I can't believe it has only been a little over a year ago since we ran that silly Dall-E event. Having people submit their horrible pony abominations with technology that felt like it was another decade from actually going anywhere felt so harmless back then. Surely this was a nothingburger right...?
Turns out we were so completely wrong.
It only took a few more months before we started getting flooded with actual reasonably good art. This is what sparked those early artist protests, as they realized their work was used to train the models. Especially when our robot overlords were sneaking little broken signatures at the bottom where you'd usually find an artist signing things.
AI had a "style" though. There were people breaking away from it slowly, but for the most part after looking at enough you could usually tell. It also had a cost to play with it, via something like NovelAI's $25 a month subscription, or required good hardware in the case of running a local version of Stable Diffusion. Either way, there was a barrier to entry. That really isn't the case anymore here at the tail end of 2023. Microsoft and OpenAI have dropped Dall-e 3 on the world, and as someone who has to sort through hundreds of drawings a day, it has become a huge struggle to separate the human art from the computer.
Today I'm going to go into the problems of dealing with this tech, give examples of some of the chaos the submit box is currently dealing with, and essentially bring everyone up to speed on what to look out for... until the next breakthrough happens in a month or something. Continue below.
I learned pretty quick as we went through the Stable Diffusion era that I was falling way behind on keeping these out of our human-created art galleries. Outside of a few days of fooling around with NovelAI and SD, I didn't keep up with the tech, which turned out to be a mistake. A good amount of submissions were sneaking into the Drawfriend posts. One or two even ended up as headers.
When I am buried in work sorting mass amounts of things from internal review, it's pretty easy to zone out. 4558 days in a row and counting of doing the exact same thing every day will do that to you. Outside of missing a questionable crotch line or scrolling too quickly through the old compiler accidentally grabbing an extra NSFW version of an image, it usually turns out ok. I quickly realized AI art was another thing I was going to have to train my brain to detect when reviewers I keep on hand aren't around to help with the flood.
I downloaded Stable Diffusion (my poor GTX 1080 needed a lot of assistance here), subscribed to NovelAI, and went on a deepdive. Turns out both of these were capable of a LOT more than what people were doing with them. Luckily most of people producing it aren't artists and can't see the glaringly obvious major errors, and almost never step out of the usual "AI" style, so the submit box wasn't too bad to manage. Some slip through these days if someone really edits things, and new fine-tuning for vastly different styles sneaks in, but for the most part we went from a several a week in Drawfriend to maybe one. I'd also allow them if they were heavily edited by an actual artist, as some were using them as references or bases to speed up the process a bit, but this kinda died out in general.
Dall-E 3 has completely changed everything though. No longer is there any cost at all to entry, and "prompt engineering" isn't required to generate something good. Space Luna up there for example was a simple as:
Princess Luna Pony, laying on her side in the seat, teal eyes, thoughtful expression, steaming coffee mug on the dashboard, On a futuristic space ship in the cockpit, looking down at a beautiful colorful planet in the window
It still makes silly mistakes quite a bit, but it also generates 4 at a
time in 15 seconds. Space Luna up there for example has a detached wing
and wrong Cutie mark. The BIG problem here is, artists make these same mistakes. Constantly. If I had a penny for every wonky hoof or misaligned horn I've seen over the years, I'd be running EQD from a Hawaiian mansion. These are things I, and others don't even bother with in review anymore. If the overall piece looks good, we run with it. Hardly anyone will notice anyway.
That's what makes Dall-E 3 so scary. Luna might have a double wing in that first image, have weird legs in that second, and have tiny hooves in the 3rd and 4th, but at first glance your average art viewer doesn't notice these things, and some are just stylistically what an artist might do anyway.
And all of these took about as much effort to generate as scrolling through an art site. Especially if you are a night owl like me and play with it at midnight when there are no queues. Bing delivers 4 in 15 seconds every time you click the button. It's almost like a freaky slot machine that costs nothing. It's Addictive to see what it will come up with based on your prompt, and if all four are derpy/aren't good? You just hit it again. I can't say I wasn't enjoying the art loot box. It probably doesn't help that every fandom I'm involved in, from pony, to furry (don't even get me started on how easy it is to generate that), to ancient video games I still play like Everquest and Ultima Online have discord servers with channels spamming DALL-E 3 all day.
