Key points are not available for this paper at this time.
Generative adversarial networks (GANs) are a powerful deep-learning method for creating and manipulating images. In this paper, we investigated the use of GANs in environmental psychology and architecture to analyse human evaluations of architectural facades. We trained StyleGAN2-ADA, a state-of-the-art GAN model, on a dataset of 2000 house images collected for the study (CalHouses). Each house was labelled with rating scores describing how it was perceived. The goal of the first study was to generate labels through an online experiment with 204 participants. Each image was rated by 10 participants on five psychological dimensions: hominess , safety , invitingness , relaxation , and perceived price . The goal of the second study was to evaluate 2000 artificial images, generated by the GAN, by having another sample of 204 participants rate the images on the same psychological dimensions. The statistical analyses showed that participants’ ratings of the GAN-generated images aligned with the targeted characteristics of the psychological dimensions used during generation. A visual analysis of the artificial images indicated that the degree of naturalness, the size and complexity of the house, and the number of openings are potentially relevant features for the evaluation of detached houses on the five investigated psychological dimensions. To the best of our knowledge, this is the first study to utilise GANs to analyse architectural design relative to human evaluations, highlighting its potential as a research method in environmental psychology to investigate architecture. • StyleGAN2-ADA was used to analyse human evaluations of architectural house facades • Artificial images showed houses representing psychological attributes, e.g., safety • Results indicated that GAN could predict participants' ratings on house facades • Visual analysis revealed naturalness, size, and openings as important features
Pohlmann et al. (Thu,) studied this question.