These AI-generated Pokemon are strangely compelling

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

I vividly remember kids at school lugging around posters of the original 150 Pokemon (some were even laminated!), and as the series nears the 900-‘mon mark, it feels like the right time to see what kind of Pokemon designs can bubble up out of a well-trained AI.

As shown in this experiment from Max Woolf, who’s a data scientist at BuzzFeed, it’s possible to create some funny, weird, and eerily accurate neural-net pocket monsters.

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

(this isn't a joke, that's actually how I made these) pic.twitter.com/MfJUWJHZoB

— Max Woolf (@minimaxir) December 15, 2021

To a dedicated Pokemon fan, many of the creatures are immediately going to register as being off-brand, but I bet I could be tricked with a few of them in a rapid-fire quiz.

After getting a lot of well-deserved interest in the art on Twitter and Reddit, Woolf posted two more batches of AI-generated Pokemon, and they’re worth inspecting up close:

Wow, you all really, really like these AI-Generated Pokémon!

As a thanks for all your support, how about ANOTHER BONUS BATCH?! 😁 pic.twitter.com/kM3Kc8bBe6

— Max Woolf (@minimaxir) December 15, 2021

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

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

The topic of generative adversarial networks came up in an ensuing conversation on Reddit, and he replied that “there have been attempts to train a GAN on Pokemon but it’s very, very hard to get coherent output. (GANs require a large amount of normalized high quality input images, which Pokemon are not.)” Maybe this will inspire other experiments!

Machines learning about Pokemon is very above my head, but fascinating, all the same. The image at the top of this article shows some of my favorite little monsters, and yes, #2 is flipping us off. #4 looks like some random NFT, and #8 is precious enough to be real.

I hope the fan art spirals out of control asap.