MMAudio VoCoderLoader

Here's the honest summary: if you're running the default 44k MMAudio workflow, you will never touch this node. It exists for the 16k path, which almost nobody uses anymore. But when the 16k path bites you, this is the node that fixes it - so it's worth five minutes to know what it does.

What a vocoder does here

A vocoder converts audio latents (or mel spectrograms) back into an actual waveform - it's the "speaker" at the end of the audio pipeline. MMAudio's two modes use different vocoders. The 44k mode pairs with NVIDIA's BigVGAN v2, which the FeatureUtilsLoader downloads automatically. The 16k mode - the older small-model path - expects a different BigVGAN (the one from the Make-An-Audio-2 family, a best_netG.pt style checkpoint), and it refuses to run without one:

AssertionError: bigvgan_vocoder_model must be provided for 16k mode

That assertion is the tell. When you see it, you've switched the FeatureUtilsLoader's mode to 16k and the pack is demanding you provide the vocoder it can't fetch for you.

What the node actually is

One required input - vocoder_model, a dropdown of anything in ComfyUI/models/mmaudio - and one output, mmaudio_vocoder (type VOCODER_MODEL). Internally it loads the file with BigVGAN's from_pretrained and hands you a ready-to-use vocoder object. Its one real job in a workflow is feeding the FeatureUtilsLoader's optional bigvgan_vocoder_model input, which is the thing that satisfies that 16k assertion.

When you'd actually use it

Roughly two scenarios. You're on a very small GPU and want the lightweight 16k model to fit - that's the legitimate reason to run 16k at all, and it needs this node. Or you inherited a workflow from before the 44k path existed, when 16k was the default. Both are edge cases in 2026; the large 44k model with the auto-downloaded NVIDIA vocoder is the well-trodden path, and this node stays dark on the canvas.

One warning if you do go 16k: the vocoder file has to be the right one. People have grabbed the 44k NVIDIA vocoder and tried to feed it to 16k mode (or dropped a random file in the folder and watched BigVGAN.from_pretrained choke on it). The 16k and 44k vocoders are different architectures and not interchangeable - grab the 16k BigVGAN specifically, put it in ComfyUI/models/mmaudio, and make sure the mode matches on the FeatureUtilsLoader.

Installing

No special treatment - it ships with the pack. Clone the repo into ComfyUI/custom_nodes, install the requirements, drop the model files into ComfyUI/models/mmaudio:

cd ComfyUI/custom_nodes
git clone https://github.com/kijai/ComfyUI-MMAudio
pip install -r ComfyUI-MMAudio/requirements.txt

ComfyUI Manager has it as "ComfyUI-MMAudio" if you'd rather click. In short: keep it in mind as the answer to the 16k assertion, ignore it the rest of the time, and let the 44k mode's auto-download do the heavy lifting.