Muse: Towards Reproducible Long-Form Song Generation with Fine-Grained Style Control

<p align="center"> 🎵 <a href="http://113.207.49.217:50000/">Demo</a> • 📄 <a href="https://arxiv.org/abs/2601.03973">Paper</a> • 📊 <a href="https://huggingface.co/datasets/bolshyC/Muse">Dataset</a> • 🤖 <a href="https://huggingface.co/bolshyC/models">Model</a> • 📚 <a href="#citation">Citation</a> </p>

This repository is the official repository for "Muse: Towards Reproducible Long-Form Song Generation with Fine-Grained Style Control". In this repository, we provide the Muse model, training and inference scripts, pretrained checkpoints, and evaluation pipelines.

News and Updates

• 2026.02.11 🎵: Demo website is now publicly available!
• 2026.01.11 🔥: We are excited to announce that all datasets and models are now fully open-sourced! 🎶 The complete training dataset (116k songs), pretrained model weights, training and evaluation code, and data pipeline are publicly available.

Installation

Requirements: Python 3.10 is required.

To set up the environment for Muse:

• For training: Install the training framework:

  pip install ms-swift -U
  

• For inference: Install vLLM:

  pip install vllm
  

• For audio encoding/decoding: Some dependencies (e.g., av) require system-level packages. On Ubuntu/Debian, install FFmpeg 4.4+ first:

  sudo apt-get update
  sudo apt-get install -y software-properties-common
  sudo add-apt-repository ppa:savoury1/ffmpeg4 -y
  sudo apt-get update
  sudo apt-get install -y pkg-config ffmpeg libavformat-dev libavcodec-dev libavdevice-dev libavutil-dev libswscale-dev libswresample-dev libavfilter-dev
  

  We recommend creating a new conda environment with Python 3.10. Note: Since omegaconf==2.0.6 is required and has compatibility issues with pip 24.1+, you need to downgrade pip first:

  pip install "pip<24.1"
  

  Then install dependencies:

  pip install --default-timeout=1000 -r requirements_mucodec.txt
  

  For more details, please refer to the MuCodec official repository.

• For data pipeline and evaluation: If you need to run data processing scripts (lyrics generation, metadata processing) or evaluation scripts, install additional dependencies:

  pip install -r requirements_data_eval.txt
  

Repository Structure

This repository contains the following main directories:

• train/: Training scripts and utilities for fine-tuning the Muse model. See train/README.md for details.
• infer/: Inference scripts for generating music with the Muse model. See infer/README.md for details.
• eval_pipeline/: Evaluation scripts for assessing model performance (Mulan-T, PER, AudioBox, SongEval, etc.).
• data_pipeline/: Scripts for building and processing training data, including lyrics generation, metadata processing, and music generation utilities.

Model Architecture

<p align="center"> <img src="assets/intro.jpg" width="800"/> </p>

Acknowledgments

We thank Qwen3 for providing the base language model, ms-swift for the training framework, and MuCodec for discrete audio tokenization.

Citation

If you find our work useful, please cite our paper:

@article{jiang2026muse,
  title={Muse: Towards Reproducible Long-Form Song Generation with Fine-Grained Style Control},
  author={Jiang, Changhao and Chen, Jiahao and Xiang, Zhenghao and Yang, Zhixiong and Wang, Hanchen and Zhuang, Jiabao and Che, Xinmeng and Sun, Jiajun and Li, Hui and Cao, Yifei and others},
  journal={arXiv preprint arXiv:2601.03973},
  year={2026}
}