YuE2 in ComfyUI: editable scores and music generation

This tutorial demonstrates YuE2 music generation in a custom ComfyUI workflow distributed through the creator's paid membership. The speaker describes a 3 billion parameter model with 48 kHz stereo output, then uses lyrics and style prompts to generate a symbolic score and render audio. An extra Qwen text model writes the lyrics and descriptors.

The piano roll reads and outputs ABC score notation. Users can preview notes as simple sine waves, move them and reuse a generated score with different prompts. The supplied description specifies that this is not MIDI-file import or export. Later, the speaker changes the tempo to 110 BPM and the key to F minor without regenerating the score.

Pop, drum and bass, metal and house examples test prompt adherence and guidance settings. The speaker hears instability at CFG 10 and compares CFG 5 with a control for house music. Score editing also produces problems: a key change appears to transpose incorrectly, and lyrics fail to come through in a later render. The description advises checking the node pack README for fixes and current requirements.

For local AI use, the speaker reports about 22 GB of VRAM without the extra text model and roughly 30 GB with it. These are observations from this workflow, not universal requirements. At recording time, the speaker says the missing audio encoder prevents training and raw-audio encoding. The metadata lists CC BY-NC 4.0 for YuE2 weights and separate terms for the node code.