Since that discussion was so busy yesterday, I figured I'd respond to some of the common things people brought up in it with a separate post. If you missed it, we asked for suggestions on filtering AI art submissions yesterday, which has been a growing problem for a while now.
So, head on down below and I'll dive into this FAQ style!
Using AI to Battle AI Art with Detectors
A lot of tools have popped up around the internet with the promise of detecting AI in artwork. I went down the rabbit hole testing loads of them on some of the art we receive every day, and the false positives were pretty common. There was one called Hive Moderation that would do better than others, but it would still fail quite a bit, particularly if the backgrounds were simple:
It has improved a quite a bit from recent testing, so it's possible these will get to a place where they are reliable. The chance at detecting a new artist as AI isn't really worth it though. I'd rather some AI slip in than ban someone for a style that looks AI and is detected as AI but isn't AI.
In the end, this will be a never ending arms race. Detectors will win for a bit, then a new model will drop that beats them, and we will repeat the cycle forever.
Allowing AI in separate AI galleries
EQD is built on human drawn art when it comes to our Drawfriend galleries, and I think it's important that we keep celebrating that. We might have goofy Dalle style one off events in the future, but as for making AI a regular feature here, I'm not interested in that. It's fine if you think that Cadance up there looks good. I think she does too. But I'd rather stick to real art on this site, just because it feels silly to feature something I know anyone can generate in a few minutes if they have a good enough graphics card.
That being said, we have no problem with using AI tools to augment human creation. BGM for example in our music category is regularly featured, and uses AI to modify his and others voices to ponies, along with 15.ai back when that was a thing to add vocals. This is just part of his overall creative process though, and he still puts crazy amounts of time and skill into producing a new track.
You could compare it to something like Suno. We wouldn't take a random Suno generated song, just like we wouldn't take a random Stable Diffusion generated piece of art.
Require Art Files/Timelapse/Etc
Honestly, EQD doesn't really have the time or resources to do deep dives on everyone's creative process, nor do we want to. Not to mention there are dedicated AI out there now that fake this entire process anyway. They are pretty obvious, but like everything with this technology, in 6 months they will probably be hugely improved. Just look at some of the recent video AI stuff being created. It was an absolute joke just 5 months ago, now we have almost cohesive pony videos floating around.
This might be something we explore for artists that come in with massive galleries out of nowhere, but most AI prompters who are actually nefarious are smart and only upload like an artist would. They post one or two pieces a week, and use very unique LoRA's trained on artists that haven't drawn anything in a decade. That's a tough one to deal with.
Ban People that Fake It
Over the past few weeks, someone has been spamming the submit box with Christmas themed AI ponies. They have somewhere around 30 email accounts, and probably run on a VPN. Banning people on the internet is about as effective as pissing in the ocean. It's not gonna happen. I'm sure this guy rotates assuming we already blocked his previous emails, but I don't even bother at this point.
I'm not going to rage about slightly wonky ponies in Christmas outfits, but I also accept that nothing can be done about it and deleting an email takes a second.
Here you go random person. Your Luna is posted. You can stop now. You win.
Hope for Laziness and Obvious Errors
This was pretty much the go-to at the beginning. Back when I was complaining about this a year ago with Dall-e. We've come a LONG way though, and models like Zoinksnoob are really shaking things up with their accuracy. There is very little artifacting going on that we saw on previous models. PONY XL could slip in with a style LoRA, but they'd almost always have lots of little errors. Zoinks still has issues, but the more simple drawings like that Marble yesterday really only have a few.
Celestia up there for example. The swirls on her neck piece are a little goofy, and her crown makes no sense, but compared to stuff submitted a few months ago this is on another level. And a LOT of these issues are things real artists do all the time too! That's the big problem. Minor defects in art from anatomical to fine details is just part of it. Whenever I commission a piece I usually go in and clean it up, since 99% of the time the artist isn't going to hyperfocus on minor things like lines bleeding through lines or a piece of jewelry sitting on an ear wrong. Hell, I've even fixed hooves bent in weird ways. No one creates perfect art.
Couple that with the fact that prompters aren't always lazy. Some will take the extra 5 minutes to go in and fix all of the major issues. . It's still an 8/10 pony in record time. There really is no way to combat that one.
The Honor System
It's already being violated to death. Humans gonna human. I'd say about 10% of AI art submitted to the EQD box is actually labeled AI art. The rest is just "I drew this thing plz post it :3" If someone thinks they can get more props by faking it, they absolutely will. If anything, editorials like this are just a challenge for people to try to break the system more. That's why humans are so successful as a species. We poke and prod everything until it breaks, which prompts others to find solutions and improvements. They also have 30 deviant art accounts and will just use a new one when one gets blacklisted.
Couple that with the fact that... a lot of people just don't care. The final product is what they want and the process to get there isn't important. Case in point the Celestia up there I posted to the EQD twitter. Almost all of the comments were just talking about how hot she was. A lot of people just turn their brains off when browsing the web and didn't read the AI blurb, and plenty of them don't mind she's AI at all.
So, What is the Solution then?!
I think a combination of everything might just be how we do it in 2025, though I'll be honest, I don't see anything working as time goes on. I'm still gathering ideas of course, so feel free to drop more in either the previous post or here. This might be a losing battle, but I'll fight it for a while. It's kinda fun to play AI detective anyway.
I've done a lot of deep dives on Civit.ai and keep tabs on AI tags and threads elsewhere to make sure I'm aware of new models and LoRA's people are using in poni. I'll keep up with it as long as I can. At least the 50 series cards coming out soon from Nvidia aren't looking too hot for the prices, though with Windows 10 being sunsetted in 2024 I'm sure a lot of people will be building new rigs, so who knows how bad the flood will get once people hit that 12-16gb of VRAM required to run these models without taking forever.
