This project is an exploration in generating an infinite number of pokemon sprites. This was done by scraping images from the games and then training a GAN to generate new Pokemon. In theory, this could be done with other art assets like floors and walls, using a conditional GAN.

This simply GAN ended up working, but I believe it is a bit overfit since one can see pokemon that already exist in the final output.

A video of the training process can be found at rthe link below. Each second is 24 frames, and each frame is from 1/10 training epochs. The resulting footage is relatively smooth but can be a bit flashy: https://www.youtube.com/watch?v=ClJe2RUw26A

• Get virtualenv with: pip3 install virtualenv This should add virtualenv to the system.
• To create a virtual env, now use python3 -m venv env_name this will create a virtual env in a folder by that name
• Activate the env by using source env_name/bin/activate
• To use a requirements file type pip install -r step1/requirements.txt

• Run the pokemon scraper to get sprites
• Run PokeGAN.py!
• Create a video of the output with functions in ProjectUtils.py

I figured that for a first foray into GANs, I should keep the images relatively small, and similar to each other. The pokemon games seemed to be a good fit. If anyone has any other suggestions, I'd love to hear them!