After nine generations of Pokémon, it can sometimes feel like the game developers are just pulling random words from a hat and seeing what they come up with. I still can’t decide if Klefki , the sentient keyring Pokémon, is a stroke of genius or madness.
Now, there’s an AI generator for that. Justin Pinkney, a machine learning researcher at Lambda Labs, created a “text to Pokémon” generator by fine tuning Stable Diffusion , an AI image generator that functions in the same vein as DALL-E 2.
Fine tuning #stablediffusion to make Pokemon!
I wrote a quick guide on fine tuning your own Stable Diffusion: https://t.co/hLWrOjEPTm
I also released my Pokemon model, you can try it out on Replicate: https://t.co/3sVQrk54wZ
or with this Notebook: https://t.co/nRU5EQt3WC pic.twitter.com/d1FIAp49nO
— Justin Pinkney (@Buntworthy) September 20, 2022
Pinkney trained the AI model on descriptions of Pokémon — for example, Bulbasaur is described as “a drawing of a green pokemon with red eyes,” while Caterpie is “a green and yellow toy with a red nose.”
On Twitter, Pinkney explained that once the model is fine tuned, it will generate images in that style no matter what. In this case, that means cartoonish pocket monsters.
Once you have a fine tuned model, it can’t help but generate Pokemon not matter the prompt you give it. So no more painstaking prompting required:
“robotic cat with wings” pic.twitter.com/OoeAyoDhOB
— Justin Pinkney (@Buntworthy) September 20, 2022
To use the generator for yourself, just link your Github account and get going. Then, you can toggle the number of image outputs, as well as the number of denoising steps and the scale of classifier-free guidance.
Because we are all the main characters of our own stories, thousands of people on Twitter inputted their own names to see what they would look like as a Pokémon. In case anyone was wondering, I am a red-haired spider that only has three legs. But when you input the names of celebrities, you get some really cursed images. For example, I have created a monster:
Image Credits: Text to Pokémon (opens in a new window)
Image Credits: Text to Pokémon (opens in a new window)
What’s even creepier is when you input the names of actual Pokémon and get an alternate universe version of characters you’ve loved since childhood. Take a look at these, in clockwise order: Pikachu, Bulbasaur, Charizard, Mew, Lucario and Sudowoodo.
Image Credits: Text to Pokémon, images generated by TechCrunch (opens in a new window)
The text to Pokémon generator is fun to poke around with, but fret not — The Pokémon Company will still shock and awe us at every turn with the increasingly bonkers (and still, often delightful! ) new creatures they come up with. This morning, Pokémon fans got their next morsel of information before the long-awaited “ Pokémon Scarlet & Violet ” games come out in November: a beach-dwelling variation on Diglett called… Wiglett . It’s like the long furby of Digletts. Pokémon fans, we are in good hands.
This startup is setting a DALL-E 2-like AI free, consequences be damned
A new main series Pokémon game is coming in late 2022
