wan-multitalk

1# WAN MultiTalk — Audio-Driven Talking Avatar
2 
3## Overview
4 
5MultiTalk (MeiGen-AI) drives a still portrait's lip-sync and head motion from an
6audio track. It runs on WAN 2.1 14B Image-to-Video via kijai's
7ComfyUI-WanVideoWrapper. Wav2Vec speech embeddings condition the WAN sampler so
8the mouth and expression follow the speech, while the lightx2v step-distill LoRA
9keeps it to a few sampling steps.
10 
11Use it for talking heads, dubbing, and single-speaker avatar clips (~10s at 480p).
12It is distinct from `wan-animate` (pose/motion-driven character animation). This
13is audio → lip-sync, not reference-video motion transfer.
14 
15Pack: `wan-multitalk` (480p, ~10s). Higher-res/longer variants exist in the source
16bundle (720p, long-context) as VRAM/duration knobs on the same graph.
17 
18## Pipeline (node graph)
19 
20```
21LoadImage (portrait) ─┐
22LoadAudio ─ AudioSeparation ─ AudioCrop ─ DownloadAndLoadWav2VecModel ─ MultiTalkWav2VecEmbeds ─┐
23                                                                                                 ▼
24WanVideoModelLoader (WAN 2.1 14B I2V GGUF) ─ MultiTalkModelLoader ─ WanVideoLoraSelect (lightx2v)
25   + LoadWanVideoT5TextEncoder (umt5) + WanVideoTextEncode + WanVideoClipVisionEncode (clip_vision_h)
26   + WanVideoVAELoader ──────────────────────────────────────────────────────────────────────────┘
27                                                     ▼
28                       WanVideoImageToVideoMultiTalk ─ WanVideoSampler ─ WanVideoDecode ─ VHS_VideoCombine
29```
30…