Ovi MMAudio VAE Loader

Ovi generates video and audio at the same time, and just like the picture side needs a VAE to turn latents into pixels, the sound side needs one to turn audio latents into an actual waveform. This node loads that audio VAE - the MMAudio VAE, plus its vocoder - for the Ovi pipeline. Its description is exactly one line and it's the truth: "Loads MMAudio VAE for Ovi audio generation." You'll place this near the top of an Ovi graph, and its output flows into the decode step that produces the audio track.

Nothing fancy here; it's a loader. But it's a required loader - without it, the audio half of Ovi has no way to become sound.

How it works

MMAudio is the audio component Ovi uses; its VAE encodes/decodes audio to and from the latent space that the joint generation works in, and the vocoder is the final stage that reconstructs a listenable waveform from the VAE's output. This node loads both and bundles them as an MMAUDIOVAE object for the rest of the Ovi chain to use.

The inputs and outputs that matter

• vae - the MMAudio VAE file to load (from ComfyUI/models, in the folder the pack scans for it).
• vocoder - the matching vocoder file. The two go together; a mismatched pair produces garbled audio.
• precision (bf16 default / fp16 / fp32) - bf16 is the sensible default; this is a small model, so precision is rarely the bottleneck.

Output is mmaudio_vae (MMAUDIOVAE), which feeds the Ovi audio decode. This is the same VAE conceptually paired with WanVideoAddOviAudioToLatents on the latent-combining side.

How to install it

ComfyUI Manager - search ComfyUI-WanVideoWrapper, install, restart. Manual:

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

then restart. Download the MMAudio VAE and vocoder files (from Kijai's WanVideo_comfy HuggingFace repo) into the models folder so both dropdowns populate. These are separate from the main Ovi generation weights - you need all of it for the pipeline to run.

Common issues & troubleshooting

Empty vae or vocoder dropdown. The files aren't where the node looks. Download both the MMAudio VAE and its vocoder, drop them in the expected models folder, and restart ComfyUI to re-scan.

Audio comes out as noise or garbled. The most likely cause is a VAE/vocoder mismatch - they're a matched pair and swapping in the wrong vocoder wrecks reconstruction. Use the pair the Ovi workflow specifies.

Where does this fit? Think of it as the audio equivalent of your video VAE loader. It doesn't generate anything itself; it's the encode/decode bridge for the sound channel. If you're not running Ovi, you don't need it - every other audio path in this pack (MultiTalk, InfiniteTalk) drives motion from audio rather than generating audio, and none of them use this.