VRGDG SheetSage2 Settings
The unglamorous node your cover depends on
- sheetsage2_config
- summary
SheetSage2 is a transcription model - it listens to a recording and writes down the melody as symbolic notation. It is not an image model, it does not live in your ComfyUI environment, and it is pinned to a different Torch and Transformers than YuE2 is. So before anything can transcribe, something has to tell the transcription node where the model is, which Python can import it, and how much of the network it's allowed to touch.
That's this node. No generation, no GPU work, no outputs you can listen to - it builds a config bundle and a human-readable summary line. It's the plumbing for VRGDG SheetSage2 Transcribe Cover, and if covers are failing for you, this node is usually why.
How it works
It assembles a VRGDG_SHEETSAGE2_CONFIG - a plain dictionary carried on a wire - and returns it alongside a summary string like SheetSage2 | <model> + <parent> | cuda | isolated process, which is your one-glance sanity check that the paths you typed are the paths you meant.
Downstream, the transcribe node reads that dictionary and launches a worker process using the Python executable you specified: not ComfyUI's Python, a separate environment's. That's the isolation the whole pack is built around, and it's the right call - audio models in ComfyUI are a chronic source of transformers/torch collisions, and SheetSage2 is one of the ones that would definitely break your image stack if you installed it globally.
The inputs that matter
python_executable is the one you'll actually fight with. It must point at the SheetSage2 venv's interpreter - ...\SheetSage2-venv\Scripts\python.exe on Windows, .../SheetSage2-venv/bin/python on Linux. The default in the node is the author's own Windows path (E:\Yue2\...), so change it unless you happen to have installed to exactly that drive. If you used VRGDG YuE2 Installer + Settings, you can skip this node entirely - the installer outputs the same typed config already filled in.
model is the SheetSage2 snapshot (local folder or Hugging Face repo id), and parent_model is the MERT-v2-FullSong encoder SheetSage2 sits on top of. Both are needed; forgetting the MERT path is a classic first-run failure.
cache_dir is where Hugging Face keeps downloaded modules and metadata. Worth setting explicitly if you installed outside the default tree, because SheetSage2 loads with trust_remote_code - it executes Python from the model repo. Which brings us to:
local_files_only is true by default here, and that's the interesting choice. It forbids network access at run time and requires complete local snapshots, which means remote code can't be fetched behind your back mid-render and your runs don't silently depend on Hugging Face being up. Turn it off only when you're intentionally letting the model download for the first time.
device is cuda (recommended) or cpu. CUDA is genuinely much faster, and CPU is a legitimate escape hatch when a video model has already claimed your VRAM - the worker also refuses to run if you configure CUDA in an environment where CUDA isn't available, rather than quietly falling back.
unload_comfy_models defaults true: before transcription starts, the pack asks ComfyUI to drop its loaded models so the transcriber has VRAM to itself. Leave it on unless you're running a graph where reloading those models costs more than you save.
Outputs
sheetsage2_config - wire it into the transcribe node. summary - the string above; wire it to a text display if you want a permanent record of what the run used, or just read it in the node.
Install
The pack installs the normal way (Manager, search vrgamedev, or git clone https://github.com/vrgamegirl19/comfyui-vrgamedevgirl.git into ComfyUI/custom_nodes), but this node is useless without the cover environment: SheetSage2's packages and its two model snapshots are not in the pack's requirements file. Use the installer node's Install Cover Tools + Models, which creates the separate Python 3.10/3.11 virtualenv and downloads m-a-p/SheetSage2 plus m-a-p/MERT-v2-FullSong. Cover-side setup on Windows additionally needs a system Python 3.10 or 3.11 with the py launcher registered - ComfyUI's bundled interpreter won't do.
Where this bites
- Wrong interpreter = import error. If the worker dies saying SheetSage2's dependencies aren't installed, you're pointing at the wrong venv - often YuE2's, which is a different environment on purpose.
- Defaults from someone else's machine. Every path here is pre-filled with the author's
E:\Yue2layout. Blank-looking failures are path failures. - Non-commercial weights. SheetSage2 and MERT are CC BY-NC 4.0. Fine for the thing you're doing tonight; read the terms before the thing you're selling.
Inputs (7)
| Name | Type | Default | Description |
|---|---|---|---|
| python_executable | STRING | E:\Yue2\SheetSage2-venv\Scripts\python.exe | Python executable from the separate SheetSage2 environment used to transcribe source audio. |
| model | STRING | E:\Yue2\models\SheetSage2 | Local SheetSage2 model snapshot or Hugging Face repository ID. |
| parent_model | STRING | E:\Yue2\models\MERT-v2-FullSong | MERT-v2-FullSong encoder used internally by SheetSage2. |
| cache_dir | STRING | E:\Yue2\hf-cache | Hugging Face cache used for model files and trusted remote model code. |
| device | COMBO | cuda | CUDA is recommended for transcription. CPU is supported but much slower. |
| local_files_only | BOOLEAN | true | Prevent network access and require complete local SheetSage2/MERT snapshots. |
| unload_comfy_models | BOOLEAN | true | Unload ComfyUI models before transcription to free GPU memory. |
Outputs (2)
| Name | Type | Description |
|---|---|---|
| sheetsage2_config | VRGDG_SHEETSAGE2_CONFIG | — |
| summary | STRING | — |