YuE
multimodal-art-projection/YuE · Python · Apache-2.0
Open music model with symbolic planning and zero-shot covers
この解説はまだ日本語がありません。英語の原文を表示しています。
YuE focuses on structure-aware music generation: melody/chord planning, covers and agent-style editing. Check weight licenses carefully — code is Apache-2.0 but some weights are non-commercial.
スコア
導入の要点
git clone https://github.com/multimodal-art-projection/YuE.git && pip install -r requirements.txt 深読み
YuE is an open music-generation model that produces complete songs — vocals, lyrics and instrumentation — from a text caption plus lyrics. The YuE2 generation is reported competitive with Suno v5/v6 on the WildSongBench benchmark, and the team publishes the full evaluation protocol.
Generation is deliberately white-box: rather than an end-to-end black box, YuE plans the song symbolically first (the SheetSage2 line of work), so the musical plan behind a track can be read, inspected and edited. Zero-shot covers reimagine a transcribed song in a new style, and agentic editing revises sections without regenerating everything.
The project ships official ComfyUI text-to-music workflow templates, hosts a public arena for listening and voting plus a demo site, and publishes technical reports on arXiv.