You've probably heard mumblings of "machine learning" for the better part of the last half a decade now, and filed it away in that part of your brain that finds something mildly interesting but way out of your skillset. I know I have. While most of us saw Google's quirky deepmind images in 2016 and thought nothing of it, others saw the beginnings of something revolutionary and ran with it. We now see startups, hobbyists, and major AI focused companies popping up all over the place.
Stuff like Talknet and 15.ai have been astounding people with pony voice work, while AI Dungeon and Novel AI take your Equestria prompts and write entire stories for you. Now we have the next big thing - art. A medium I doubt anyone really expected to make so much progress so quickly.
I dove into the Purplesmart.ai Discord server to see exactly what is going on in this crazy new world.
Read about it all below!
I bothered the leader of the Purplesmart.ai Discord server for a brief history lesson on how this new format came about and they ended up saving me a bunch of time with a complete timeline. Here's how we went from no AI pony, to the image you see above in almost no time at all!
In 2015, Elon Musk and a host of other investors pledged over a billion dollars to starting up a company called OpenAI. From there, the first major releases happened with something called GPT, (Generative Pre-trained Transformer), a machine learning model that used deep learning to try and create text responses as human-like as possible. If you've used AI Dungeon, you've probably run into this series, as their Dragon, and Griffon options both used later versions of GPT.
While the first versions of GPT (known as GPT-2) were released publicly OpenAI decided not to follow up with bigger and more capable models making them into a commercial product hence an open source version of larger GPT models was formed by a collective of researchers under the banner of EleutherAI. This lead to an AIdungeon competitor, NovelAI.
From there, we venture into the world of DALLE, literally just a few months ago. We even ran an event for it here on the site, though that was using the public Dall-E mini, which was nowhere near the power of the actual AI image generator that was available internally.
Like AI Dungeon and NovelAI, art needed it's open source option too, and we ended up with what is generating all of these amazing ponies you see popping up on various websites. StabilityAI unleashes Stable Diffusion on the world, giving your everyday programmer the ability to fool around with this new tech that is nearly as strong as what is internally available with the full version of DALLE.
You now see both NovelAI ponies, as well as our in fandom training models via PurpleSmart.AI using Stable Diffusion, both of which are making more and more progress every single day.
You have two easy options and a few complicated ones if you want to give this new technology a shot for yourself. I'll just be covering the primarily pony focused easy one at the Purplesmart.ai server.
After you've joined, they have separate channels for regular pony, anthro pony, furry, anime, and more. Each one lets you enter a "/dream" command followed by a prompt which will attempt to generate something based on what you tell it to do. Until you know what you are doing, it's a good idea to either read a detailed guide linked in the #start-here channel or just copy someone else's prompt and swap the inputs to what you actually want to create as there are a lot of small variables in there that have already proved successful. Things evolve pretty rapidly and people are always discovering new ways to "talk" to the machine, but this is the most basic way to get going. Chances are you won't be generating the images you see here on your first or even 10th try, but that's how this all works out. The machine is learning, and you are getting to know it too in a weird, cyberpunky way.
The other easy option is the paid service NovelAI, where you can also go on some pretty epic text adventures in Equestria if you are so inclined, or use it to help write your masterwork in horse fanfiction. Their Discord server has all of the information needed for generating images on that side. NovelAI is a subscription service though, so you do pay for that extra clean interface and pony adventure.
For the more technologically inclined among us, you can use your own hardware to join in on it. Their models are open source available over here. A graphics card with 4gb+ of VRAM is suggested, though people are doing much slower generations on weaker hardware. The more VRAM the faster it is.
Regardless of what you choose to do, be it text and image adventures in NovelAI or generating painted horses with everyone else in Purplesmart.AI, there will be a learning curve. This technology is still relatively new. From my experience in text adventures on some of these, learning how to teach the machine what you want is a process in itself. Especially when you need to convince it you are surrounded by 4 hooved anime-eye mares who very much do not have hands.
There is no doubt about it; this is a disruptive technology. The furry fandom in particular has been dealing with an influx of AI driven problems over the past few months as more and more people flood their fandom with high quality versions of their original characters that they may have paid an artist to create for them otherwise. Over on reddit, the Artist Lounge sub has posts every few hours with an artist panicking about their livelihood being taken away by a machine in the coming years. The last twenty years have been amazing for artists around the world who suddenly have an entire market of people needing art for everything from their indie game to a personal commission of their favorite pony. Not to mention the professional sector and their ravenous need of digital artists. This has given a lot of people freelance and stable corporate jobs that they worry they may lose.
Right now, most places in the fandom are doing their best to be respectful of the amazing artists that have kept us going for so many years. On Derpibooru, they've set "machine learning generated" to automatically be a hidden label and require actual effort to be put into anything uploaded.
