This chapter presents the project of architectural machine translation. The rationale of the project was that many tasks within the architectural design process could be reformulated as translation tasks between different modes of representations, such as sketches to plans, and that machine learning, specifically techniques for image-to-image translation, could serve as a tool in such tasks. The output was, among other things, a set of conditional generative models for architectural plans, sections, and elevation drawings used in an interactive installation with a desk and a webcam connected to a computer. The image-to-image training sets for the models were quite small and of low resolution. The source images were manually curated and compiled into different data sets for plans, sections, and elevations.

This paper describes a general method for synthesizing discrete 2D and 3D output by building probabilistic models of rasterized or voxelized training data, and subsequently synthesizing new data iteratively by substituting cells or groups of cells in accordance with a learned transition matrix. The process is non-deterministic, stochastic and unsupervised. The size of the source data and output is arbitrary, and the source and output data can have an arbitrary set of cell states. Possible variations of the process are discussed, as well as possible applications in design processes on multiple scales.