ComfyUI MOSS-SoundEffect v2 Nodes

Native V3 ComfyUI nodes for OpenMOSS MOSS-SoundEffect v2.0.

Installation

Clone this repository into your ComfyUI custom_nodes directory:

cd /path/to/ComfyUI/custom_nodes
git clone https://github.com/envy-ai/ComfyUI-MOSS-SoundEffect-v2.git moss_soundeffect_v2

Install the upstream MOSS package into the same Python environment used to run ComfyUI:

cd /path/to/ComfyUI
pip install -r custom_nodes/moss_soundeffect_v2/requirements.txt

Restart ComfyUI after installing dependencies.

Nodes

• Load MOSS SoundEffect v2: loads the pipeline and downloads OpenMOSS-Team/MOSS-SoundEffect-v2.0 to models/moss_soundeffect_v2/MOSS-SoundEffect-v2.0 if it is not already present.
• MOSS SoundEffect v2 Generate: generates native ComfyUI AUDIO from a prompt. Connect it to ComfyUI's built-in Preview Audio or Save Audio nodes.

Workflow Template

The included MOSS SoundEffect v2 workflow appears in ComfyUI's Template Browser under Custom Nodes after ComfyUI is restarted.

The loader defaults disable_torch_compile to enabled. This follows the upstream inference script fallback and avoids TorchDynamo graph-break errors when SageAttention calls non-traceable CUDA helpers during torch.compile.

On CUDA 13 torch builds, the node also prepends and preloads the matching nvidia/cu13/lib runtime libraries before importing or running MOSS. This keeps NVRTC VAE decode kernels from failing when libnvrtc-builtins.so.13.0 is installed but not visible to the ComfyUI process.

The loader has an advanced weight_quantization option:

• auto: uses int8_convrot for the MOSS DiT on supported CUDA ComfyUI installs, otherwise falls back to original weights.
• off: uses the upstream MOSS loader and original weights.
• int8_convrot: forces streaming DiT quantization and raises an error if ComfyUI's int8 quantization support is unavailable.

The int8_convrot path streams transformer/diffusion_pytorch_model.safetensors one tensor at a time and quantizes DiT linear weights as they are loaded. This avoids materializing the full DiT state dict in RAM. The text encoder and DAC VAE still use the upstream loading paths.

Dependencies

The upstream package requires Python 3.12+ and has strict dependency pins. It uses the existing ComfyUI torch install because this requirements file does not install the upstream torch CUDA extra.

Model Location

The loader prefers local models in:

models/moss_soundeffect_v2/

When auto_download is enabled, the default Hugging Face model is downloaded into:

models/moss_soundeffect_v2/MOSS-SoundEffect-v2.0/