VRGDG YuE2 Settings

The pack has two ways to point YuE2 at a working install. VRGDG YuE2 Installer + Settings builds one for you - source, virtualenv, weights, the lot. This node assumes you already have one. Same job, no buttons, no installation side effects, every path visible.

Reach for it when you set YuE2 up yourself from the official instructions, when you're sharing a workflow with someone whose folder layout differs from yours, or when you're moving between two installs and want the graph to be explicit about which one it's using.

What it produces

A VRGDG_YUE2_CONFIG bundle for the generation nodes, plus a summary string that spells out exactly what you configured - something like YuE2 isolated_process | <model> | <vae> | cuda | 24 GiB | torch | none | verify_hashes=False. Wire yue2_config into Generate Song, Create Plan, or Render ABC. Read the summary when a run behaves oddly; a typo'd path is invisible until something tries to use it.

The inputs

runtime_mode is the first real decision. isolated_process is the recommended one: each run spawns a worker using the Python interpreter you point at, does the work, and exits. YuE2's pinned Torch never touches ComfyUI's. in_process_experimental imports yue2 directly into ComfyUI's interpreter - faster across repeated runs because the model stays cached in memory, but it requires a Torch/Transformers stack that satisfies both YuE2 and ComfyUI at once, which is a genuine conflict and why the mode is labeled experimental.

python_executable is required for isolated mode and must be the Python of a YuE2 environment where yue2 is importable - typically <root>\.venv\Scripts\python.exe (Windows) or <root>/.venv/bin/python. In-process mode ignores it.

model and vae take a local folder or a Hugging Face repo id, defaulting to m-a-p/YuE2-3B and m-a-p/YuE2-Vae. The VAE is the decoder that turns latents into the 48 kHz stereo audio you actually hear, which is why the pack calls YuE2-Vae the recommended listening decoder - swap it and you'll hear the difference in the render, not in the plan.

Then the VRAM policy, which is the part you'll revisit:

• memory_budget_gib defaults to 24, and YuE2 keeps a 2 GiB reserve inside that number. On a 32 GiB card the tooltip's own suggestion is to try 30 for long covers; lower it if you're sharing the GPU with anything else.
• offload_ar moves unused autoregressive modules to system RAM during synthesis. It's the difference between a long cover finishing and an OOM - at the cost of transfer time.
• backend stays on torch (CUDA graphs, the fast path). torch-eager is a troubleshooting fallback and is substantially slower for autoregressive generation; you don't want it as your daily driver.
• quantization stays on none, the validated BF16 path. fp8 is real but experimental, and "experimental" in a 3B autoregressive music model means quality drift you'll have a hard time attributing.
• device offers cuda, auto and cpu. Take the tooltip at its word: CUDA is required for practical generation, and CPU is for diagnostics. Auto just picks CUDA when it's there.

local_files_only is worth a second look because its default here is false - the opposite of the installer node, which ships true. False means a missing snapshot triggers a download at run time; true means the run fails fast on incomplete local files instead of quietly leaning on the network. unload_comfy_models (default true) asks ComfyUI to release its own models before YuE2 launches, which is usually the right call on a single-GPU box.

Optional: cache_dir for Hugging Face snapshots and metadata, and verify_hashes if you want every model weight re-hashed before each run. The installer already verified once, and re-hashing a multi-GB model on slow storage costs minutes per run, so leave it off unless you've just replaced weights.

Install

The pack itself: ComfyUI Manager, search vrgamedev, or

cd ComfyUI/custom_nodes
git clone https://github.com/vrgamegirl19/comfyui-vrgamedevgirl.git

Then restart. Important distinction: installing this pack does not install YuE2. The requirements file covers the pack's own dependencies, not the model. YuE2 needs the official source (yue2-infer) importable in a separate environment plus the weights and VAE - either built by the pack's installer node, or by following the upstream instructions yourself and typing the paths here. Either way, plan on an NVIDIA BF16 GPU with real VRAM headroom; the pack's runtime notes put the official target at 24 GB, with native Windows described as experimental.

If it fails

"This executable was not found" means a path typo or a venv that never got created. "YuE2 is not importable" in isolated mode means the interpreter you chose doesn't have the package. If everything loads but you OOM, work the ladder: offload_ar on, memory_budget_gib down, shorter lyrics, and only then consider fp8. And if you're in in_process_experimental mode and generation is fast the second time but something else in ComfyUI broke, you've just met the dependency conflict the isolated mode exists to avoid.