The Spire of AI - Voxel-based 3D Neural Style Transfer

Abstract

In the architecture field, humans have mastered various skills for creating unique spatial experiences with unknown interplays between known contents and styles. Meanwhile, machine learning, as a popular tool for mapping different input factors and generating unpredictable outputs, links the similarity of the machine intelligence with the typical form-finding process. Style Transfer, therefore, is widely used in 2D visuals for mixing styles while inspiring the architecture field with new form-finding possibilities. Researchers have applied the algorithm in generating 2D renderings of buildings, limiting the results in 2D pixels rather than real full volume forms. Therefore, this paper aims to develop a voxel-based form generation methodology to extend the 3D architectural application of Style Transfer. Briefly, through cutting the original 3D model into multiple plans and apply them to the 2D style image, the stylized 2D results generated by Style Transfer are then abstracted and filtered as groups of pixel points in space. By adjusting the feature parameters with user customization and replacing pixel points with basic voxelization units, designers can easily recreate the original 3D geometries into different design styles, which proposes an intelligent way of finding new and inspiring 3D forms.