Introduction

Currently this is only a minimal port of the inference functionalities of the original SyncTalk, which uses NeRF based technology that creates a talking head with synchronized lip movements etc.

The custom nodes contain only wav process and inference functionalities. If you want to train your model, please see the guides in SyncTalk.

NOTE: This is a wip repository tested only on Ubuntu22.04 with python3.12.5, pytorch2.4 and cuda12.4.

Installation

Open “custom_nodes” folder and clone this repo.

At first start, some requirements will be installed, and dependent repos (pytorch3d, a forked SyncTalk for newer pytorch version) will be cloned to “custom_nodes/ComfyUI-SyncTalk/Repos/”.

Specially for pytorch3d and four extensions (freqencoder, gridencoder, raymarching, shencoder) in SyncTalk, since there are usually no built wheel, they will be build on the machine and then install, which might take many minutes.

Pytorch3d may need some dependencies to compile, please refer to pytorch3d’s install guide for more information.

Build wheels and install them in advance is also a choice. I have also put some wheels built on my environment in the release.

Workflow

For quick start, please download training data and trained checkpoints from SyncTalk in advance, and put them in correct locations under “custom_nodes/ComfyUI-SyncTalk/repos/SyncTalk/”.

The following video contains a workflow to use in ComfyUI. Or you can use workflow.json in demo.

https://github.com/user-attachments/assets/e017206d-1a68-4334-859f-9a74891d7de3

This workflow contains nodes from ComfyUI-VideoHelperSuite. You may need to install it in advance.