Windows play a vital role in occupants’ well-being, yet quantitative methods for analysing how people interact with window views remain limited. This paper introduces a novel computer-vision framework for assessing both movement and gaze dynamics in window views. First, we analyse real-time eye-tracking data from 84 participants viewing 35 high-resolution window-view recordings, revealing a clear hierarchy of attention drivers: semantic and social objects dominate viewing patterns (57.3% importance), followed by perspective features (32.9%), and movement (9.8%). Furthermore, areas of true fixations exhibit elevated perceptual features, including energy (+17.1%), contrast (+15.4%), and colourfulness (+14.4%) relative to random viewing patterns. Second, we introduce STAMP-GNN, a deep learning model designed to capture the continuous evolution of viewer attention in dynamic window views. Our approach outperforms baseline methods across multiple metrics, achieving improvements in correlation coefficient (+0.020), normalized gaze path saliency (+0.037), and similarity (+0.012). By enabling temporal, data-driven assessment of view dynamics, this framework opens new opportunities for evidence-based architectural design, offering a path to more occupant-centric and health-supportive building solutions.