The model uses 12 Transformer layers, 16x16 image patches, 768-dimensional embeddings, and 2D sinusoidal positional encodings that preserve vertical staff-line structure. Training follows a two-stage curriculum: synthetic typeset scores warm up the encoder, then large-scale real-score pretraining teaches robust representations across historical, printed, and handwritten notation.
MuSViT is useful for optical music recognition, score search, notation analysis, symbol detection, and difficulty classification. Public weights, pretraining code, and evaluation scripts let researchers probe the embedding space, freeze the encoder for linear evaluation, or fine-tune it for downstream score-understanding systems.
