WanVideo Decode Ovi Audio

Wan has no native audio at any version - the KB is blunt about this, and it's one of the few things about Wan that hasn't changed across every closed-weight release since 2.2. If you want sound, the standard answer is bolting on a separate model like MMAudio as its own pipeline stage. Ovi takes a different approach: it's a joint audio-video generation model, and this pack ships kijai's own integration of it, with dedicated Encode/Decode nodes for the audio half. This node is the second half of that pair - it turns the audio latents Ovi generated back into a playable waveform.

What it's for

Ovi generates video and audio together rather than audio being an afterthought bolted on after the fact - that's the whole pitch of a joint model. Internally, the audio side runs through the same kind of encode/latent/decode loop as the video side, and it borrows its audio VAE machinery from MMAudio rather than inventing a new one from scratch (an early community fork of Ovi's ComfyUI support noted this directly - the only extra folder it needs beyond Wan's own VAE and text encoder is MMAudio's). This node is the decode step: it takes the audio latents your sampler produced and runs them through that VAE to get actual audio out.

The inputs and output that matter

Two required inputs, both already-prepared objects: mmaudio_vae (MMAUDIOVAE) - the audio VAE, loaded elsewhere in your graph, and samples (LATENT) - the audio latents coming out of your Ovi-capable sampler. There's nothing to tune here; like most decode nodes in this ecosystem, it's a straight, deterministic conversion.

Output is a single audio (AUDIO) - a standard ComfyUI audio object, ready for a Save Audio node, a video-muxing node, or anywhere else your graph expects audio.

How to install it

Via ComfyUI Manager, search WanVideoWrapper and install, or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/kijai/ComfyUI-WanVideoWrapper

then restart. You'll need Ovi's model weights on top of your base Wan setup, plus the MMAudio VAE component specifically - these aren't part of the standard Wan T2V/I2V download, since Ovi is its own model family layered on Wan's VAE and text-encoder architecture.

Common issues & troubleshooting

No mmaudio_vae loaded means this node can't run at all. It's a required input with no fallback - make sure you have an MMAudio VAE loader node upstream, separate from whatever loads your video VAE.

Audio comes out but sounds wrong or empty. That points back to the sampler, not this decode node - decode is a fixed transform, so if the latents it received were noise or malformed, the audio will reflect that faithfully. Check that your Ovi sampling actually ran the audio branch, not just the video branch.

Running the unofficial fork's setup notes against this node's requirements. Because Ovi had an earlier, independent community wrapper before kijai's official integration landed in this pack, you may run into setup instructions online written for that other version - attention-backend flags, folder layouts - that don't map cleanly onto WanVideoWrapper's own nodes. If something a guide tells you to do doesn't match what you see in this pack's node list, you're likely reading instructions for the other implementation.

VRAM adds up fast on a joint model. You're effectively running two generation pipelines - video and audio - off the same Wan backbone at once. If you're already tight on VRAM with video-only Wan generation, budget extra headroom before adding Ovi's audio branch on top.