YuE2 in ComfyUI: setup, score editing and covers

This tutorial follows YuE2-3B through a ComfyUI workflow for generating full songs with vocals and accompaniment. The speaker describes an open source model that writes an ABC notation score before producing audio, so users can inspect and edit the musical plan. The reported output is 48 kHz stereo, with support for English, Mandarin, Japanese and Korean.

Setup starts with downloading a checkpoint and an audio encoder into the appropriate ComfyUI model folders. The speaker distinguishes the official version's reported 24 GB VRAM requirement from a GGUF version said to run with 8 to 9 GB. The int8 checkpoint's file size, under 4 GB, is a separate figure from runtime memory requirements.

The practical focus is score editing. Style and lyrics feed the ABC generation node, whose output normally passes straight to audio generation. To review it first, the presenter bypasses sampling and copies the score into a native multiline text node. The example changes F minor to C minor before reconnecting the edited text and generating another version with the same lyrics.

For reference audio, the workflow uses SheetSage2 transcription to supply ABC notation instead of generating a random score. The speaker also cites benchmark results above Suno V5 and V6 and a 71-second generation time for a 3.6-minute song on an RTX 4090. These are reported figures, rather than independent tests in the walkthrough.