(NSFW) Some art and commentary

Time for a potentially controversial post. While I've been taking a break (before a proper holiday) I've been messing around with AI art generation the past few days - I'll be returning to writing for the next couple of weeks, though (I want to send some alpha chapters for Neural Wraith 2 out soon and finish Unseen Banking). This was both personal and research for Neural Wraith (as I look into new tech stuff to get ideas for the series).

As a challenge, I tried to make AI art for Demon's Throne, so I'm posting what I created. Due to the ethical dilemmas around AI art, I wanted to make this post public but it has NSFW content so it must be patron-restricted. You can treat all these images as public and repost them elsewhere, however, if you like.

If you just want to see some interesting art and don't care about my mixed feelings and commentary on AI art generation, feel free to check out the pictures (there's a short list below of what each are) and leave a comment below. For those with burning opinions that can't wait, I'll simply say that while I'm not terribly comfortable with how AI art is being created and how it will be used, it's probably inevitable, like most technology.

The art, in order:

• Vallis, portrait, semi-realistic digital painting style
• Alsia, thighs-up, heavily stylized painted style
• Fara, portrait, heavily stylized painted style
• Maria, portrait
• Grigor in human form, portrait, a terrible attempt at creating him
• Alsia, NSFW, naked, older anime style
• Mina, NSFW, naked on the bed asking for attention, western anime-ish style
• Fei, NSFW, riding a dildo, western anime-ish style (yes, she's not from Demon's Throne. Sue me)
• Mina and Sarae, NSFW, double blowjob/titjob, modern anime/pixiv style

As you can see, there's a variety here. Each of these pictures took between 45 minutes and 2 hours to make using Stable Diffusion with a variety of models. NovelAI recently released an image generator (which is making waves due to the spam of AI-generated anime art it can easily create) but I didn't use it.

Commentary

I started this experiment because I saw the thicc Victorian catgirl images blowing up on twitter, and wondered how easy it was to actually create stuff like that. Unlike previous AI image generators, Stable Diffusion runs locally on your graphics card, is free and open source, heavily customizable, and very techy. So, it kind of leans into my personal preferences - especially given the questionable nature of AI art, the fact this isn't some paywalled project was a plus.

The end result kind of blows my mind. At first, I struggled to produce even decent looking stuff and it paled in comparison to what good artists can make, let alone make it specific. By the end... well, you can be the judge. The double blowjob piece was the last one I made, but the pieces aren't necessarily in the order I made them.

Each of the images have flaws. Some are obvious, others require you to look closely. None of them have hands (well, except Fara, and it's not a great hand). We're safe from Skynet for now. It doesn't really understand hands.

Before I jump into talking about the ethical dilemma, I'll address the elephant in the room.

No, I'm not trying to replace my book cover artist(s). Literally every time I've brought this up to someone, they bring this up. I have a big backlog of book covers and haven't really had issues getting book covers done. While it can be annoying finding new artists, I've had a pretty good run with my current one and I get some fantastic covers with elements I might not think to include. I'm paying for the expertise, and I don't see that changing anytime soon.

I'm also not really that interested in interior art for the books right now.

However, there is one part of art commissioning that has a huge level of friction (well, two parts): character art, specifically NSFW stuff.

My suspicion is that merely bringing this topic up will cause some other authors to be upset (if this gets talked about more broadly), because there are a number of them that attract people to their Patreons by commissioning NSFW art (and putting in the effort, money, and sometimes duplicity), and the idea that AI NSFW art can circumvent that will feel cheap.

This unfortunately falls into the camp of author ethics, which I dislike talking about. There's a Tom Scott video where he says that "Everyone draws the moral line of what's acceptable just slightly below what they're actually doing." That describes selfpub so well I don't really need to elaborate. I've seen authors react extremely negative to AI writing assistants (which I'll discuss in a moment) and AI art, but I also know some of these same people use certain selfpub practices I consider pretty scummy. That's about all I'll say on that topic - remember that authors aren't good arbiters of morality, and that it's a pretty personal thing.

But commissioning character art sucks for both sides. Many artists stop taking commissions as soon as they get an alternative income source (Patreon and streaming are popular choices, and full-time artist gigs have their own caveats) because they hate dealing with commissioners. As a commissioner, you spend time looking for artists, finding out they don't do commissions or have a 1yr+ waitlist, have extortionate prices, or are kind of... not great with details (if you're a self-employed artist, you don't necessarily want to follow your client's details to the letter).

In modern corporate lingo, this is called "friction." The friction between both sides of the transaction reduces how often it occurs unless there's a driving need on both ends. Many artists are only doing commissioned art (especially NSFW) because they hate the alternative, and commissioners really want what they're commissioning (or are trying to make money, which sometimes results in them forgetting to tell the artist what they're using it for).

