
YuE
YuE2: frontier music generation with symbolic planning, zero-shot covers, and agentic music editing.
- Category
- AI
- Website
- github.com
- Language
- English
- Listed
- 5d ago
- Stars
- 9,394
- Forks
- 1,012
multimodal-art-projection/YuE ↗, Python, Apache-2.0, Last commit 1d ago
YuE is an open music generation model whose current release, YuE2, starts from a score rather than from raw audio. Given lyrics and a style prompt, it first writes a plan of melody and chords, then renders that plan as a full song with vocals and accompaniment. Because the plan is explicit, the composition can be read, played and edited before any audio is produced, which the team describes as white-box generation. The same checkpoint handles zero-shot covers, where a transcribed recording is reinterpreted in a new style, and agentic editing, where a song is revised through a conversation about its score, arrangement or lyrics.
On quality, the authors report YuE2 as competitive with Suno v5 and v6 on WildSongBench and cite a best-of-8 SongBench average of 6.9632 as the highest mean among the settings they evaluated. Their editing demo takes one song, The Last Train, through nine steps and fourteen versions. The code and weights are released under Apache 2.0, and the original YuE remains available on a separate YuE-v1 branch. The repository also links companion models, MERT2 and SheetSage2, hosted on Hugging Face.