That's the struggle here on EQD. We are getting loads of these with DA and Twitter links from brand new "artists". Some upscaled with other AI and modified to escape obvious resolution, others generated in ChatGPT, which uses the same DALL-E 3 generator while allowing any shape you want. Here's a piece that was sent in that almost fooled me:
Aside from the badly photoshopped cutie mark and bench with an extra leg, it seems like anything you'd find on DA. Again though, these minor things are so difficult to catch when viewing hundreds of images.
Lets talk about style too, because that's a big one with DALL-E 3. NovelAI and Stable Diffusion are capable of just about any form of art, but it's an extra technical leap to really reign in the models for something unique, usually with a trained LoRA. This is not the case at all with DALL-E 3.
It can do pretty much anything, from cartoony, to watercolor, to Japanese watercolor, to Art Nouveau, and all it takes is telling it to do it in that style. No special techno wizardry required. This is where we go from "I can handle this" to the verge of giving up.
I don't know how it will be possible to police this work in the future. Tiny errors are out, because witch hunting artists for minor flaws is ridiculous unless it's an obvious extra wing like that Luna above. The style is no longer "that ai style" so recognizing it is out. DALL-E 3 isn't perfect, but what about in 5 months? A year? In July of the last one this was what we had when it came to easy, free pony generations. That's technological progress at a chaotic rate.
And these are just the low effort submissions. My first time trying it, I casually threw in "Princess Luna drinking milkshake in cafe". Someone told me to add "Pixel art" and out popped the right image. People with photoshop or general art skills can remove any derps, fix the cutie marks, and come out on the other side with something that looks perfect. Either that or just spin the slot machine 20 times and you are bound to have some near perfect ones even for complex poses and scenes.
I guess as I've been writing this, I've started to realize how almost hopeless it all is to stop the people out there pretending to be actual artists and fooling thousands. There are Twitter and Deviant Art accounts right now posting AI art with no labels at all and an armada of followers who think they drew it, and because Twitter is such an awful discussion platform, calling them out does nothing.
At the end of the day all we can do is educate. This technology is here and the companies that are producing it have thrown caution to the wind in order to compete with each other. There are legal battles happening, but trying to stop this would be about as successful as trying to stop people from pirating music. The tools are open source and already out there, and even if the entire internet miraculously snuffed them all out, the super computers producing the base models now will be consumer grade and sitting on your desk in 10-20 years.
Hopefully this has been eye opening so you know to be a bit more wary when commissioning brand new people or following someone. I'd love to trust my fellow human to be honest about how they created something, but that trust is already out the window as I see another sitting in the EQD submit box as I write this. Sometimes "don't care, more pony" isn't the best response, and I wish people would at least label it AI.
One thing I want to stress more than anything is, don't attack artists over this. I absolutely despise how shitty people are being to them over on Reddit when it comes to AI. It's frustrating seeing your work trained on the thing that the corporate world says will replace you. Personally, I think people will always commission their favorite artists anyway. There's a lot more to it than just receiving pretty art. Knowing a fellow person put work into something beautiful is special in a way. I love me some Luna. I generated hundreds as I learned what DALL-E 3 is capable of. But that beautiful Magnaluna art I have as my phone wallpaper triggers so much more emotion than something made by a machine.
A lot of good for humanity can come out of LLM's, from medical advances to rescuing people from bland repetitive computer work. You might even get your cyberpunk moon-horse wife future. It's important to be aware of what they are fully capable of though. Support real artists, be wary of fake artists but also do your absolute best to research their work before calling them out as fake. We are about to enter an era where images, video, sound clips, and basically anything else digital will be easily fabricated by anyone with the ability to type a sentence. I'll tell you all right now, I have 0 faith that EQD will be able to keep up with this stuff and filter everything perfectly as the technology continues to advance. We will try our best to keep Drawfriend posts human, but simple DALL-E prompts already look like stuff straight from Deviant Art. In the end though, the more prepared we are as a society to deal with the rise in machine generated media, the better.
