Multi/InfiniteTalk Model Loader

This node loads MultiTalk (or its successor InfiniteTalk) - the model that lets Wan lip-sync a character to an audio track. Feed it a face and a voice clip and you get a talking-head video where the mouth actually matches the words. Both come from the MeiGen-AI team; InfiniteTalk is the newer one, built specifically to push talking-head generation to long, continuous clips instead of short snippets.

The loader itself is dead simple - pick the model, get a MULTITALKMODEL out. The interesting part is where that output goes: into the multitalk_model socket on WanVideoModelLoader, which fuses the talking-head capability onto your base Wan model. This node is one piece of a small cluster; you'll also need the wav2vec audio encoder and the long-video MultiTalk sampling node to actually produce anything.

How it works

MultiTalk/InfiniteTalk conditions Wan's generation on audio features so the facial motion - mouth shapes, jaw, some head movement - tracks the speech. It doesn't work alone: the audio has to be turned into features first (that's what DownloadAndLoadWav2VecModel is for), and the generation runs through a windowed long-video method (WanVideoImageToVideoMultiTalk) because talking clips are usually longer than Wan's native 81 frames. This node just gets the talking-head model into memory and attached to Wan.

The inputs and outputs that matter

• model - the MultiTalk/InfiniteTalk model file. That's the only input. Pick the weights you downloaded.

Output is a single MULTITALKMODEL, which plugs into WanVideoModelLoader's multitalk_model input.

How to install it

Part of the pack. Via ComfyUI Manager: search WanVideo Wrapper, install, restart. Manually:

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

then restart. The MultiTalk/InfiniteTalk weights are a separate download (Kijai packages them alongside the rest of the Wan collection). You also need the wav2vec model and, of course, a base Wan model to attach to.

Common issues & troubleshooting

Empty dropdown. The MultiTalk/InfiniteTalk weights aren't downloaded into the model folder yet. Grab them and restart. Note the history here: the InfiniteTalk repo actually went up, got pulled within about half an hour, and reappeared once the team sorted the release - so if a link you found is dead, get the weights from the current MeiGen-AI / Kijai packaging rather than an old mirror.

It loads but there's no lip-sync. This node alone doesn't produce talking video. You need the whole chain: the wav2vec model to encode audio, this model attached to the base Wan model, and the WanVideoImageToVideoMultiTalk node driving the windowed generation. Missing any link and you get silent, non-synced output.

Generation is slow. Expected - talking-head clips are long, and long means many windows. Early testers reported 30+ minutes for a default-length clip even with a speed LoRA, because the windowing multiplies your step count across the clip. Keep resolution and length sensible, and use the low-noise speed LoRA trick.

Which one, MultiTalk or InfiniteTalk? InfiniteTalk is the newer, longer-form successor from the same team. If you're starting fresh and want continuous talking video, prefer InfiniteTalk; MultiTalk is the earlier model that the pipeline was originally built around.