Overview
YuE2 is a current open music-generation system that places an editable symbolic score between lyrics and style on one side and the final recording on the other. The series follows its score modes, four-stage planning-to-decoding lifecycle, AR–NAR architecture, covers, bounded edits, agent workflows, and reported evaluations. It shows where score-level control improves inspectability and selective influence, while performance details, waveform identity, human preference, and broad superiority remain unproven.
Learning outcomes
- YuE2's score makes selected melody, timing, structure, and harmony inspectable before audio is rendered.
- Full, melody, and off modes place different musical decisions under constraint; none is a universal quality ranking.
- Editing changes the score and regenerates a complete performance; it does not patch or preserve the old waveform.
- Reported results support conditional claims about cover identity and selective edits, not universal quality or human preference.
Keywords
Preview every chapter · Unlock full text, audio, and video from chapter 4
PlansFAQ
What is YuE2's main difference from a one-shot music generator?
YuE2 places an editable symbolic composition between the request and the generated recording. The score can be generated, imported, compared, or edited before semantic and acoustic realization, while performance details remain uncertain.
Which YuE2 score mode should I use?
Use full when melody and harmony need inspection or preservation, melody when the melody should anchor a style change, and off when direct generation is acceptable. The choice sets a control–adaptation trade-off rather than a universal quality ranking.
Can YuE2 edit an existing song without changing the rest of the audio?
No. YuE2 edits a symbolic composition and regenerates a complete recording rather than patching the existing waveform. Score checks can support declared invariants, but the new performance may vary outside the intended change.
What do YuE2's benchmark and editing results prove?
They support protocol-specific claims about competitive aggregate quality, stronger cover identity with explicit scores, and selective influence in a tested editing study. They do not establish universal ranking, waveform preservation, or human preference.