Gotta Catch 'em AIl

Co-creative Pokémon Generation

By Christina Irakleous, Nikki Rademaker, and Yanna Smid

Overview

For the course Computational Creativity, we created a system that lets users design new Pokémon with the help of generative AI tools. The project combines creativity with generative AI tools, allowing users to pick types, body shapes, and an inspiration theme. Based on these choices, the system generates both a textual description and a visual representation of a new Pokémon. The goal was to build a simple, creative tool that makes the Pokémon design process more accessible and interactive.

To evaluate the output of our tool, we evaluated generated images using creativity measures (SPECS) and a small test based on the Turing test.

The Project

The Pokémon generator uses AI to generate new Pokémon designs based on user input. The tool works as follows:

• Design input: users can select Pokémon types and a body shape. They can also provide a short textual description as inspiration for the Pokémon design.
• Pokémon description: GPT-3.5-Turbo generates a text description of the new Pokémon.
• Pokémon visuals: DALL·E 3 creates a matching image based on the generated description.

The images below show the interface and what users are expected to do to use the tool.

The Pokémon generator successfully creates convincing and creative Pokémon. Most users could not tell the AI designs apart from real ones, suggesting it passed a Turing-test-like evaluation. According to the SPECS evaluation, the system was found creative, although some Pokémon looked slightly less realistic. You can find a short report on the evaluation of the project here.

Below are a few example outputs from the Pokémon generator.

Tools and Development

We built the interface in Python using Tkinter and connected it to OpenAI's GPT-3.5-Turbo and DALL-E 3 APIs. After refining the prompts to generate more fitting Pokémon designs, we evaluated the tool using creativity measures (SPECS) and a small Turing test.