Generated Adventure — postmortem of a game made with chatGPT and Midjourney (prompts included)

Generated Adventure — postmortem of a game made with chatGPT and Midjourney (prompts included)

Also, we use chatGPT to make tutorial texts more comprehensive, concise, and easy to understand. We ask the system to rewrite the tutorial text in the ELI5 (a.k.a. ‘explain to me like I am five years old’) format. A few times we generated descriptions for our games with chatGPT, mixed them with human-made versions and bought some ads to check which one would prove to be more attractive. Sometimes chatGPT’s texts performed better, sometimes they did not.

The internet is full of articles on using cases like ours but this story is not about them.

As most of our players know, we are huge fans of game jams and experiments. Even itself is a huge experiment of mixing learning and video games. With this passion for experiments in mind, we joined the 72 hours-long game jam Ludum Dare 53, which took place from April 29th to Tuesday, May 2nd, 2023 with the theme “Delivery”.

Our idea was to try generating the whole game using AI-based systems. Not the game engine, of course — we use Defold because it’s super lightweight and web-ready which is important for the Ludum Dare rating system (everyone can easily play the web game and rate it, but not everyone is ready to download executable files, and launch it).

Defold is a lightweight game engine Generated Adventure made with

Since it’s all about relying on generative systems we decided to choose the visual style and genre which would be more achievable with Midjourney and Stable Diffusion.

From our experience, we know that these systems can easily generate something beautiful, but tuning, and maintaining a consistent style of dozens of tiny elements can consume a lot of time. That’s why we started experimenting with concepts that would work in a limited number of locations, one big generated picture per location. Isometric flying “islands”, as seen in Into the Breach, looked like a perfect fit.

Source link