Composer Vector: Style-Steering Symbolic Music Generation in a Latent Space

Scan to Fill Survey

Help Us Improve!

We're conducting a listening test to evaluate style-steered music generation. Your feedback is valuable!

Abstract

We introduce Composer Vector, a training‑free, inference‑time steering method for controllable symbolic music generation. By injecting a style vector into hidden states with a tunable coefficient, we enable single‑style control, multi‑style blending, and continuous interpolation. Please see the paper for details.

Method

Audio / Symbolic Examples

Listen to examples of original compositions and their style-steered versions using our Composer Vector method.

Experiment 1

Bach — Original

Bach — Steered to Mozart (0.5)

Beethoven — Original

Beethoven — Steered to Mozart (0.5)

Chopin — Original

Chopin — Steered to Mozart (0.5)

Experiment 2

Bach — Original

Bach — Mixed with Mozart

Beethoven — Original

Beethoven — Mixed with Mozart (0.8)

Chopin — Original

Chopin — Mixed with Mozart (0.5)

Experiment 3

Bach — Original

Bach — Mixed with Mozart (0.1/0.9)

Beethoven — Original

Beethoven — Mixed with Mozart (0.3/0.7)

Chopin — Original

Chopin — Mixed with Mozart (0.3/0.7)

BibTeX

@misc{jiang2025composervector,
  title={Composer Vector: Style-Steering Symbolic Music Generation in a Latent Space},
  author={Jiang, Xunyi and Others},
  howpublished={NeurIPS 2025 Workshop on AI4Music},
  year={2025}
}