Player not loading? Watch on YouTube
Vladimir Chopine walks through local AI lip-sync workflows in ComfyUI, using the Windows portable 0.3.66 release with embedded Python 3.13. He tests built-in Humo and Wan 2.2 audio-driven video templates before installing a separate Wav2Lip workflow with ReActor face restoration.
His Humo examples show why prompts matter: the default text can override the supplied character image. He reports degradation when extending clips and shows a non-human character moving without matching the speech. In his tests, Wan 2.2 handles the Muppet example better, though it still misses some mouth movements.
The setup covers FFmpeg configuration on Windows 11, Git cloning into custom_nodes, and installing requirements with ComfyUI's embedded Python. Chopine explains where to place the Wav2Lip checkpoint and how ComfyUI-Manager finds missing nodes. For ReActor installation failures, he checks the Python version, installs a matching prebuilt wheel, then reruns requirements.
He also describes a soxr/librosa import failure and a sitecustomize workaround for his specific error, warning that it may not solve other failures. A separate lip-sync wrapper requires Python 3.11 in his account; using ComfyUI 0.2.3 means giving up newer nodes and updates. A sponsored Domo AI segment demonstrates an online alternative. The final local pipeline loads video and audio, restores faces with ReActor, and combines the output.
