Human-centric AI news and analysis

This AI spit outs endless new My Little Pony characters

The creations could give Twilight Sparkle a gallop for her money

My Little Pony has a curiously broad set of fans, from little girls to adult male “bronies” and, err, Nazis. But since the cartoon series Friendship Is Magic ended in 2019, there’s been no new adventures of the Mane Six gang for fans to enjoy.

Luckily for them, an AI system called This Pony Does Not Exist is helping fill the void, by spewing out an endless stream of new pony characters.

The site was created by a developer known as Arfa, the same mind behind This Fursona Does Not Exist. Arfa built the system by training a generative adversarial network (GAN) on around 104,000 pictures from Derpibooru, the world’s largest imageboard for My Little Pony artworks.

[Read: Nvidia teaches AI to create new version of Pac-Man just by watching gameplay]

Arfa used transfer learning to fine-tune the final model from This Fursona Does Not Exist on the pony dataset for 13 days. It was then scaled up to 1024 x 1024 resolution, and trained for another 200,000 iterations to produce the final model.

Here’s a taste of what it spits out:

Credit: This Pony Does Not Exist

While some of the ponies resemble the images they were trained on, Arfa says no one has provided an example of an exact copy:

It may be helpful to remember that the GAN isn’t simply copying parts of existing artwork and replicating them. Rather, it is looking at hundreds of thousands of distinct pieces of art and ‘learning’ an internal representation of what constitutes a pony, similar to how a human artist looks at hundreds of thousands of images while learning to draw.

The GAN gang

This Pony Does Not Exist follows in the illustrious footsteps of This Person Does Not Exist, This Rental Does Not Exist, This Foot Does Not Exist, This Meme Does Not Exist, and, of course, This Waifu Does Not Exist. I could go on, but you get the idea.

All of their masterpieces are crafted by GANs, which create new content by pitting two neural networks against each other: a generator that produces new images, and a discriminator that determines whether they look genuine by comparing them to the original dataset. The feedback loop helps the generator refine its designs until the discriminator decides they can pass as the real thing.

The framework can be used in some seriously nefarious ways, from putting celebrity faces on the bodies of pornstars to making politicians say whatever you want. But at least it can also create some pretty cute/creepy ponies. Just check out these two rascals:

 

Published July 15, 2020 — 17:25 UTC