Introduction

In part one, I analyzed the current forays with AI in the comic book industry, and explored reception of AI through industry polls and studies. Visual artists are generally against the introduction of AI due to copyright and economic concerns. Several online comic communities have voiced displeasure over the introduction of cover art generated by AI, and have restricted AI imagery from being posted. However, general population studies have found positive reception toward AI’s introduction, and that further exposure to AI heightens that positivity. Meanwhile, studies focused on the “artistic” aspects of generative AI reveal negative bias, and a preference for human creations. None of these studies, however, account for the multimodality of comic books: text and image. Is a human-written story a mitigating factor for readers? Is it feasible for writers to become “AI comic book artists" or will consumers drive AI imagery out of the market altogether? Given the multitude of information making it difficult to discern how comic book consumers perceive AI imagery, I moved to first-person research through the generation of an AI comic and a survey conducted on unsuspecting comic book customers. 

Can you create an AI comic book? 

AI models advance at breakneck speed, so reviews, articles, or media from even six months ago can be outdated regarding current capabilities. To truly answer one of “Is the technology there yet to adapt comic books using AI models?” I gave it a try using Midjourney. For a more general overview and description of Midjourney’s mechanics, please refer to AMT Lab’s review article. 

The Goals 

There were a few objectives when creating the comic. 

1. Avoid using any image not made by Midjourney as part of any prompt for the comic, or at least any artist’s image.

2. Avoid referencing any specific person or work in a prompt when generating images to be used in the comic. 

3. Create a comic that is difficult to discern as AI generated. 

Ethically, it was best to make efforts to avoid copyright infringement, particularly since Midjourney V6 has been criticized for generating images that are close to existing works. Restricting existing images and intellectual property would challenge the model to create images derived from a user’s text inputs. Then, for the purpose of the survey, the comic had to be consistent and without obvious error to prevent one of the test groups from realizing the comic used AI images before they ranked it. 

The Process 

It was important to begin with a story and a concrete idea of what each panel should look like. The inspiration came quickly: adapt two scenes from an arc of my homebrew Dungeons and Dragons campaign, Tales of the Five, conceived by Sean Murphy. The first scene would be simpler to adapt and not require the characters to be within the same image. The second scene would lean into the horror elements that AI imagery can excel in, as highlighted in part one. The resulting narrative came to 32 panels. Ultimately only the first scene, originally imagined as 12 panels, was acceptable for use. It was adapted as 17 different panels along with the cover art. For ease of consistency, the author and I decided the art style should be black and white, minimalistic, and manga-inspired. This removed the issue of color variation and made the comic more achievable if there was to be a second comic adaptation made by a human artist. This would allow a fair, lighter workload for a human artist while still being comparable to the AI comic.

Ultimately, the core character reference image I settled on for Leo, an original playable character and hero in distress, is below. This reference image had a lasting impact on the art style of the rest of the comic, which was not the intended hand-drawn aesthetic but a more digital look. 

In order to create consistency in characters and art style from panel to panel, Midjourney relies far more on image referencing than text prompts. The basis to create consistent panels were the character and style reference parameters. Additionally, parameters for prompts were often set to be the lowest quality and stylization setting. This was in order for Midjourney to more accurately follow the prompt and avoid adding inconsistent details. 

Overall, Midjourney was quite successful when rendering scene one with the characters in separate panels. However, issues immediately arose when trying to generate panels for scene two. Despite any combination of reference images (or lack thereof), or organization of the text prompt, Midjourney was unable to put characters side by side. In a majority of instances, aspects of characters would be blended together. 

Additionally, Midjourney struggles to understand complex prompts. Considering you have to dedicate much of a prompt to describing your characters to remain consistent, this exacerbates the challenge of rendering a more complex shot. 

Final Results 

Overall, it took the better part of twelve hours, 4.5 hours of GPU time, and 250 prompts to create the comic. Goals 1 and 2 were compromised slightly in order to render an image of a dungeon cell. A manga establishing panel had to be used as a style reference as well as a photograph of a dungeon. I considered the images different enough to be used. 

In this author’s opinion, it’s evident that AI diffusion models have primarily been used for cover art because it’s too soon to make comic books on par with human artists. A few more years are needed for newer models capable of differentiating different characters and creating more engaging shots beyond character portraits and backgrounds. It is not a matter of if it will happen, but when. 

To help assess goal three, read the comic and tell us your opinion below in the poll.

Comic Book Survey 

Population(s) of interest

The survey was intended to be taken by those who have demonstrated an interest in the comic book industry. Three “locations” were chosen to distribute the survey: New Dimension Comics, a large comic book store in the south side of Pittsburgh, Phantom of the Attic Comics in Oakland, Pittsburgh, and online comic book communities. In total, there were 108 participants across the three locations: 60 from New Dimension Comics, 36 from Phantom of the Attic Comics, and 12 online participants. 

