Mob Sorcery Pixel Art (and discussion)

First up, these are AI generated. That also means there'll be a lengthy discussion later in the post (which you can skip if you just want to look at the pixel art of the characters).

I could have told you some story about how these were actually created by my girlfriend who just used AI as a base for sketches and actually worked painstakingly for days to make them, but that's obvious bullshit. I don't care to tell such obvious lies, so I'll leave bullshit to the bullshit artists.

Also, this is only in the $5 tier because the final character shown is a major spoiler. Otherwise, these are public, so feel free to share them. There are 23 in total, for reference, so you'll need to scroll.

If you haven't read at least the first chapter of Mob Sorcery 4, don't look at the final character image. I won't name her below for your sake.

Also, I've included a Google Drive folder with all the images in case you don't like Patreon's image viewer. There's no protection, because I don't care if these are shared (they're AI art lmao):

https://drive.google.com/drive/folders/1ESIWnyEPwg-Ag9CPBM-x5tYl5o5AMO-C?usp=sharing

List of characters in order:

• Nina

• Alessia

• Ally

• Anzu

• Ashley

• Bastet

• Fia

• Erik Frost

• Gabriela

• Juliet

• June Marcellus

• Kigenai

• Kiyoko

• Livia

• Lucia

• Mei Suwa

• Momo

• Nicki

• Pola

• Quintus

• Salome

• Wagner

• Spoiler Character

Also, these are "artistic interpretations" of each character. They're based on the cover art images where available (except Ally, whose sundress version come out much better than the kimono one), but you'll notice they vary from those because, well, AI. The cover art remains the definitive look for each character.

From here on in, I'll talk about the AI shit and why these exist. So you can ignore this part if you don't care. I'm not going to talk about ethics in AI or anything, as that topic was done to death years ago and, predictably, it got tied up in the courts with no decision in sight.

I like to keep up with technological developments, and AI is no exception. Part of the reason I could include so many random tech developments and "peeks into a potential future" in Neural Wraith is that I keep an eye on what speculative developments are happening, although that's arguably become a lot harder because the AI hype cycle is crowding out so much regular tech development and reporting lately (much like blockchain did in the mid-2010s).

I haven't touched AI art since my last big post about it, when I did versions of a bunch of Spellblade characters using the then best models available (for my preferred AI art style). That was nearly two years ago. Since then, I've messed around with LLMs (i.e. AI chat bots) a fair bit, and even scrapped a big post on them last year where I would have talked about their accessibility and the poorly kept secret that a fair few authors use them.

I've avoided talking about AI for one big reason. It's a drama generator. People have strong opinions about it. So do I, but I don't fall into the pre-defined internet camps of "AI is the future and opponents are luddites" and "AI is the devil and needs to be destroyed" which tends to overwhelm all other voices. I dislike a lot of the discourse surrounding it, especially the way any attempt to talk about is rapidly poisoned by dipshits saying "it's inevitable" to avoid responsibility in a very Nuremburg-esque way.

... I said I wouldn't talk about ethics, so let's move on.

I've avoided AI art for a couple of years primarily because it hasn't really improved. The underlying technology was already stalling out when I made my last post, as investment shifted to ChatGPT and LLMs, major firms (and nVidia) hired open source programmers that had been pushing things forward, and Stable Diffusion nearly imploded. Also, AI art got much, much worse press than ChatGPT, despite both technologies doing the exact same thing. Formerly text-based LLMs integrated image generation into their AI suites, and the focus shifted to improving photorealistic image generation, which has been improving to a frightening degree.

But I don't care about photorealistic bullshit. 2D art generation had still stalled.

Over two years, I had noticed a bunch of slow developments, particularly in the commercial space. AI art has infiltrated a lot of commercial work, including ruining stock photography libraries almost completely. The overly smooth, shiny, and almost plasticine "AI sheen" of a lot of AI art somehow became desirable, as a lot of people identified it as high quality (it doesn't help that there are a tiny handful of artists who do draw that way, but most cases are people just misidentifying art... or real artists using AI as a helping hand for certain parts, even if they draw most of it themselves). There are numerous cases of artists transitioning to AI assistance, with their art styles changing completely overnight.

The common factor in AI's popularity, whether it be for art, writing, or corporate bullshit, is that it can do something faster and more consistently, even if the quality is lower or it changes the output. If you're an artist whose number of paying fans on Patreon or fanbox depends on your output, then being able to pump out an image every day, instead of every week, will increase your income two, three, or even ten times.

I wanted to check on AI art to check on a lot of assumptions and myths people make about it. Is the AI sheen still there? Are hands still shit? Can it generate more than one person? What about consistent patterns or even generating straight lines? Was I right to assume the growth in odd pixel art games in certain places is due to AI art? How good have the models gotten? And what about loras, which were still in their relative infancy two years ago? Etcetera, etcetera.

Short answer? Unless you're personally deep in the weeds of AI art, you probably have the wrong idea.

Hands have been a solved issue for years. Loras (small, specifically trained additions to AI models) that focused on fixing hands were being used when I made my last post, an extension came out in late 2023 that started fixing them after image generation, and new models in 2024 often generate accurate hands. The one I'm using now only generates shit hands rarely, or if I'm pushing it too hard (typically because it doesn't understand the prompt).

The same goes for stuff melting together. In the old days, hair, jewelry, limbs, weapons etc would melt together given any excuse. If something was occluded, it would often merge with the thing occluding it, instead of emerging on the other side? AI had issues with object permanence.

