Hierarchical Music Generation with Explicit Structural Planning via Graph-Guided Variation Attention

​

This research proposes a hierarchical music generation framework that achieves both long-term structural coherence and musically meaningful variation. Although Transformer-based models can generate fluent local phrases, they often fail to maintain thematic consistency over an entire composition and tend either to repeat previous material verbatim or to generate unrelated content. To address this limitation, the proposed framework separates music generation into two stages: a Structure Planner and a Content Generator. The Structure Planner represents a composition as a directed graph of sections and predicts a structural blueprint specifying the reference section, pitch shift, and time-stretch ratio for each section. A GNN–VAE architecture and self-supervised metric learning are employed to learn variation relationships without requiring explicit edge annotations. The Content Generator introduces graph-guided variation-aware attention, which transforms retrieved Value representations according to the planned transposition and temporal scaling rather than merely copying them. Training combines constructive denoising on GiantMIDI-Piano with structure-aware fine-tuning on POP909. The framework will be evaluated using self-similarity matrices, interval-based edit distance, temporal alignment accuracy, ablation studies, and blind listening tests. By explicitly visualizing and controlling the structural blueprint, this research also aims to improve the interpretability and human controllability of AI-assisted composition.