Methodology
In part one, two studies of interest were Humans versus AI: whether and why we prefer human-created compared to AI-created artwork, which had participants rank various pieces of art labeled as either human made or AI generated, and Foregrounding Artist Opinions: A Survey Study on Transparency, Ownership, and Fairness in AI Generative Art, which assessed the sentiments of artists regarding generative AI. A survey for comic book consumers was developed using these studies as a reference point. This enables the study to discern if comic book consumers differ from general populations or artists, or if results are congruent with their findings. 

The first portion of the survey consisted of participants reading the AI comic book and answering five questions on its performance, modeled after Humans versus AI. The questions assessed enjoyment, conveyance of narrative, emotional resonance, consistency, and dollar value. 

The study was conducted over the course of three weeks in April 2024. Two versions of the survey were created and distributed as google forms via QR code or physical copies for participants in each location. Participants were randomly assigned to Group A or B. Group A participants had no prior knowledge that comic panels were generated by AI. Group B participants received written or verbal notification that all images were generated by Midjourney prior to reading and ranking the comic. After ranking, Group A received notice that the images were generated by AI and were given the option to discontinue. Only one participant opted out of the study. 

The second part of the survey data, given the comic book assessment was complete, was analyzed collectively without distinguishing between Group A and B participants. The second portion of the survey consisted of questions to identify other possible variables (knowledge of AI, reading frequency, and artists vs nonartists), and opinion questions modeled after those in Foregrounding Artist Opinions. All survey questions can be found at the bottom of this page. 

To compare responses between groups, likert scale responses were converted to numeric values on a scale of 1 to 5 for t-testing or anova testing, depending on variable. Questions with three possible answers were assigned scores of 1, 3, or 5 for testing. 

Key Findings 

Across Groups 

In the aggregate, some data matches expectations while other responses do not. Overall, sentiments regarding ownership, disclosure, and threat status match that of artists. Interestingly, participants did not see AI advancements as a positive development in the comic industry, which defies expectations that one may have if considering the Twitter study discussed in part one, in which a majority of analyzed posts indicated generative AI developments to be positive and beneficial (Miyazaki et al 2024). 

Surprisingly, more participants were potentially willing to purchase AI comics than online discussion has indicated. 43.52% of participants responded they would not consider purchasing comics with A.I. generated panels. 44.44% replied “Maybe” while 12.04% replied “Yes.” 

When considering responses to the AI comic in the aggregate, only one of the questions had statistical significance that supports an AI negative bias hypothesis, emotional resonance. Group A had an average of 3.05 (somewhat resonant) while Group B had an average of 2.67 (between “a bit resonant” and “somewhat resonant”). A one-tailed t-test reported P<.05 (P=0.041). A lack of statistical significance is a result of different responses between the online and in-person locations.

Online Discussion 

Attaining online participants was difficult, resulting in a small sample size (n = 12). Initially, inquiries were sent to various subreddit moderation teams for permission to post the survey. Some moderation teams replied “We do not allow A.I. in any way,” ending the discussion. Only one subreddit permitted the study, garnering little attention. Six more participants were acquired through comic-related Discord servers. Attempts were made to expand the study to Facebook groups, but the survey was removed or failed to be approved in three groups with no explanation. After bot response issues, all further attempts to participate in the online forum ceased. 

Although the online sample size is much smaller than the comic book store locations’, the contrast in responses is striking. 

Based on the data visualizations of the comic rankings, it appears that the online group had a pro AI bias. Group B ranked the online group more highly than Group A, the opposite of the other locations. However, I hypothesize this is due to only six participants in each group, and Group A’s keen awareness. Three of the online participants in Group A, prior to ranking the comic, notified me that the comic was AI-generated. One participant said “That (generative AI) s*** does not have any place in the comics industry.” These participants gave the comic poor marks. 

There were still marked differences beyond the comic ranking, however. 

Such data represents that online communities may feel more strongly about the introduction of AI and its ethical and economic ramifications for comic book artists. Unfortunately, these ardent sentiments made it difficult to accumulate a larger sample and more study is needed. 

Phantom of the Attic Comics vs New Dimension Comics 

There were not many differences between the two physical locations. Differences between the comic rankings were not statistically significant beyond differences in group B responses to comic book enjoyment. Phantom of the Attic had more negative perceptions, with an average of 2.78 vs New Dimension’s 3.36. A one-tailed t-test reported P<.05 (P=0.021). A comparison of the stores’ comic rankings is below. 

There were some significant differences in three of the opinion questions, visualized below. 

