Machine Learning 1

This was a project to train a machine learning system in terrain erosion, to speed up the process of landscape creation. It used a dataset from Washington in the US, and a custom version of ‘Pix2pix’. While the results looked promising, the resolution (256×256) wasn’t high enough to be useful, and turned into more a learning exercise than a production-ready tool.

Machine Learning 2

This was another ML project, aimed at generating building footprints for my ‘Greenhouse’ tool. I eventually ended up writing a differentiable SVG renderer in an attempt to make the training fast enough to run on my home machine. Another interesting exercise, which turned into more of an academic challenge than something robust enough to actually do anything useful (an ongoing theme with ML tools).