Evaluating the aesthetic quality in computer-generated renderings via a comparative analysis

Abstract

In architectural competitions, patrons often receive a multitude of design submissions. Often, it is hard for reviewers to evaluate all submissions in a fair and balanced manner. The research aims to investigate how computational models can assess aesthetic quality in architectural renderings by comparing human-judged scores with algorithmic predictions. Using a dataset of crowdsourced architectural competition entries from Arcbazar, different deep learning models are trained to predict and compare aesthetic scores and generate attribute-based heatmaps. These heatmaps visualize the regions that contribute positively or negatively to the perceived quality of the designs, offering explainable AI outputs. The method includes preprocessing the images, extracting perceptual features, and evaluating model performance through metrics. The results show a high correlation between predicted and actual scores, validating the model’s effectiveness. By using machine learning algorithms, a fair and efficient method to assess aesthetics across a large number of submissions is tried to be achieved. This study aims to contribute to the field by providing a transparent and replicable framework for aesthetic evaluation in architecture, bridging human perception and machine analysis. It also demonstrates how explainable AI tools can support assessments in design competitions and stimulate critical dialogue on aesthetics in computational design processes.