We love Pokémon's, don't we? What if we could make more of them? We present to you the ultimate Pokémon generator!! After training on 2000 images of Pokémon, this deep learning model is able to generate chef's kiss Pokémon's from scratch.
What it does
It produces images which look exactly like how a Pokémon would look, but the Pokémon's do not actually exist. The model has successfully learn a latent space of different Pokémon like species by training on a corpus of "real" Pokémon images. We then combine these processed images into the card layout to make a new card based on the Pokémon. We also trained a minGPT on existing Pokémon names and asked it to generate new names! Magale, Pimate, Garenige and Popet are some of the Pokémon names that don't exist.
How we built it
We built it with the help of variation of StyleGAN model which we fine-tuned to generate images. We trained a minGPT model(a minimal implemetation of GPT by Karpathy). We then manually prepared the cards(of course we used python, duh) and put them on a website.
Challenges we ran into
Collecting data was crucial for the models performance. We had to collect images from various sources, augment the data and then train the model on a huge corpus of 2000 Pokémon images.
Accomplishments that we're proud of
That we were able to fine-tune the model and were able to produce images which exactly looked like Pokémon's but were not the ones which already exist.
What we learned
This project gave us insights to how a GAN works. We had to dive deep into the theory to make sure the performance of the model isn't compromised.
What's next for These Pokémon Cards do not exist
Due to computation constraints we had to pre-process the images. We would love to host the website on a GPU server and show the Pokémon's generated real-time.