Understand what travels from the original
The native cover workflow transfers a transcribed melody, rather than feeding the original waveform straight into the generator. Mistakes in the transcription can carry into the cover, so review the score before judging the final arrangement.
Use source recordings and lyrics you have the rights to work with. A model-weight license does not grant permission to use somebody else’s song or voice.
Start with the Music Cover template
Open YuE2: Music Cover in ComfyUI’s Template Library. It needs the YuE2 checkpoint and sheetsage2_bf16.safetensors in models/audio_encoders/. Load the reference, enter the new style and lyrics, then queue the workflow.
Keep new lyric lines close to the original phrasing and syllable count. A melody written for a short line will not automatically make room for a long paragraph.
Keep Python transcription separate
The official cover guide uses a separate SheetSage2 environment because its dependencies differ from YuE2’s. SheetSage2 uses Python 3.10 or 3.11 and requires FFmpeg 6.1 with its shared libraries; the YuE2 environment uses Python 3.12.
Exchange the ABC and audio files between environments. Run transcription and generation one after the other to share a GPU. Follow the complete installation and transcription commands for your platform.
Edit a copy of the composition
Generate and save the original artifacts, copy score.abc, and edit the copy. Use cot="melody" to keep a melody while allowing freer accompaniment, or cot="full" when you supply melody and harmony.
For a chord-only experiment, preserve the notes, durations, bar boundaries and tempo. The repository’s abc_tools.py compare helper can check supported invariants. Listen to both versions in full: a matching symbolic score does not guarantee unchanged singing or timbre outside the edit.
Sources & further reading
Documentation checked against the linked sources. We have not benchmarked generation on our own hardware.