Phantom of the Attic Comics is a pedestrian-friendly comic book shop situated in Oakland, Pittsburgh in proximity to University of Pittsburgh and Carnegie Mellon University. While it was predicted the location would have many university participants, many of the participants were frequent visitors to the store, diverse in age groups. A minority of participants identified as “casual” readers in the study (4/36). A majority of respondents from Phantom of the Attic would not purchase a comic book with AI generated panels. 

New Dimension Comics is the largest comic retail chain in the Pittsburgh area. This location was situated on the Waterfront development in Homestead, PA, making it a more car-friendly location. New Dimension Comics, being a larger location, also sold many items unrelated to comics and attracted fans of Dungeons & Dragons, Magic: The Gathering, video game series, etc. This led to many participants identifying as “casual” readers in the study (26/60). Only 31% of respondents from New Dimension Comics would not purchase a comic book with AI generated panels. Such specific differences may mean reading frequency has an impact on responses.

Reading Frequency 

The frequency at which a respondent read or purchased comic books appeared to have an impact on comic rankings. For comparison, frequency was divided into three categories: casual readers (0-1 time a month), average (2-5 times a month), and enthusiasts (6-9+ times a month). More frequent readers would respond more negatively in Group B, significantly so for conveyance and consistency.  The reading enthusiasts acknowledged the truth of the comic: it wasn’t that good. They were harder to please than casual readers. 

There was also a significant difference in responses to the question of if AI advances were a positive development in the comic book industry, visualized below. This parallels the zeal of online respondents, indicating that die hard fans that wish to discuss the latest comics are more inclined to disavow generative AI. 

AI Understanding 

There were no significant differences in responses when analyzed by familiarity with artificial intelligence. Only a small number of participants, less than 17%, responded that they do not understand how generative AI models work. The sample size of those that responded as such in group B was 5, so more investigation is necessary. 

Artists vs Non-artists

There were no significant differences in responses when analyzed by artistry. Participants were asked to answer if they would consider themselves to be a visual artist and/or comic book artist (professional, amateur, or hobbyist). Responses were divided into two groups. Those that responded “Yes” or “Somewhat” were categorized as artists, and “No” as Non-artists. Given that responses align with our current knowledge of illustrators, artists, and comic book artists as outlined in part one, we can surmise that comic book fans agree with artists about how generative AI should be treated in the industry. 

Limitations of the Study

1. Small samples: As previously discussed, online responses were too slim to be considered concrete evidence, warranting further study. Further evaluation is needed to see if familiarity with AI is also an impact. 

2. Evaluation inaccuracies: As these were self-evaluations, some responses may not be accurate. For example, while many participants asked about how the comic was made and how diffusion AI models work, they replied they were “somewhat” familiar with models. 

3. Dollar value: There was a question assessing the dollar value of the comic. However, a large number of participants had no idea how they should answer the question. This was due to multiple factors: 

   1. Many participants read comics free online and do not typically pay for them. 

   2. Readers are more likely to read and purchase longer works or volumes, rather than single issues. 

   3. The comic itself is much shorter than the standard 20 page comic issue, making it difficult for evaluation. Some respondents treated it as it was, others made a guess at the value of a full issue. 

   Given the huge variance in responses and participant confusion, results have not been reported. It is worth noting the most frequent answers were between $3 and $4, roughly the value of a comic issue. Enthusiasts valued the comic between $2 and $3 on average. 

4. Location: Survey responses are indicative of sentiments of comic book fans in the Pittsburgh area, but it should not be assumed that this study applies on the national level. 

5. Ownership: Participants were only given three options regarding ownership of AI generated art. This was to discover which party should have “primary” ownership, but some participants stated that they wished there was more nuance available. 

Group A Reactions

While Group A participants were not explicitly tasked to express if they suspected the comic had AI images, no in-person participants stated they had suspicions during the ranking portion of the survey. A few participants double checked the cover page to see who the artist was and asked. Some people considered the art to be “fantastic” while others said they felt there was “something off” about the art in hindsight. Overall, a majority of Group A participants were surprised and impressed that Midjourney was able to generate panels with such a high level of consistency. 

Final Thoughts 

Based on the results of this survey, it remains unclear if generative text-to-image AI will become ingrained in the comic book industry. Most likely, consumers will connect less with comics they know have AI imagery, but many might be willing to make a purchase under the right conditions. 

Many participants expressed that generative AI needs to be a tool, not a replacement. There was recognition that many would-be creatives are held back by financial constraints, and access to AI models allows them to create proofs of concept. Some were aware that AI models could be efficient for artists to use by being trained solely with their own art, especially for those that have developed drawing difficulties over time. 

Regardless, the overwhelming majority of participants believed that generative AI requires transparency and can pose a threat to artists without additional copyright protections and regulations in the generative AI industry. Artists can rest easy knowing that fans are invested in their artwork and stories.