Multi-granularity Music-guided Movie Trailer Generation via Lightweight Foreground-background Decoupling

Music-driven movie trailer generation aims to produce coherent and engaging trailers guided by music. However, existing methods typically rely on either global or fine-grained music features and treat video scenes holistically, overlooking the distinct roles that foreground and background play in music-visual alignment. To address this, we propose a novel framework that enables deep interaction between multi-granularity music features and hierarchical foreground-background visual representations. We introduce a lightweight disentanglement module that separates visual features into semantically distinct foreground and background components with minimal computational cost and performs adaptive fusion. Furthermore, we design a pseudo-supervision strategy to provide weak semantic guidance for foreground-background separation. Experimental results on the CMTD dataset and our curated MovieNet-M dataset demonstrate that the proposed MDT model outperforms existing mainstream methods in both generation quality and music alignment.