ComfyUI-FL-YuE2 is a custom node pack for ComfyUI that lets users compose songs from lyrics and style, build or import an ABC musical score, and render 48 kHz stereo audio using YuE2-3B, the music generation model released by Multimodal Art Projection (M-A-P) at github.com/multimodal-art-projection/YuE; this repository only adapts YuE2 into ComfyUI nodes and is not the model itself. It installs through ComfyUI Manager or by cloning into ComfyUI/custom_nodes and running pip install -r requirements.txt, reusing ComfyUI's existing Torch rather than installing the upstream YuE2 package or downgrading it. It requires an NVIDIA GPU supporting BF16 and a current ComfyUI build with Comfy Kitchen's split-half RMS/RoPE operation; the README says it was validated on Windows with Python 3.12, Torch 2.11, and an RTX PRO 6000 Blackwell GPU, with other devices unvalidated. The first queued Load Models node downloads roughly 7.8 GB of weights into ComfyUI/models/yue2/, with resumable, manifest-verified transfers, nothing fetched at startup, an offline mode via a download_missing toggle, and support for an external model root via extra_model_paths.yaml. The core graph chains Piano Roll, Compose, Render Music, and Decode Audio around a shared model loader: Compose takes style, lyrics with [Verse]/[Chorus] sections, a planning mode of full, melody, or off, and seed; Render Music generates tokens and acoustic latents with max_duration as an upper bound; Decode Audio outputs standard ComfyUI AUDIO for native preview, save, and processing nodes. The bundled piano roll editor offers note-snapped mouse entry, multi-note group selection, an optional per-bar harmony/chord lane, key transposition, and ABC import/export, plus a browser-only synthesizer preview that does not invoke YuE2; it edits a two-melody-plus-chords score, not MIDI files or raw audio. An optional training path covers legacy AR and joint AR/NAR LoRA recipes plus paired audio-to-audio adapters, with training dependencies installed separately. Per the README, the pack's own code is Apache-2.0, while the YuE2 weights it downloads remain CC BY-NC 4.0 under their original authors' license, separate from the code license.