MMAudioSuite VoCoderLoader

Every loader in this pack has a job, and this one's job is: be the vocoder for the 16k path. That's it. MMAudioSuiteVoCoderLoader loads a neural vocoder (BigVGAN) as a VOCODER_MODEL so the FeatureUtilsLoader can reconstruct a waveform - but only when you're running MMAudio in its lighter 16k mode. If you're on the default 44k mode, this node is dead weight and you can ignore it entirely.

The situation, concretely: MMAudio comes in two flavors. The 44k route is the good one - full 44.1 kHz output, and the FeatureUtilsLoader auto-downloads Nvidia's bigvgan_v2_44khz_128band_512x for you on first run. The 16k route exists for lower-spec machines: a smaller model, faster generation, less VRAM - and noticeably rougher audio quality, because 16 kHz just doesn't carry the detail. It needs a vocoder supplied externally, and that's where this node comes in.

How it works

Nothing exotic. It calls BigVGAN.from_pretrained on a folder in ComfyUI/models/mmaudio and hands back a VOCODER_MODEL. One input, one output.

• vocoder_model - a dropdown of what's in ComfyUI/models/mmaudio. Pick the BigVGAN directory.
• Output: mmaudio_vocoder (type VOCODER_MODEL).

The wiring rule: connect this output to the FeatureUtilsLoader's optional bigvgan_vocoder_model input, and set that loader's mode to 16k. Forgetting either half gives you a model-shape or assertion error when the Sampler runs. And if you're in 44k mode with the vocoder plugged in, unplug it - it's not used, and the auto-downloaded 44k BigVGAN already lives inside the VAE loader's flow.

When to actually use it

Honestly, most people won't. The 44k path is the default for a reason, and this suite's example workflow never touches this node. Reach for 16k only if your card is tight and you want the smallest MMAudio footprint - or if you specifically have the 16k model family on disk and want to use it. In that sense it's the rare "utility node you should know exists but usually skip," which makes it a good sanity check: if someone tells you a MMAudio workflow needs four loader nodes, they're talking about the 16k setup.

Install

Same as every node in the pack: ComfyUI Manager (search "ComfyUI-MMAudio-Suite") or git clone https://github.com/Takenoko3333/ComfyUI-MMAudio-Suite into custom_nodes, restart, and on portable Windows run python_embeded\python.exe -m pip install -r ComfyUI\custom_nodes\ComfyUI-MMAudio-Suite\requirements.txt if dependencies don't auto-install. Model files go in ComfyUI/models/mmaudio, from Kijai's MMAudio_safetensors repo - grab the 16k variant's vocoder alongside its model and VAE. If the dropdown is empty, the folder is, too.