Muse
yuhui1038/Muse · Python · MIT
Fully open long-form song model with style controls
Muse emphasizes reproducibility: open data, weights and fine-grained style control. Smaller community than ACE-Step, but strong research transparency — a hidden gem for serious builders.
Score
Setup at a glance
git clone https://github.com/yuhui1038/Muse.git Deep dive
Muse is a long-form song generation project built for reproducibility: it generates full songs with fine-grained style control, and — unusually for this space — releases both the weights and the datasets behind them.
The paper was accepted to ACL 2026, the demo website is public, and all datasets and models were released together so results can be reproduced rather than just admired.
The repo separates training and inference documentation, making it a solid reference for anyone building their own song-generation stack.