These AI-generated Pokémon designs are weirdly compelling

Some of Pokemon’s designs are uncomfortably close to reality (at least from a safe distance)

I vividly remember kids at school carrying posters of the original 150 Pokémon (some have even been laminated!), And as the series nears 900 months, it’s a good time to see which type of Pokémon designs can bubble up. of a well-trained AI.

As this experiment by Max Woolf, data scientist at BuzzFeed shows, it is possible to create funny, weird, and eerily precise pocket neural network monsters.

I forced a bot to look at each Pokemon and told it to generate its own. Here are the results.

(this is no joke, this is actually how I made them) pic.twitter.com/MfJUWJHZoB

– Max Woolf (@minimaxir) December 15, 2021

to a dedicated Pokemon fan, a lot of critters will immediately register as off-mark, but I bet I could be tricked with a few of them in a quick quiz.

After garnering a lot of well-deserved interest in art on Twitter and Reddit, Woolf posted of them Following lots of AI-generated Pokémon, and they’re worth a close inspection:

Wow, you all really, really love these AI-powered Pokémon!

As a thank you for all your support, how about ANOTHER BONUS LOT? ! ?? pic.twitter.com/kM3Kc8bBe6

– Max Woolf (@minimaxir) December 15, 2021

Writing more about the project on Reddit, Woolf said that “the AI ​​used here is a ruDALL-E refined on official Pokemon images (i.e. not VQGAN + CLIP or Wombo Dream). The way the AI ​​works is that it generates the top-right images in 8 × 8 chunks. It samples the next chunk somewhat randomly so that the image is consistent, the fine-tuning process learning. to the AI ​​to better recognize the pieces of a Pokémon.

While it would be amazing to have an “interactive demo” (much like the easy-to-use Pokemon Fusion tool), as Woolf puts it, “it’s not very portable / easy to use”.

The subject of generative accusatory networks was brought up in a subsequent conversation on Reddit, and he replied that “there have been attempts to form a GAN on Pokemon but it is very, very difficult to get a consistent result. . (GANs require a large amount of high-quality, normalized input images, which Pokemon doesn’t.) ”Perhaps that will inspire other experiences!

The machines that learn about Pokémon are way over my head, but fascinating nonetheless. The image at the top of this post shows some of my favorite little monsters, and yes # 2 makes our heads spin. # 4 looks like a random NFT, and # 8 is valuable enough to be real.

Hope the fan art gets out of hand as soon as possible.