MultiTalk Silent Embeds

MultiTalk is the audio-driven talking-video approach a lot of the Wan community rates highest - it's what powers continuous talking-head content (InfiniteTalk builds on the same idea, letting people make 20-second clips with no visible transitions). Normally you feed MultiTalk audio and it drives the mouths. This node is the opposite: it produces silent MultiTalk embeds - placeholder conditioning for frames where there's no audio and nothing should be talking.

Why would you want that? Because real sequences have silence. A pause between sentences, a stretch of the video where the character is listening rather than speaking, a lead-in before the voice starts. Silent embeds give you correctly-shaped MultiTalk conditioning for those frames so the graph runs cleanly and the face stays still instead of flapping at phantom audio.

How it works

MultiTalk expects a stream of audio-derived embeddings, one span of conditioning per chunk of frames. This node fabricates the "no speech here" version of that conditioning for a given frame count - a valid MULTITALK_EMBEDS object that carries no talking signal. Downstream, the model reads it as "these frames have no speech to sync to," so the mouth stays closed and neutral.

The input and output that matter

There's exactly one input:

• num_frames (INT, default 81) - how many frames of silence to generate embeds for. Set it to the length of the silent stretch you're covering. The Wan default clip is 81 frames, hence the default here.

The output is multitalk_embeds (MULTITALK_EMBEDS) - wire it into the MultiTalk conditioning path of your Wan graph, in place of (or combined with) the audio-driven embeds.

How to install it

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

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

then restart. This node itself downloads nothing - but a working MultiTalk graph needs the MultiTalk model and its audio encoder, so it's only useful as part of that larger setup.

Common issues & troubleshooting

Frame-count mismatch. The num_frames here needs to line up with the segment it's covering. If your silent stretch and the embeds don't agree on length, you'll get a shape error or misaligned timing where the silence lands in the wrong place. Count the frames of the silent section and match them.

Mouth still moves during silence. If you expected stillness but the face keeps talking, check that the silent embeds are actually reaching the model for those frames and haven't been overwritten by audio-driven embeds later in the chain. This is conditioning plumbing - order and routing decide what actually gets used.

You don't have a talking pipeline yet. This node is meaningless on its own. It's a companion to a full MultiTalk workflow (audio encoder + MultiTalk model driving Wan). If you're just starting with audio-driven video, get the talking path working with real audio first, then use silent embeds to handle the quiet gaps.