Enter AI art. You can create your own art, on your own computer, using your own input. Just type in a text prompt and you get amazing art out. Art has been "democratized." Or, well, that's the sales pitch.

The problems with creating AI art have been covered in excruciating detail in tons of articles. I recommend checking those out if you care. An artist whose art has been heavily used in Stable Diffusion, Greg Rutkowski, has done a lot of interesting interviews and commentary and I do recommend considering his point of view. I'll keep things brief.

AI art is produced by scraping public repositories of art. This includes metadata tags, such as those included on boorus such as danbooru and accessibility/SEO metadata that some artists include (like the aforementioned Rutkowski). The AI models then "train" on this art in order to "learn" it, which they are then able to do in new variations, effectively creating new and unique art.

This comes with lots of problems. Mostly ethical. Legally, this is basically unlitigated. People call it a gray area, but the reality is that nobody knows what the courts will decide. There's like one case on AI creations and copyright, and nothing on the legality of using publicly available (if copyrighted) information to train an AI. Given many algorithms have been trained on copyrighted data, I wouldn't bet the house on this stuff being made illegal without specific laws being enacted. There are other areas of AI generation that are more likely to have problems, but AI art isn't one of those.

But that's the legal argument. Ethically, AI art effectively weaponizes artist portfolios against them. It also brings up the automation arguments that have been bouncing around for a while. Again, this is discussed extensively elsewhere. Automation will cost jobs. Neural Wraith is literally a cyberpunk harem series about automation (that's a funny sentence). I address it in a balanced way, because the topic is complicated.

My only real take here is that AI generation is going to be inevitable. Short of laws stopping it (which, again, I wouldn't bet on if you look at some of the companies developing this stuff). I mentioned above that morality is a personal choice, but that applies for lots of reasons. I've largely avoided Facebook my entire life after High School, but that had no effect on the company and I was basically forced to create an author account. I'm usually not a fan of "individual action" doesn't matter, but sometimes it literally doesn't.

I mentioned AI writing assistants above. I looked into NovelAI (who just released a high quality, if very limited, anime image generator) earlier this year. Not because I want to get an AI to produce my books for me, but because I knew that other authors were already using it (and some were really mad about it).

My personal view was that AI writing assistants simply aren't good enough for the type of books I write. My dialogue and plots rely too heavily on background events, worldbuilding, and complex developments. But if I was writing something simpler, and in a simpler, modern setting? I absolutely could use one, right now, and write a lot faster. Make far more money, too.

When AI writing improves, will a time come when selfpublishing becomes difficult if you're not using one because a book every 3 months is too slow, because everyone else is using AI to publish far faster? If the choice is between being an author and not, what happens to "morality" if I personally have a distaste for replacing my writing with an AI's? It's a much harder decision.

And what happens when the AI gets good enough to write novels without an author? Honestly, nobody knows.

Did I Learn Anything?

Honestly, the AI image generation kind of blew me away. The first couple of days involved me banging my head against a wall as I struggled to make anything decent.

The fact is that the popular perception of AI art as "enter text, get amazing art" isn't really true. Every single image up there was heavily modified. I didn't just type in text. I created a pose, iterated on that image to get a better image, then modified portions of that image using a process called inpainting to add specific details (such as Fei's sapphires). It's a frustrating, often painful process (and funny, as the AI creates some ridiculous stuff). But then you sit back and realized you created an entire image in an hour or so, without any of the artist skills necessary.

With that said, it's only a matter of time before the AI improves enough that those steps aren't as difficult. Stable Diffusion is already a massive step forward from DALL-E2 and Midjourney, and the fact it's open source and can create NSFW art is huge.

NovelAI's new image generator showed this off this week. It's much better at creating good art, good poses, and uses stock backgrounds instead of AI monstrosities so it's easily the best character creator around right now. What it mostly lacks is good customizability and flexibility in art styles, but that will likely come with time.

Which leaves me with very mixed feelings. I started this experiment expecting to find that AI image gen had improved, but still had a ways to go. I ended it while creating art that I would genuinely be happy to find on pixiv, and it's tailored to my own characters. That is possible right now, in under a week, and I can produce more picture like that whenever I want, in like an hour or two. Less, as both I and the software gets better.

So I don't know where this or I go. A lot of my professional career involved automation, or selling automation solutions. I don't quite have the same ethical block for it, especially as it's not affecting my current behavior. I'll still commission my book covers, and I'll still support a range of artists on Patreon/fanbox/fantia.

But I'm likely going to still create this stuff for personal use at minimum. The fact I can create scenes using my own characters in the same time it would take to find an artist, gather references, and describe it in an email/discord message is kind of nuts. I don't know if I'll share much more of it, or how.

Anyway, that's my long, kind of rambling commentary. I'm definitely interested in thoughts on the topic. Hopefully this doesn't result in people screaming for my head or something.