Pitch MMAudio Feature Utils Loader

In the four-node MMAudio chain this pack ships, PitchMMAudioModelLoader is the headline act and PitchMMAudioFeatureUtilsLoader is the real heavy lifter. This is the node that assembles everything the sampler needs to condition on your video and your text: the audio autoencoder (VAE), the Synchformer model that watches the video, the CLIP text encoder that reads your prompt, and the BigVGAN vocoder that turns latents back into waveforms. Four models, one node, one MMAUDIO_FEATUREUTILS output that plugs into PitchMMAudioSampler.

It's also the node where most of the setup pain lives, because "assemble four models" means four model files - and in the 44k mode, one of them downloads itself.

How it works

The required inputs are three file pickers, all reading from ComfyUI/models/mmaudio:

• vae_model - the MMAudio audio autoencoder checkpoint.
• synchformer_model - the video-understanding model that extracts sync features from your frames.
• clip_model - the CLIP encoder (the DFN5B-CLIP-ViT-H-14-384 config ships in the pack) that encodes your text prompt.

Then the settings:

• mode (16k / 44k, default 44k) - the sampling rate the whole chain runs at. 44k is the better-quality default; 16k is the lighter option.
• precision (fp16 / fp32 / bf16, default fp16).
• bigvgan_vocoder_model (VOCODER_MODEL, optional) - required in 16k mode. More on that below.

The 44k/16k split is where the behavior diverges. In 44k mode, the node auto-downloads NVIDIA's bigvgan_v2_44khz_128band_512x vocoder into ComfyUI/models/mmaudio/nvidia/ on first run - a huggingface_hub snapshot download, no action needed, but you'll watch it pull a couple hundred MB the first time. In 16k mode, there's no auto-download: you must wire a BigVGAN vocoder into bigvgan_vocoder_model, which comes from the pack's PitchMMAudioVoCoderLoader. Leave it unplugged and the node asserts - it hard-requires that input for 16k.

Output: mmaudio_featureutils (MMAUDIO_FEATUREUTILS), feeding the sampler.

Installing it

Same pack, same install as the rest: ComfyUI Manager, search "Pitch Mask Fallback," or:

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

Restart, and gather the model files into ComfyUI/models/mmaudio - the VAE, Synchformer, and CLIP checkpoints from the official MMAudio release, plus the BigVGAN vocoder if you plan to run 16k. The pack's README is empty (a theme), so the file layout above comes from reading the source. Dependencies (accelerate, open_clip_torch, huggingface_hub, omegaconf, einops) install automatically with the pack.

Common issues

• First run in 44k mode seems "stuck." It's downloading the NVIDIA vocoder into models/mmaudio/nvidia/. Let it finish; you'll only pay it once.
• "bigvgan_vocoder_model must be provided for 16k mode." You switched to 16k but didn't wire in PitchMMAudioVoCoderLoader. Either add it, or stay on 44k where the vocoder auto-fetches.
• CLIP load hiccups. The code retries the open_clip model config with a fallback flag for version differences, so a mismatch usually self-heals - but if it doesn't, make sure you're on a reasonably current open_clip_torch.
• VRAM pressure. This node holds four models. The sampler's force_offload default (on) moves them back off the GPU after generation, which is the designed escape hatch - keep it on unless you have VRAM to spare.

This is the node that makes MMAudio's two conditioning streams - video and text - possible. Get its four files right and the sampler becomes a one-stop generate node.