A machine-learning engineer has harnessed a powerful machine-learning technology to develop a web application that lets you create images of fake pocket monsters. The results are hilarious, and I love how simple the tool is.
Liam Eloie, who works in machine learning at e-commerce marketer Programmai, published the “Nokémon” generator back in January. Eloie was inspired by BuzzFeed data scientist Max Woolf, who last December forced an AI bot to create weird-looking Pokémon. Eloie did something similar, but instead of just sharing the results after the fact, he fine-tuned the process so that you too can create fake monsters.
According to a January blog post, Eloie detailed the technical wizardry that went into the Nokémon generator. It sounds complicated, but essentially, the AI program DALL-E, which creates images from text descriptions, was trained on a dataset of all Pokémon and their attribute types, and used its accrued knowledge to spit out a legion of strange new ones. Now, in the web app Generate Fakémon using AI, you can specify a base Pokémon or element and it’ll show you a fitting example from its ever-expanding menagerie of strange-looking pseudomons.
Unlike some other, similar image generators, there’s not much you can tweak: Just choose a Pokémon or type and hit generate. What you get is a Nokémon based on your selection. For example, I asked for a fire-type monster and ended up with a creature that looked like a riff on the Gen II fire starter Cyndaquil.
Not every Nokémon created looks that good. Some are very bizarre, like this “dark and normal Pokémon” based on the Gen I grass and poison-type monster Gloom. It looks like a crude amalgam of Cubone and Sandshrew. Not sure I like it all that much.
Eloie, who’s a big fan of Arcanine, told Kotaku over email that the program is still being worked on. He said that he tested many ideas and models, including one that can learn how to evolve these Nokémon. That addition wasn’t successful–yet. Eloie also said the new feedback buttons should alleviate some of the repeats and help him produce even better-quality fakes. Each one takes the AI model about 30 seconds to generate, and Eloie adds the “best” ones to the web interface.
It’s curious seeing Pokémon fans using AI technology to create new monsters. Arceus is Game Freak’s latest in the long-running series, which holds some 242 Pokémon altogether. If the developers ever start running out of creature ideas, maybe they can tap an AI to generate them instead.