MMAudioSuite FeatureUtilsLoader

MMAudio doesn't just generate audio from nothing. It conditions on three separate things at once: your text prompt, the video frames it should match, and its own audio autoencoder/vocoder stack. In the original ComfyUI wrapper you wired those up piecemeal and hoped the versions lined up. MMAudioSuiteFeatureUtilsLoader collapses all of it into a single node with one output - which is the "suite" convenience in a nutshell, and the reason a beginner should start here instead of the original pack.

What it actually loads

Three models, all from ComfyUI/models/mmaudio, all combined into one MMAUDIO_FEATUREUTILS object that the Sampler consumes:

• The VAE (mmaudio_vae_44k_fp16.safetensors) - the autoencoder that compresses audio latents, and it wraps the vocoder inside it.
• The Synchformer (mmaudio_synchformer_fp16.safetensors) - turns video frames into the sync tokens that make the sound land on the right beats. This is the heart of the video-to-audio trick.
• A CLIP model (apple_DFN5B-CLIP-ViT-H-14-384_fp16.safetensors) - the text (and image) embedder for your prompt. Apple's DFN5B, the same CLIP family image pipelines use.

The CLIP config isn't fetched from the internet - the pack ships its own configs/DFN5B-CLIP-ViT-H-14-384.json and instantiates the model from it, which is one less thing to drift.

The mode switch and the vocoder

There are two MMAudio routes, chosen with the mode input:

• 44k (default) - the good one. Full 44.1 kHz quality. On first run the loader auto-downloads Nvidia's bigvgan_v2_44khz_128band_512x snapshot into ComfyUI/models/mmaudio/nvidia/. That first queue can look like a hang while it pulls the weights; it isn't. If the download ever fails, grab the folder manually from the nvidia/bigvgan_v2_44khz_128band_512x HF repo and drop it in place.
• 16k - the lighter, lower-sample-rate path for tight VRAM. Here you must supply the vocoder yourself via the optional bigvgan_vocoder_model input, fed from the pack's VoCoderLoader node.

So the wiring rule of thumb: 44k mode needs only this loader plus the ModelLoader and Sampler; 16k mode adds the VoCoderLoader into the chain.

Inputs and output

• vae_model, synchformer_model, clip_model - the three dropdowns, populated from models/mmaudio. Pick the matching _44k_fp16 safetensors for each.
• bigvgan_vocoder_model (optional) - a VOCODER_MODEL, only needed for 16k mode.
• mode - 44k or 16k, default 44k.
• precision - fp16 default; leave it alone unless you hit a weird error, in which case bf16 is the friendlier next try on modern cards.

One output, mmaudio_featureutils (type MMAUDIO_FEATUREUTILS), into the Sampler's feature_utils input.

Install and gotchas

Install is the shared pack story: ComfyUI Manager (search "ComfyUI-MMAudio-Suite") or git clone https://github.com/Takenoko3333/ComfyUI-MMAudio-Suite into custom_nodes, restart, and on the Windows portable build run the requirements.txt install with python_embeded\python.exe -m pip install -r ... if dependencies don't land. All four safetensors files come from Kijai's MMAudio_safetensors repo - grab the fp16 variants and put them in ComfyUI/models/mmaudio.

Where people get burned: forgetting the first-run BigVGAN download isn't a freeze; mixing the 16k vocoder into 44k mode (or vice versa) and getting a model-shape error; and the rare open_clip version mismatch where the CLIP model refuses to initialize - the code has a fallback for exactly that, but a newer open_clip_torch in your environment is the usual culprit. When in doubt, keep the precision at fp16 and the mode at 44k until you've heard a successful generation, then experiment.