Generative artificial intelligence and building design: early photorealistic render visualization of façades using local identity-trained models

Abstract

Abstract This paper elucidates an approach that utilizes generative artificial intelligence (AI) to develop alternative architectural design options based on local identity. The advancement of AI technologies has increasingly piqued the interest of the architecture, engineering, construction, and facility management industry. Notably, the topic of “visualization” has gained prominence as a means for enhancing communication related to a project, especially in the early phases of design. This study aims to enhance the ease of obtaining design images during initial phases of design by drawing from multiple texts and images. It develops an additional training model to generate various design alternatives that resonate with the identity of the locale through the application of generative AI to the façade design of buildings. The identity of a locality in cities and regions is the capacity for the cities and regions to be identified and recognized as a specific area. Among the various visual elements of urban and regional landscapes, the front face of buildings may play a significant role in people’s aesthetic perception and overall impression of the local environment. The research proposes an approach that transcends the conventional employment of three-dimensional modeling and rendering tools by readily deriving design alternatives that consider this local identity in commercial building remodeling. This approach allows for financial and temporal efficiency in the design communication phase of the initial architectural design process. The implementation and utilization of the proposed approach’s supplementary training model in this study proceeds as follows: (i) image data are collected from the target area using open-source street-view resources and preprocessed for conversion to a trainable format; (ii) textual data are prepared for pairing with preprocessed image data; (iii) additional training and outcome testing are performed using varied text prompts and images; and (iv) the ability to generate building façade images that reflect the identity of the collected locale by using the additional trained model is determined, as evidenced by the findings of the proposed application method study. This enables the generation of design alternatives that integrate regional styles and diverse design requirements for buildings. The training model implemented in this study can be leveraged through weight adjustments and prompt engineering to generate a greater number of design reference images, among other diverse approaches.