yue2-music

Create a request JSON with style, lyrics, cot mode, and seed, then run the generation helper or the YuE2 Python pipeline. Use cot="full" for editable melody and harmony, cot="melody" for a chord-free plan, or cot="off" for direct text-to-audio generation.

Transcribe the recording with SheetSage2 to obtain ABC notation, inspect and correct the score, strip chord symbols, then render with YuE2 using cot="melody" plus your target style and lyrics. This preserves the melody symbolically but not the original singer's identity.

The supported baseline is an NVIDIA GPU with 24GB VRAM and BF16 support, running one request at a time with default settings. SheetSage2 requires a separate Python environment because its dependency pins differ from YuE2.

No. MERT-v2 encoders return continuous features, not the discrete semantic tokens YuE2 uses internally. YuE2 also exposes no audio-reference, phoneme-alignment, or local-inpainting arguments in its public interface.

Render a baseline from the full plan, edit a copy of its score.abc under a defined invariant contract, validate the changes with the ABC comparison helper, then regenerate audio with the revised score, style, and lyrics. Old latents cannot implement musical edits.

The YuE2 and related model weights are released under CC BY-NC 4.0, meaning non-commercial use only. The skill's own code is Apache 2.0, but that license does not relicense the model weights or remove their non-commercial terms.