Image based Dungeon Generator

 These dungeons were made using the system's developer tool mode.

This project was done as my dissertation piece for my final year of university entitled "Optimising repetition in procedural generation through computer vision." The system is able to scan and trace a round an image to create a single 'cell' which can be used for dungeon crawlers and roguelikes. Each cell can be connected to others, which the system takes a hold of

The dissertation was made in realisation to how many procedural generated games have repeating geometry, which breaks the illusion. Due to that, It was important that the system had enough variety in its generation that it would cause very little repetitions. This included: Image layering, inner images, image rotation, mirroring and sizing.

The system included both a procedural runtime system, and a designer tool that allowed the user to create the dungeon they desired. The tool includes the ability to connect multiple cells and bend the connections to create a complex network of rooms for the users dungeon, as well as manually adjust each of the previously names customisation options of the procedural tool.

The system was made to study how computer vision may help remove repetitive apects in procedural dungeon design. The system layers multiple images of different sizes and scales them down to fit. A trace of the outline is used to get the shape and options such as smoother edges, connections and inner shapes are given to the user.

The fully runtime generator uses a data table with a plethora of settings to see if an image is big enough to hold an inner image as well as reccomended connection points and compression settings that the system uses.

An issue in developing this system was how the data was stored. UE5 stores data for an image differently depending on the compression settings. If I attempted to read the image wrong with the wrong compression settings the engine would constantly crash. Despite an option for the engine to grayscale an image, I decided I would do that process and use 'vector displacement' compression settings as it stored and rgb value that could be easily read. I needed rgb values as I have future plans for the system, which uses each channel for a different result to create inner walls looping throughout each cell.

The image is divided a specified amount and then each pixel is grayscaled. The pixels are then taken through multiple processes such as flipping or a Difference of Gaussian effect

The system takes the image and traces it edge by edge. If it reaches a dead end then it will go back pixel by pixel in order to find a new edge

Some images were used for generation made with Bing Image Generator as well as my own images.

All images had background subtraction done in Adobe Photoshop