YuE2 in ComfyUI: song and cover generation tutorial

This tutorial uses YuE2 in ComfyUI to generate songs from lyrics and style prompts, then create a cover from an existing track. The presenter describes the open source model as faster than MiniMax Music 3 with comparable audio quality. Those judgments reflect the speaker's tests rather than a controlled benchmark.

The workflow has automatic and manual ABC notation modes. Automatic mode creates notation from the supplied style and lyrics; manual mode accepts a musical score to guide the arrangement. Leaving the manual input blank still produces a song. The duration setting is a ceiling: shorter outputs remain shorter, while longer songs are cropped.

The examples reuse lyrics across 130 BPM soft rock and 1950s rock and roll. A further test pairs the melody of Jingle Bells with a 160 BPM heavy metal prompt. For the cover, SheetSage2 extracts ABC notation from the first generated song, and YuE2 uses that structure with a different style. The supplied notes clarify that reference audio does not directly condition YuE2.

Music generation requires one YuE2 checkpoint; the cover workflow also needs the SheetSage2 encoder. The listed local AI test system runs Linux with an RTX 3060 12 GB and 48 GB of system RAM. These are the demonstrated specifications, not stated minimum requirements.