cs.SDApr 3, 2026

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

Authors: Xunyi Jiang, Mingyang Yao, Jingyue Huang, Julian McAuley

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Symbolic music generation has made significant progress, yet achieving fine-grained and flexible control over composer style remains challenging. Existing training-based methods for composer style conditioning depend on large labeled datasets. Besides, these methods typically support only single-composer generation at a time, limiting their applicability to more creative or blended scenarios. In this work, we propose Composer Vector, an inference-time steering method that operates directly in the model's latent space to control composer style without retraining. Through experiments on multiple symbolic music generation models, we show that Composer Vector effectively guides generations toward target composer styles, enabling smooth and interpretable control through a continuous steering coefficient. It also enables seamless fusion of multiple styles within a unified latent space framework. Overall, our work demonstrates that simple latent space steering provides a practical and general mechanism for controllable symbolic music generation, enabling more flexible and interactive creative workflows. Code and Demo are available here: https://github.com/JiangXunyi/Composer-Vector and https://jiangxunyi.github.io/composervector.github.io/

Figures & tables

Figure 1: Inference-time control pipeline and style control diagrams
Figure 2: NotaGen: CLAP and CLaMP similarity before and after steering.
MethodBach MozartHaydn BeethovenHaydn LisztBeethoven MozartBeethoven BachSchubert RavelSchumann BachChopin BachDebussy Beethoven
NotaGen5.926.318.74.911.76.611.214.933.7
+vector14.034.226.216.017.914.429.222.552.2
ChatMusician3.44.55.06.025.91.229.523.01.4
+vector16.430.115.113.254.314.657.053.420.8
Table 1: Prediction probabilities (%) of target composer , where the target composer is the one specified by the Composer Vector. Columns are ordered by prompt era (Classical to Romantic). Headers show the prompt composer rr (italic) and the target composer cc (bold). Each value indicates the classifier’s predicted probability that the output matches the target composer.
Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 3: t-SNE visualization of piece-level embeddings across composers.
ComposerMIDI Perf.Audio Perf.MIDI/XML Scores
Bach16915259
Balakirev1031
Beethoven27112057
Brahms101
Chopin28910834
Debussy332
Appendix
Table 2: ASAP dataset summary: 1,067 performances and 236 distinct scores spanning 15 Western classical piano composers.
Figure 4: Comparison of layer-wise composer style localization across different models.
Figure 5: ChatMusician: CLAP and CLaMP similarity before and after steering.
Figure 6: Steering improvement heatmaps for (a) NotaGen and (b) ChatMusician.
Prompt (B / S)BachBeethovenChopinDebussyHaydnLisztMozartRach.RavelSchubertSchumann
Bach48.4% /4.3% /3.1% /7.6% /1.6% /9.3% /5.9% /2.5% /2.6% /11.8% /3.0% /
 59.3%13.3%12.0%10.7%2.3%16.5%14.0%18.2%10.0%17.9%3.1%
Beethoven11.7% /35.8% /6.4% /4.5% /1.2% /12.3% /4.9% /7.2% /4.7% /4.3% /7.0% /
 17.9%35.3%9.5%10.3%3.1%21.2%16.0%8.3%10.2%7.6%7.7%
Chopin14.9% /28.6% /12.8% /4.9% /0.6% /9.9% /4.0% /10.2% /3.3% /1.1% /9.6% /
 22.5%34.3%23.6%9.3%0.9%13.0%6.9%11.4%10.3%3.1%12.2%
Appendix
Table 3: All Classifier Prediction Results for NotaGen. Baseline (top, ends with “/”) and Steering (bottom). Numbers with green background indicate steering probabilities greater than 10%. “Rach.” = Rachmaninoff.
Prompt (B / S)BachBeethovenChopinDebussyHaydnLisztMozartRach.RavelSchubertSchumann
Bach15.1% /31.9% /0.8% /9.1% /0.1% /12.6% /3.4% /4.4% /9.7% /6.1% /6.7% /
 51.7%25.0%1.6%18.8%0.4%18.4%16.4%6.5%13.6%3.1%4.3%
Beethoven25.9% /2.5% /0.9% /22.6% /0.4% /7.4% /6.0% /7.4% /19.5% /2.7% /4.6% /
 54.3%20.7%3.2%16.7%0.4%7.9%13.2%4.2%16.7%5.4%6.4%
Chopin23.0% /6.2% /3.8% /9.9% /0.4% /10.4% /15.3% /5.1% /10.3% /7.2% /8.5% /
 53.4%25.4%2.5%19.9%0.4%12.7%13.0%4.3%13.7%2.3%5.5%
Appendix
Table 4: All Classifier Prediction Results for ChaMusician. Baseline (top, ends with “/”) and Steering (bottom). Numbers with green background indicate steering probabilities greater than 10%. “Rach.” = Rachmaninoff.
Figure 7: Correlation Between Steering Coefficient and Prediction Probability
Figure 8: Probability of Steering Composers for Style Fusion
More in this topic

Timbre-Conditioned Music Generation

See all 10 papers

Explore similar work

CardsList

  1. Diff-Symbo: Text-Controlled Long-Duration Symbolic Music Generation Using Autoregressive Latent Diffusion Model

    Aug 5, 2026Zhiwei Lin, Jun Chen, Boshi Tang +5Text-to-Music GenerationLatent Diffusion Models

    Shenzhen International Graduate School, Tsinghua University, Shenzhen, China · Huawei Technologies Co., Ltd., Shenzhen, China · The Chinese University of Hong Kong, Hong Kong SAR, China

  2. Getting Motif-ated: Controllable AI Compositions from Injected Motif Prompts

    Sep 26, 2026Chao Peter Yang, Cynthia Rudin, Yue Jiang +2Controllable Music GenerationSymbolic Music Generation

    Yale University New Haven, CT · Duke University Durham, NC · Chinese University of Hong Kong, Shenzhen Shenzhen, Guangdong

  3. Generative Modeling of Bach-Style Symbolic Music: A Comparative Study of Autoregressive, Latent-Variable, and Adversarial Approaches

    Jun 11, 2026Dezhi Yu, Kyuil Lee, Yongkang HuangVariational AutoencodersLatent Variable Models