ML Castles
# Machine Learning Supported Procedural Castles
is a project I did for my bachelor thesis. The goal was to find out how machine learning could support procedural content genertation. I created a procedural castle “generator”, which is enhanced by two types of ML systems.
The project files and code can be found on GitHub: H_ML_Castles
# ML 1: Training a Neural Net to Understand my Sketches
A Machine learning network was used to provide a new kind of user interface to control parameters of an HDA, which procedurally generates towers.
more details can be found here: Tower Sketcher
# ML 2: Predicting Heightfield Data using cGANs
This idea isn’t anything new and was already shown in this SideFX demo. Since heightfields are essentially images (2D arrays) you can use conditional GANs to translate between one image to another.
You can for example translate height data from a base terrain to an eroded version of that terrain, like shown in the SideFX demo.
In the same way you can then use the erosion data (height, water, sediment etc.) to predict where certain types of vegetation should appear. The model spits out different colored masks, which are used to scatter points for different plant or tree models.
more details can be found here: Vegetation Erosion Prediction
# Results
