The Wonderful, Weird World of AI Generated Pokemon

How an emergent field of mathematics can reshape the familiar franchise characters

In mid-October, Twitter user @Mushbuh tweeted out a link to the 3,020 Pokémon they generated with the help of an AI algorithm called GAN, or Generative Adversarial Networks. As a designer and artist, their other works spanacross 3D rendering, stickers, clothing, 2D art, patches, and more. When I first saw their foray into AI sprite generation, I became completely entranced by the empty and abstract vessels – a few of which can be seen above – just waiting to have meaning assigned to them.

pokemon mushbuh 1

Published on December 18, 2018

In mid-October, Twitter user @Mushbuh tweeted out a link to the 3,020 Pokémon they generated with the help of an AI algorithm called GAN, or Generative Adversarial Networks. As a designer and artist, their other works span across 3D rendering, stickers, clothing, 2D art, patches, and more. When I first saw their foray into AI sprite generation, I became completely entranced by the empty and abstract vessels – a few of which can be seen above – just waiting to have meaning assigned to them.

Needing to know more about what I was seeing, I found Matthew Guzdial through a post in the comments. He’s a Computer Science Ph.D. Candidate at Georgia Tech, where he studies Creativity and Machine Learning. From his place as an academic, he voices a sentiment that Mushbuh also shared with me – that utilizing Pokémon and their mass recognition factor serves as a gateway for the public into finding other endeavors that a person is doing with GAN – even if it’s not the most technically challenging application of the algorithms. “We’re at a point now where it’s fairly straightforward to plug existing, sufficiently large datasets into a GAN and just see what happens without a lot of fiddling or programming required,” Guzdial said.

Matthew Guzdial's GAN Pokémon
Matthew Guzdial’s GAN Pokémon

To help me gain at least a cursory understanding of GAN, Syafiq Kamarul Azman gave me a useful metaphor that centered around cops and criminals. Azman, who currently works as a research engineer at his postgrad university and generates batches of Pokémon in his free time, explains the mathematics as a struggle between criminals who are making fake banknotes in a city where a bunch of real money (the original dataset of the real Pokémon) just appeared out of nowhere. “But more importantly, both the cops and criminals are totally oblivious,” Azman says. “So like baby cops and baby criminals.”

At the end of each day, the baby cops get a stern talking to by their police chief about the mistakes they made, meanwhile an informant is assisting the criminals in making their fake bills better and better. These day cycles repeat themselves – with both sides getting better at their objectives – until the cops can’t improve their abilities to pick out the fakes from the real ones and the informant has no more information about how to improve the counterfeits. If you’re into seeing the really interesting, albeit inscrutable, training images that a GAN produces, Mushbuh has uploaded a whole heap of pictures onto their Flickr.

“The system in the city is analogous to the GAN itself and, curiously, computer scientists are very interested in what the criminals are creating,” Azman said. For him though, these quirky, whispered renditions of Pokémon aren’t what’s paying the bills, so they don’t go much farther than getting posted on his Github or Twitter. But each sprite serves as a marriage between his adult, professional passion for computer science and his childhood love of the series that was prematurely cut short when he wasn’t able to afford a Game Boy Advance. “I used to be under the cover some nights with a torch, volume at zero, nodding off from time to time, catching em’ all,” Azlan said. “I vividly remember my friend showing me the Missingno glitch in a magazine article and actually trying it out, failing a couple of times and finally getting it; that was insane.”

DCGAN learning to generate “fake” Pokemon #PyTorch #DeepLearning pic.twitter.com/ZfLPr7vGya

— Syafiq Kamarul Azman (@syaffers) October 10, 2018

As a relative layperson, these generated Pokémon are just another example of the kind of creations that provoke hard questions about the murky boundaries we place on art and authorship where so much of the process is necessarily reliant on AI and propelled by the constant good-faith sharing of community members in GAN and artificial intelligence circles. Mushbuh’s creative partner, Robbie Barrat, very recently saw his code used to create a piece of AI-produced art that sold for $432,500 at Christie’s Auction House (the first of its kind to be sold at a major auction house) without adequate crediting or compensation. In the specific case of the generated Pokémon, it’s worthwhile to note that the original Pokémon art, used as the input for the GAN to do its work, is never credited and is extremely hard to find in the first place.

As for the question of whether these Pokémon or any other GAN output are “art,” Azman believes it’s highly subjective and relies on the personal history that an individual brings to an interaction with these images. “Is the GAN making art by imagining new Pokémon or is it having blurry flashes of its brainwashed network weights?” Azman said.  “Is it really expressing something? What does that say about us and our perception? The question is profound.” Mushbuh told me that, “for the stuff I want to make and for most AI artists, in general, this isn’t a replacement it’s a tool- GAN’s can be a good jumping off point and help you think differently!”

But regardless of how we classify the images, they can continue their life after generation in the ways they inspire creativity in the minds of other humans. After Mushbuh’s original post, several commenters began posting their interpretations of the noise – sparking animation and life.

these weird generated #pokemon that @mushbuh made are so cool. they hugely inspired me and i decided to make some models and renders of them#screenshotsaturday pic.twitter.com/Xhw5VDohaC

— HELLO (@hello_aberdeen) October 20, 2018

These Pokémon are almost the pawns in a bigger campaign, whose end goal is to bring attention to all of the awesome things that can be done with GAN. And for me, it worked. Upon Matthew Guzdial’s recommendation, I am now keeping up with the work of Janelle Shane and other folks who are engaged with really compelling projects that use GAN to complement their extraordinary creative visions.

GAN rendering of runway fashion used by Mushbuh and Robbie Barrat for their latest project.
GAN rendering of runway fashion used by Mushbuh and Robbie Barrat for their latest project.

I’m excited to see where artist applications of GAN take us next – whether video game related or not. As gamers, we have a unique perspective on what it means to experience art mediated through technology. These Pokémon offer us a chance to discover the creatives and their burgeoning fields that seek expand our horizons of what these mediations can look like.

Categories: Features

D.W. Wallach

D.W. writes about video games and how to cherish our moments with technology. D.W. is non-binary and uses they/them pronouns. Twitter: @gaiaonline420