Although recent film studies encourage a computational perspective, poster design is still primarily evaluated using qualitative visual grammar. To address this limitation, we propose an interpretable hybrid model that quantifies character prominence in movie posters. First, the model uses YOLOv8 and U2-Net to detect people and derive three mid-level cues with explicit design meaning: visual share, positional score, and saliency score. Along with actor billing, these cues are entered into a transparent, white-box regression model in which every weight can be interpreted. A user study with twenty-one participants calibrated the visual component, achieving a correlation of R-squared equal to 0.498 with human judgment. We then applied the model to a curated set of 2,527 films released between 2003 and 2023. Our analysis revealed that character prominence is strongly influenced by genre, that most designs adhere to a standardized template of moderate prominence, and that poster design has shifted toward larger character sizes over the past two decades. Additionally, we found that the final prominence score only weakly correlates with market performance. By converting qualitative visual grammar into measurable variables, our study provides a practical tool for large-scale visual media mining and offers new insights into how movie marketing strategically captures attention.