Heretic Spellblade - Character Art (plus more commentary)

A batch of SFW character art for some of the champions of Heretic Spellblade. Yes, they're AI generated. If you're just here for the art, the list is below and the pictures should be above. Otherwise, there'll be some commentary attached. I won't really go into the ethics, as I talked about that last time. This is more about what's changed since.

Note that the canon appearances remain those of the book covers. Except maybe Fei (and Sen, as I know a lot of people have their own headcanon appearance for her).

Characters in order, all depicted as of Spellblade 5 (before the final battle):

• Fei
• Ciana
• Narime
• Sen
• Seraph
• Sunstorm
• EDIT: I added a second Narime

I tried to keep everything to a fairly consistent style, both as a test of the updated models and also because it's fun to do so. You'll notice that the most consistent art style revolves around the girls with animal ears, and there's a good reason for that I'll get to. For reference, no artist embeddings or artist tags were used in making these. I didn't steal Greg Rutkowski's soul. This also doesn't take away from any artists, as I don't really commission non-cover art.

This wasn't a post I planned to make. The situation remains controversial and this platform has refused to take an actual stance despite making some moves against it. So I won't really retread all the ethics stuff. Check out the last post if you care (warning NSFW).

But other authors are happily plowing ahead, even using it for cover art. AI art remains incredibly easy to spot, especially if you're behind the curve and using older models.

At the same time, other AI stuff is bursting onto the scene and making far less ruckus. Apple launched AI audiobooks, not long after Google, and they're improving rapidly. Microsoft's Azure TTS is commercially available and sounds really impressive, and will be used widely in companies very soon, bringing pie-in-the-sky IT projects I once worked on to fruition. Machine translation is being rebranded as AI translation, and I'm seeing it used much more these days in paid products (I genuinely can't imagine being enough of an ass to charge for MTL). I even had a conversation with some friends about ChatGPT: one is using it to help him with his programming, and the other plans to use it to write his papers. Ironically, both are against AI art.

The lack of buzz over the other AI stuff is the principle reason I decided to post this recent batch. I have my own opinions and wariness over AI, but other, brighter minds have already summed things up: it will destroy industries and jobs, but you can either adapt or remain an idiot. Such is life. I try to depict all the automation issues in Neural Wraith with an open mind because technology isn't black and white.

What's Happened Since Last Time

Anyway, let's talk about the stuff that's changed in the AI art space, in case you haven't followed it.

When I last posted, producing anything good took a lot of painstaking labor. Hours per picture if you wanted something coherent and high quality, due to the amount of inpainting and trial and error required. (For reference, inpainting is how you modify part of an existing image with AI).

Then NovelAI (NAI) arrived and changed everything. It could reliably generate accurate pictures of anime girls all day long. The internet got flooded with AI anime art, particularly on all the art sites, and that's where all the anti-AI art drama started.

My prediction of a new level of mediocrity ended up being largely true. While you could use the NAI model to produce some very interesting stuff, it involved the same level of work as before, but for even better results. Most people wouldn't do that. Hell, most of the AI book covers on Amazon are that level of laziness, with maybe some Fiverr artist editing the hands for $5.

But there's a new wave of models coming out. There are newer, better systems to add to the "library" used for the AI, adding new poses, art styles, and concepts. The generation method is improving, making it possible to generate far higher quality images with more reliability. You can spit out highres images all day long with the new models, whereas it took a lot of bullshit before.

Honestly, given it took me like an hour to produce the base for the Fei image after ignoring AI art for close to a month, it's kind of sad to see people still spitting out the bad stuff online.

The images above still took a little while, but I created them between doing other things. The base image rarely took more than five to ten minutes, and sometimes less. Even taking that long was because I was looking for certain qualities in each image, before working on it further. You can spit out amazing stuff a couple of times a minute.

Weaknesses

But there are still a bunch of weaknesses. I mentioned one above.

Style remains a bit of a bitch. Ciana, Fei, and Narime all share a largely similar style, because tags like "animal ears" and "fox tails" exclude a lot of non-anime art. This pushes the model's art style away from, say, the 3D rendered look popular with Riot or Blizzard and closer to anime. Seraph benefitted a little from her clothing. But both Sen and Sunstorm look so different because they lack any stereotypical anime attributes. Hell, the fact they have smaller breasts made it difficult to make them look older. Ironically, that is probably the fault of anime.

This can be rectified with a style embedding/LORA (a LORA is a new thing that makes it possible to add concepts/characters to the AI art model). But a style embedding comes with deep ethical issues, as it almost always involves using the style of a specific artist. By contrast, the telltale "NAI look" so common to AI anime art is the result of how NovelAI trained their model, and is best compared to the way that an amalgamation of faces ends up looking incredibly generic.

