Validation of Affective Evaluation of Architectural Images Using Large Language Models

This study investigates the applicability of current large multimodal models (LMMs) for affective evaluation of architectural images. Interior and exterior images were evaluated using a GPT-4-based large multimodal model with selected affective adjectives, and compared with large-scale human questionnaire surveys employing a seven-point semantic differential method. Although the absolute evaluation values did not directly correspond to human responses, a stable linear relationship was observed after calibration. Exterior images generally exhibited higher agreement with human evaluations, while for interior images, higher agreement was observed for visually grounded adjectives such as “heavy,” whereas agreement decreased for adjectives involving experiential aspects.