AI Image Generation: Impressions and Examples (Patreon)

As Maverick noted in the December State of the Patreon, we recently subscribed to an AI image generation service, NovelAI. Our objective was just to learn a bit more about it via some personal experience and experimentation. We know this type of AI tool is a subject of controversy, particularly among illustrators and other creator types. Asaragi, an illustrator whose assistance we’ve commissioned before, even made a cute, on-point little sketch about the issue a while back. Indeed, the controversy itself is part of what sparked our curiosity.

In the remainder of this post, I’ll share with you a brief summary of my own overall impressions of the NovelAI image generation functionality, as well as some highlights of the more interesting results from several particular experiments. This is all based on perhaps a dozen or so hours of experimentation across several days. Maverick may have some additional insights of his own to share, so maybe keep an eye on the comments for those.

Summary of Impressions

First, let’s go over its limitations. At least when it comes to the NovelAI tool, perhaps the most notable “feature” is that its image generation functionality is essentially limited to an “anime” style. Further, NovelAI seems to be chiefly limited to creating single-character images. Because of these limitations, and probably because of the scope of the source material it was “trained” on, the images NovelAI produces tend to have a “same-y” kind of feel. You’ve likely experienced this yourself if you’ve seen the large number of AI-generated illustrations popping up on DeviantArt and elsewhere.

The tool is also somewhat unpredictable and inconsistent in terms of how it responds to various keyword prompts. On the one hand, this can be helpful because you can repeatedly re-create an image based on the same set of prompts, which helps to mitigate the often weird, anatomically jumbled images the AI assembles via algorithm. Essentially, you can treat it like a slot machine and keep cranking it until you get something you like. On the other hand, this randomness also makes it difficult to retain a consistent character appearance and details across multiple images. In my experience, that inconsistency remained the case even when leveraging features ostensibly designed to improve consistency, like an image “seed” or an uploaded image. When it comes to illustrating kinky weight gain stuff, where sequential images showing progressive gains are a mainstay, I think this a pretty big shortcoming.

That said, NovelAI did surprise me with its adaptability. While entering almost anything at all as a keyword prompt would produce “something,” the AI did seem able to understand and accommodate a lot of prompts for poses, backgrounds, and basic clothing. It also could accommodate guidance for body types and weight distribution, but with a bit less flexibility and reliability. Notably, NovelAI seems to skew towards extremely busty females. Toning that down requires an active effort by the user. Love handles and potbellies were easy enough to tease out after some trial and error with keywords. “Hanging belly” seemed to be one of the most reliably effective prompts, which struck me as odd, but whatever.

Also intriguing was the ability of NovelAI to recognize the names of specific individuals as a prompt. This works not only for identifiable illustrated characters, like Samus or Zelda, but also for real celebrities. While the fidelity of the results from these prompts varied, it was good enough to catch my eye. Similarly, using an uploaded image as a prompt could enable the AI to produce a surprisingly faithful interpretation of the same scene—subject to its usual quirky variations.

Putting all of that together, here’s my overall takeaway. NovelAI can be a fun toy, capable of producing some pretty good, kinky imagery tailored(ish) to your desires, with minimal user effort required. It has a substantial element of variability that’s largely beyond control. When just trying to make a single, titillating image based on some basic prompts, that variability is a boon. It’s easy to lose a few hours playing with this combination of user-influenced randomness and not get bored. But that variability also limits its potential for creating sequential images or anything else requiring consistency. Trying to produce a comic or a set of story illustrations with NovelAI, at least as it currently exists, seems like it would be somewhere between frustratingly inefficient and futile, with an end result that, at best, will look tellingly similar to thousands of other AI-generated images floating around.

Experiments and Examples

Here are a few curated results from my various experiments with NovelAI, showcasing some of the considerations I discussed above. I’ve also provided a Google Drive link to download a zip file with additional images if you’re curious to see more. I’ll attach the zip file to the Patreon post itself, soon, too, in case the link gets overtaxed.

Generations Based on Uploaded Images

Here are some examples of what NovelAI produced when given an uploaded image to replicate.

First, using a photographed model.

And here, using illustrations of the character Miranda Lawson, depicted in various styles. These also reflect NovelAI’s struggles to replicate details, such as Miranda’s iconic outfit.

CGI Miranda

Original by Hutt-Slayer

Original by SquishComplex

Generations Based on Names of Specific Individuals

Next, here are some examples of what NovelAI produced when given the name of a specific individual. Generally, these were generated either with only the name itself as a prompt or, sometimes, with a few additional prompts (to get a bit of that fun pudge!). In many cases, the result turned out more recognizable than I anticipated.

Emilia Clarke

Liv Tyler

Megan Fox

Jennifer Connelly

Kirstie Alley

Molly Ringwald

Taylor Swift

Intrigued by these results, I decided to see what NovelAI would make of “Allison Stokke” as a prompt. Ms. Stokke, and particularly her notorious freshman year pork-out, were substantial sources of inspiration for my work on “The Slowest Champion: Revisited.” To my surprise, the AI picked up on her just about as well as more typical “celebrity” types.

Amusingly, here’s what the AI came up with given just the bare bones prompts of “Allison Stoke, pole vault, college, freshman 15.” Perhaps that embarrassing reputation she won for rapidly chunking up has filtered into even her AI associations? I like to imagine so, anyway!

That image still seemed a bit “generous” relative to where Ms. Stokke wound up nearing the end of her freshman year, though, so I felt compelled to help the AI remedy the situation. ;)

An Attempted Image Sequence: Kara from “The Slowest Champion”

Using all the things I’d learned, I tried using NovelAI to illustrate some images of the character Kara based on various scenes from “The Slowest Champion: Revisited.” This proved to be quite a time sink, and I’m not sure the results were worth the effort. Taken as-is, though, maybe it’s decently appealing? Here’s a collage, arranged in an order to roughly track with the story. You be the judge. Bonus points if you can identify which images relate to which scenes from the story!

Closing Remarks

If you’ve got thoughts you’d like to share about AI generated content generally, or about how Mav and I could (or shouldn’t) use that type of tool in our work, please do feel encouraged to share via comments on this post.

Thanks again for helping to enable another fun and productive year of creating kinky stuff under the MAVRIP banner. I’ll check in again with you all soon in the new one!

-Riptoryx

Google Drive link to download zip file of additional AI-generated images: LINK