Hands are also still a disaster, but they've gotten better. I actually spent a fair bit of time on each picture trying to ensure each character had "passable" hands, rather than eldritch horror stuff. The current state of hands means there are going to be a bunch of people selling AI art on Fiverr and just redrawing hands.

Getting a very specific character still requires actual work. Depending on how much or little you give a shit about people going "man, those hands look like shit" or even leaving out basic details of the character because it involves effort, the process takes from thirty seconds to two hours. I put in the effort because using a text prompt and half-assing everything else is boring as shit, and I enjoy the technology itself.

Photoshop is still necessary to force in certain things. Ciana's unicorn horn took a fair bit of nudging. Most of the gemstones could be forced in, but there's no consistency and none of them are in the collarbone. An actual artist can clearly do this better.

Finally, AI art is still really obvious, despite all the improvements. Overexposure to AI art allows you to pick a lot from the thumbnail. But even the better stuff have a bunch of tells, and not just hands. Patterns tend to be nonsensical, and details tend to bleed into each other (which is why painterly styles are popular for AI). If you see something in a high enough res, you'll also spot too much detail - actual artists can't paint with the precision that computers do. Humans create the illusion of detail when you view something in a smaller resolution, and the images often look a bit shit when zoomed in, but AI art is too perfect.

Oh, and while backgrounds look really cool, AI art still has issues with recursion. Buildings and castles are the hands of backgrounds. Mountains and trees can recurse, but most buildings need some level of logic.

Using It Commercially

I'm being super critical here, because the reality is that the more I play with AI art, the more obvious it becomes that there is a massive gap between it and human artists. Will it eventually catch up? Almost certainly. But I suspect some of these issues are harder to solve than others.

Also, AI art really fucking sucks at making anything for Neural Wraith. I think the issue is that most of the booru tags for scifi stuff overlap with fantasy and are drastically outnumbered, due to anime's preference for fantasy (especially in the last decade).

So lets take a step back. I've said before that I have no real interest in using this for book covers yet, and that remains very true. Everything I've created still falls short of my current artist. She knows composition, color balance, and can include all the stuff I need her to if I ask for it. Stuff like the regalia and armor consistency between Book 5 and Book 7's cover art in Spellblade is huge, and basically impossible with AI. Plus, I don't need to spend hours doing it myself. At this point, the money for the cover art isn't a problem. I value having high quality cover art, which is why I've splurged a little for the full jacket art for the Neural Wraith series.

(And, yes, there's the ethical argument, but I'm avoiding that. Mostly because it annoyed me to have somebody lecture me that I need to hire artists because it's the right thing to do when he planned to use AI to do his own job in secret. You can't argue from a point of principles if people don't share them.)

But if I were a new author, I'd definitely be using this. If I created that Fei image and was preparing Spellblade 1 right now, it'd be my book cover. I sure as hell don't begrudge the newer authors for using AI art. Especially if they put in the effort to create nicer stuff. There are ways to create stuff that leans more toward the rendered or semi-realistic artstyles, too, while retaining this level of detail.

As for bigger authors... Well, if you're doing well enough that cover art expenses aren't that important but still use AI art, then they're not the type of people to give a shit about what anyone thinks, anyway. Part of the point of being indie is to do your own thing. Not everyone cares about good cover art. I hate marketing, and it's always been amusing to see the veiled grumbling at me for not doing it. Personally, I'll judge them if they produce shit cover art with AI, given how easy it is to make better stuff.

At the same time, community reaction is what it is. AI art caused drama because so much mediocre garbage flooded the internet. If people get upset over it, then that is what it is.

This is, ironically, where traditional publishers have gotten in trouble. They often produce extremely generic covers at absurdly high prices (several times what indies pay), and AI can spit out way better stuff. Community reaction might keep them at bay, or it might not. It was kind of funny to see the major publishers steam ahead with blatant AI covers while the indie community was umming and ahhing over whether to use them or not.

To sum up, nothing has really changed. As expected, AI art has improved a lot, but hasn't gotten over a lot of the biggest stumbling blocks. People are still mad about it, but there's so much AI stuff progressing at once that it's almost farcical to pretend it's not inevitable. I don't really blame people for being worried or upset about it, although I am still concerned about the degree of misinfo and vitriol used.

As mentioned last time, I'll still play with this stuff myself. I need to write Spellblade 6 and do other IRL things, so it's unlikely I'll play much with it until April.

Let me know your thoughts, including if you'd prefer not to see this stuff. There's been a lot more discussion and publicity since I last posted on it.