Great info!!!

panos

WtF?!!! I actualy wrote 5000 WORDS?!!!! ๐Ÿคฃ๐Ÿคฃ๐Ÿคฃ I meant characters, of course! Some of my latest prompts on chat gpt - because of limitations on reference images - have forced me to copy paste detailed parapgraphs on charactes, places etc. looks. That makes some prompts HUGE! Over 5000 CHARACTERS, - heh - some times.

panos

9 referecne sure beats 4 to 5 of GPT. BUT the disasrtous text - and in a text heavy page like my own tend to be - makes it irredeemable to post edit.