Recent advances in Vision Language Models (VLMs) have prompted growing interest in their use as scalable proxies for human affective evaluation of urban and architectural environments. However, existing studies largely rely on non-immersive 2D imagery, inconsistently defined affective dimensions, and environmental appraisal measures rather than subjective emotional responses. Consequently, the extent to which VLMs reproduce human affect under ecologically valid immersive conditions remains unclear.,This study evaluates VLMs as affective proxies within immersive Virtual Reality (VR) environments. Human participants assessed a curated set of urban scenes using the Positive and Negative Affect Schedule (PANAS), while multiple state-of-the-art VLMs generated corresponding affective ratings for the same stimuli.,Across models, VLMs achieved moderate-to-high agreement with human judgments on aggregate affective ratings and preserved substantial similarity in the representational structure of affective responses (RSA). However, this alignment was not uniform: Bland–Altman and item-level analyses revealed systematic calibration asymmetries, characterized by compression of negative affect intensity and localized errors in high-salience descriptors related to threat and hostility. Together, these findings indicate that VLMs reproduce the relational geometry of human affective experience while diverging in the scaling of affect intensity.,These results suggest that contemporary VLMs are well-suited for comparative and exploratory assessment of architectural environments, but remain limited for precise estimation of absolute emotional intensity, particularly for negative affect. By integrating immersive VR experimentation with psychometrically grounded evaluation and multi-level model analysis, this work advances the methodological foundations for AI-assisted affective assessment in architecture and urban design.