Again, a thing of the past, mostly. Things still melt sometimes, particularly complicated hair (largely with itself), but AI can consistently separate the various parts of characters and objects. In the Gabriela picture above, it has her horse ear poking up behind her unicorn horn with no issues, for example, and that's a common occurrence.

Multiple people was a massive issue two years ago. The solution at the time was a lot of inpainting. Now, good models and the use of an extension can generate images with six, seven, or more characters in reasonably complicated poses, with backgrounds.

As you can probably guess from the art I posted, pixel art is definitely good, but there are caveats. This is a specific lora, based on a specific art style. It's arguably the best lora I've used. Most AI pixel art is... less than good as it tends to look more like pixelated real art, and I used a few simple tricks to cull the color palette and make the pixels look blockier. But there's still cases of AI artifacting and pixels with more than one color, plus non-square pixels.

Not all models are this good. A lot of people are probably still using old ones, because it suits their workflow. I'm using a finetune of Illustrious, which is the anime model to use right now and compatible with a lot of good loras. Pony and Illustrious are basically the only good 2D art models imo, but they tend to toward anime and cartoon art, and a lot of western preferences are toward the pseudo-3D semirealistic plasticine look that AI art has become famous for. There are probably finetunes that create that look, though.

The breadth of finetunes and loras is... kind of insane now. While the underlying technology has largely stagnated, save for a brief burst of activity with the emergence of Flux (the newest iteration of Stable Diffusion, which doesn't have a wide variety of finetunes and models based on it), the AI art community has created an obscene amount of infrastructure to support it. The NSFW community is probably the biggest, as is the fanart community.

Way too many finetunes. There are artstyles for a ton of artists and games. Character loras for way too many characters of reasonably popular games and anime. Loras for different articles of clothing, poses (sexual and otherwise), concepts etc. The AI models themselves can only be trained on so much, but the loras mean you can add a lot.

For example, when generating the Pola image, her Chinese dress always came out as a long one that hung on the ground. This is because Illustrious is trained on basically the entire Danbooru database, which combines short and long qipaos, cheongsams etc. So I found a lora for a short cheongsam and bam, I had the right dress to (mostly) match her cover art.

But now it's time for the negatives, after all the glazing.

To continue with the clothing, AI art still kind of sucks at being specific. Because it needs to be taught what things are, and clothing is incredibly hyperspecific and detailed, AI is very limited in what it can dress people in. I can give somebody a neckerchief, but I can't hang that neckerchief to the side. Pola wears a form of a cheongsam on her cover art, but it's sure as hell not the version I found in that lora, because my artist can create her own clothing.

AI is fundamentally limited to what it is trained on. It has very limited ability to combine concepts and go beyond that. There have been a few times where I've tried to create poses that are slightly different to the ones that are more popular in the model, and while it tries, it usually starts to melt down. The solution is to create a lora and teach it the new pose, but that reveals the weakness - AI needs to be taught every little thing. If I tell a human artist to take this pose, but draw it from an external perspective instead of point-of-view, they might find it harder than if I found them an example, but they could. In fact, I could find them an example. An AI would need 30-50 examples to train a lora.

The pixel art has a ton of problems, too. For one thing, it's still not real pixel art. It can't really be animated without a real pixel artist stepping in (AI pixel art is easy to spot due to really awkward animation imo), and I imagine they'd find it a pain to work with. There is a method I found last night to remove the crappy not-quite-the-same-color backgrounds, so i won't bitch about that, but it's more time intensive than it should be. I made this for the characters because it looked cool, and don't plan to use it commercially.

Finally, patterns and consistency. It, uh, doesn't exist. In fact, when I spot an artist I suspect has started using AI to brush up their art, they quickly learn to keep doing their own backgrounds or at least fix the errors.

AI will create the fuckiest jewelry, tattoos, patterns etc known to man. Get it to generate a throne and the patterns in the metal will be nonsense, and they won't match on both sides unless it just mirrored it symmetrically (which it loves to do as a way to cheat). I saw a Halloween piece where the pumpkins just looked like nonsense. City skylines will be a mess, with windows not even lining up etc. This has not been fixed, although it has gotten a lot better than the early days.

Ultimately, AI art has improved a lot (if you use the good models and offline generation, and don't hire morons on fiverr). It can generate really high quality stuff, has a lot more variety than it used to, doesn't ghost you, is free (if of questionable ethics), and fast. If you don't want something specific, and don't particularly care about composition or the other neat things artists can do, it's great. Especially as a growing number of people actively like it, and there's a political aspect to supporting it that I won't touch on.

Artists haven't exactly been replaced, though. There are things they can do that AI still can't. And if AI is better at niche things, I wouldn't be surprised if I learned that artists have been using it (although the problem there is that there is still a learning curve to using all the bullshit involved in AI, especially as there's now an insanely complicated UI with more utility that was designed by people who should never get within 100 feet of a UX designer for fear of their life).

I'm not about to make any predictions, though. Maybe AI will rapidly improve. On the other hand, progress on AI has been glacial despite the predictions initially made. So far, AI progress reminds me a lot of the talk about both VR and blockchain. Lots of Ethereums and Quests, not a lot of revolutionary improvements that automate everything to match the pie-in-the-sky rhetoric. Yet.

As always, don't expect me to post much AI art, if any. I know these get mixed responses. I posted some of it in Discord, and got the muted reaction I largely expected. There'll continue to be cover art and the NSFW bonus art from real artists as always. I actually hadn't intended on making a huge post like this, but, eh.

Thoughts are welcome, even if they're focused on the art, rather than my ramblings, or just about whether you'd prefer if I just posted the art instead of rambling.