Pitch MMAudio VoCoder Loader

Of the four MMAudio nodes in this pack, PitchMMAudioVoCoderLoader is the one that looks redundant until you understand the 44k/16k split - and then it makes perfect sense. It loads a BigVGAN vocoder checkpoint from ComfyUI/models/mmaudio and hands it out as a VOCODER_MODEL. That's the entire job: one input picker, one output, no settings.

The reason it exists is the asymmetry baked into the pack. In 44k mode, the FeatureUtilsLoader doesn't need you to load a vocoder at all - it auto-downloads NVIDIA's bigvgan_v2_44khz_128band_512x on first run and builds the vocoder internally. In 16k mode, there's no auto-download path: the FeatureUtilsLoader requires a VOCODER_MODEL wired into its bigvgan_vocoder_model input, and this node is the only way to produce one. So you reach for PitchMMAudioVoCoderLoader exactly when you've decided to run MMAudio at 16k.

How it works

Under the hood it loads a BigVGAN vocoder from the checkpoint you pick and returns the model path in the VOCODER_MODEL wrapper type. Inputs and outputs:

• vocoder_model (enum) - the BigVGAN checkpoint in ComfyUI/models/mmaudio.
• mmaudio_vocoder (VOCODER_MODEL) out - wire this into the bigvgan_vocoder_model input on PitchMMAudioFeatureUtilsLoader.

The complete chain for a 16k setup looks like: PitchMMAudioModelLoader (16k checkpoint) → PitchMMAudioFeatureUtilsLoader (mode 16k, with this node's output plugged into bigvgan_vocoder_model) → PitchMMAudioSampler. If you run 44k, you can leave this node out of the graph entirely.

Installing it

It's part of pitch7900/ComfyUI-PitchCustomNodes, installed the same way as its siblings - ComfyUI Manager, search "Pitch Mask Fallback," or:

cd ComfyUI/custom_nodes
git clone https://github.com/pitch7900/ComfyUI-PitchCustomNodes

Restart, and drop the BigVGAN vocoder checkpoint into ComfyUI/models/mmaudio. The pack's README is empty, so that folder location comes from the node source, which registers the mmaudio model folder itself. Dependencies are shared with the rest of the pack and install automatically: accelerate, open_clip_torch, huggingface_hub, omegaconf, einops.

Common issues

• "bigvgan_vocoder_model must be provided for 16k mode." You know this one by now - it's the 16k requirement. Wire this node's output into the FeatureUtilsLoader, or switch to 44k and skip the vocoder loader entirely.
• Nothing in the dropdown. The checkpoint isn't in ComfyUI/models/mmaudio, or ComfyUI hasn't refreshed the folder list. Drop the file in and refresh the node.
• "torch.nn has no attribute Buffer" style errors. The underlying BigVGAN code needs a current torch; a pinned-old torch breaks the vocoder first. Update ComfyUI's torch before troubleshooting anything else.

Real talk: for most people the answer is just use 44k and never install this node. It exists for the 16k path, it's clean at that one job, and that's all there is to it.