FL YuE2: Piano Roll Music Studio for YuE2 in ComfyUI

What the pack adds

ComfyUI landed native YuE2 support earlier this month, which covers lyrics-to-song generation and ABC score conditioning. FL YuE2 builds a production layer on top of it: you compose songs from lyrics and style, build or import an ABC score, and render 48 kHz stereo audio with YuE2-3B. The nodes plug into ComfyUI's standard audio preview, saving, and processing nodes, and the pack uses ComfyUI's existing Torch installation instead of installing the upstream YuE2 package.

The node set is small and linear:

• Load Models fetches the YuE2-3B model and the separate audio decoder (about 7.8 GB on first run) into ComfyUI/models/yue2/, with resume support and weight-manifest verification.
• Compose takes musical style, lyrics with [Verse] / [Chorus] sections, planning mode, and seed, and outputs an ABC score.
• Render Music generates music tokens and acoustic latents; acoustic_steps=32 matches the released midpoint solver.
• Decode Audio returns ordinary ComfyUI AUDIO, with a tile_frames option to reduce decoder memory use.

The piano roll is the hook

The standout piece is the score editor workflow: a MIDI-style piano roll backed by YuE2 ABC, with two monophonic melody lanes and a separate chord lane. You drag notes, connect the editor to Compose's score_abc input, and render exactly the melody you drew. Two monophonic melodies and a chord lane is the current scope; MIDI-file import and audio transcription are not included.

Planning modes map cleanly onto it: full generates melody and chords, melody leaves the accompaniment freer, and off skips the symbolic score entirely. An unchanged composition keeps its exact token prefix through the composition socket, so re-rendering with a fixed seed reproduces the same take. A supplied ABC score bypasses score generation and retains the full or melody instruction, and blank lyrics never override your planning choice — useful for instrumental prompts where you describe the instrumentation in the style field.

AR LoRA training built in

The pack also ships a YuE2 AR LoRA trainer: reviewed datasets, checkpoint previews, and the training loop running inside ComfyUI on the same encoder. That closes the loop from style reference to custom voice or genre adapter without leaving the app, which no other YuE2 integration currently does.

Workflow

The score-editor-to-song example wires Piano Roll → Compose → Render Music → Decode Audio → Preview/Save with a shared model loader:

Availability

Install ComfyUI-FL-YuE2 through ComfyUI Manager, or clone filliptm/ComfyUI-FL-YuE2 into ComfyUI/custom_nodes and run pip install -r requirements.txt in the ComfyUI environment. It needs a current ComfyUI with Comfy Kitchen's split-half RMS/RoPE operation and a BF16-capable NVIDIA GPU; the author validated it on an RTX PRO 6000 Blackwell under Windows, and 24 GB configurations are untested. YuE2-3B remains a research license (CC-BY-NC-4.0) from the m-a-p team.