YuE2 Song Generator One-Click ComfyUI Windows Installer
Give it lyrics and a style and it generates a complete song with vocals and accompaniment in 48 kHz stereo. YuE2-3B is the m-a-p team's open lyrics-to-song model — it outscores Suno v5 on the WildSongBench benchmark — and this one-click installer sets up ComfyUI Portable, the song workflow, my SongScribe style-preset node and both models in a single double-click.
No subscription and no credit meter — it runs entirely on your own machine, and every track stays there.
Two ways to create: write from scratch by picking a style from the SongScribe dropdown and adding your lyrics, or copy a melody — upload an existing song and the SheetSage2 encoder pulls its background music and melody into an editable melody/chord score to seed your new track.
What the installer does: sets up ComfyUI Windows Portable with the CUDA 12.8 PyTorch build, installs the SongScribe node, preloads the YuE2 workflow with the melody-copy option built in, then pauses to ask before downloading the models. Git, 7-Zip and the correct PyTorch build are handled automatically.
What you get
A one-click Windows installer that sets up ComfyUI Portable, the YuE2 song-generation workflow, the SongScribe style-preset node, and both YuE2 models.
Model files (downloaded automatically into the right folders):
yue2_3b_bf16.safetensors (song model) — placed in checkpoints/
sheetsage2_bf16.safetensors (audio encoder, used for melody copy) — placed in audio_encoders/SongScribe node: pick a style from a dropdown and it fills in the style description for you, so you don't write it out every time. It's from my SongScribe project, which is an active work in progress — the installer sets it up for you, and development is on GitHub at TheLocalLab/ComfyUI-SongScribe.
Workflow: preloaded, with the melody-copy path built in and switched on or off from the Fast Groups Bypasser.
System requirements
OS: Windows 10 or 11, 64-bit.
GPU: NVIDIA RTX 30XX / 40XX / 50XX. 24 GB VRAM recommended — peak usage is around 11 GB. A song a few minutes long takes roughly 1–2 minutes on an RTX 4090.
RAM: 24 GB system RAM recommended.
Disk: at least 25 GB free.
Handled for you: Git, 7-Zip and the correct PyTorch build are installed automatically.
Getting started
1. Put the install .bat in its own dedicated folder and double-click. Use a simple local path like C:\YuE2\ — avoid Desktop, Documents, and cloud-synced folders.
2. The installer pauses before downloading models — press Y to download or N to skip. The window waits for your input.
3. Load the workflow and confirm both models are selected.
4. Pick a style from the SongScribe dropdown, add your lyrics, and generate. For melody copy, load a reference song and enable the SheetSage2 path in the Fast Groups Bypasser.
5. Keep first generations short to dial in the style and lyrics before committing to a full track.
Troubleshooting
Install fails or behaves oddly: check the install path first. A cloud-synced folder (OneDrive, Dropbox) or a deep path with spaces is the most common cause. Move to C:\YuE2\ and re-run.
A model isn't showing in a loader: confirm you pressed Y at the download prompt, and check that yue2_3b_bf16 landed in checkpoints/ and sheetsage2_bf16 in audio_encoders/.
Melody copy has no effect: the SheetSage2 group has to be enabled in the Fast Groups Bypasser, and a reference song loaded.
Out of memory: close other GPU-heavy apps and test with a shorter song first.
The result doesn't match the style you wanted: try a different SongScribe preset, and iterate on short clips before a full-length track.
Licence & credit
YuE2-3B is by the m-a-p (Multimodal Art Projection) team. ComfyUI YuE2 support is by Comfy-Org. The SongScribe node is my own. This product is the installer and workflow packaging — it automates the setup of their model into a one-click process.
Licence note: the YuE2 model weights are released under CC BY-NC 4.0 — non-commercial. Personal use and experimentation with attribution are fine; commercial use is not permitted under that licence. Read the terms on the model's Hugging Face page before using anything you make commercially.