Way back in October of last year, OpenAI released its third iteration of Dall-E, bringing on a wave of scarily accurate pony art in a variety of styles. It got to a point where we were getting flooded with them here on EQD, and had to post up a little guide to help people tell the difference between AI and real art.
Today we focus instead on Stable Diffusion, which is being used to directly copy artist styles via trained LoRA's. This is something that has always been possible, but only really exploded with the new version of Pony Diffusion and its rising popularity as people moved to more complex ways to generate art than Dall-E.
Continue on below as we delve into yet another aspect of the AI art war, and educate all of you on what to look out for now that even your favorite artists are being compromised.
LoRA's essentially give people the power to influence large AI models to specifically target certain styles and characters. Something that would usually take a huge amount of computing power can be done with your average gaming graphics card made in the past few years. It's not as perfect as a full blown custom model, but it comes damn close, which is why this has become a growing issue. We've already received fake submissions from people pretending to be these artists, and over on Discord people already get loads of messages from fake artists looking for commissions with AI art.
Most people use LoRA's to specifically generate a character they like, so early ones focused on stuff like Autumn Blaze here, who's unique kirin features were somewhat difficult to get right without targeted training. You'd use a kirin LoRA for it which plugs right into your prompt.
Unfortunately, people have stopped just training characters in the pony fandom, and started training the actual artists. Right now, the biggest pony specific model, Pony Diffusion, actually blocks and removes artist tags entirely, but with a LoRA this is easy to circumvent while still using the powerful technology behind its MLP generations.
All it takes is a dataset of ~20-100 images and a graphics card with enough memory. A typical GeForce RTX 4080 for example can create extremely powerful LoRA's in just an hour to half a day depending on what parameters are set and how many steps you have it take. People then distribute this LoRA so anyone can generate ponies with that stolen style by simply dropping it in a folder and adding it to the prompt.
Right now the number of artists that have been compromised has reached around 25 that we know of. Most are people that have retired from pony, but a few are still actively posting. Yesterday I slapped a warning up on Twitter and a few people commented that they could hardly tell the difference.
There are of course still loads of little AI errors that I mentioned in that earlier editorial, but like before, your average internet user isn't nitpicking details in art they like. They see cute pony, hit the heart button, and move on.
We need to be realistic about this one. Stable Diffusion is open source and always improving because of it. There will absolutely be a time in the future where there are no more errors to spot, especially as more powerful hardware is released. There's also no realistic way to stop it. Dall-E got slightly nerfed due to people raising arms about certain aspects of it, but there isn't a manager to complain to for SD. You'd have about as much success as stopping music piracy, and the companies behind that have legal teams leaps and bounds more powerful than anything we could muster.
The only thing we can really do is educate. Let people know that this is a thing now. That you might get a DM from someone posing as your favorite artist willing to draw your OC on Discord because they need money for vet bills. Or see prints of brand new art from that same person on Etsy and Deviant Art, who both seem to be doing absolutely nothing at all about the AI spam.
Always double check where the source of something is. Message artists directly for commissions and keep supporting them. Never assume they will message you directly asking for commissions unless this is expected for some reason.
Especially double check accounts. There are thousands of fake Twitter accounts out there tricking people every day. Check their followers, check their replies, make sure it's who you actually want to work with.
The world is changing rapidly with this technology. For all the potential medical breakthroughs and scientific advancement we might get out of it, AI at the ground level is still going to cause all sorts of disruptions like this. Humanity has decided to open the floodgates as major corporations all compete viciously to be the best of the best with this their different brands of LLM. Stay informed and keep your friends educated. EQD might not be the juggernaut it used to be, but I'll do my best to keep you all updated on anything else that comes from these crazy times.
