Low-resource GAN-stack for high-resolution floor plan generation with enhanced evaluation and contextual validation

Abstract

The fast and reliable prototyping of floor plan layouts is a crucial element in the early stage of the building construction cycle. The purpose of this study is to ameliorate the major difficulties associated with the use of generative adversarial networks (GANs) for automating high-resolution floor plan generation: vast computational and data requirements, training instability, and problematic result evaluation. A stable resource-efficient multi-module GAN-stack framework is proposed comprising pre-processing (denoising and 4$\times$ down-sampling), floor plan image generation, and up-sampling modules. Each module is individually optimized. An innovative holistic evaluation framework of the generated building floor plans is presented covering image quality, diversity, truthfulness, overall training time, energy spent during training and associated carbon emissions. A novel validation framework is introduced, involving contextual building functionality, data privacy, capability to manipulate design generation to suit the designer’s desires, usability in BIM downstream tasks, and inference speed. Results demonstrate that the apropos network architecture choices allow for significantly cutting down the wall clock training time ($<$60 h) while maintaining superior generated image quality and contextual meaningfulness, given only sub-thousand 1024 × 1024 training images of single-story residential Danish homes and a limited computing resource (a single RTX-3090 GPU). The proposed computer-aided pipeline may support decisions between architects and their clients. It may broaden access to the GAN-based research on the automation of building design. The presented automated tools could find application in other industries which have the same driving needs and resource constraints for adopting GANs, and that also lack ways of validating their end product.