The recent GenAI products seen on Kuaishou are all quite interesting. For example, KeLing Video (a large video generation model with powerful video generation capabilities, allowing users to easily and efficiently create artistic videos), Huison (sing karaoke together, using AI audio correction to complete the KTV experience on a mobile phone). Today, I also saw an open-source project from last week —— LivePortrait, a video-driven Avatar animation framework.
Practical scenarios:
Generating portrait animations from static images： Portrait video editing： Eye and lip redirection：
Eye redirection effect Lip redirection effect
Usage
You can use the LivePortrait project on Huggingface:
Upload original images and expression videos。 
Generate video images corresponding to expressions。 
Introduction
the framework remains highly effective. The generation speed significantly reaches 12.8ms on an RTX 4090 GPU using PyTorch.
Experimental methods
First-stage process: Basic model training
Optimize the appearance extractor (𝐹), motion extractor (𝑀), deformation module (𝑊), and decoder (𝐺). At this stage, the model is trained from scratch.
Second stage process: Concatenation and redirection module training
After training the base model in the first stage, freeze the appearance and motion extractors, deformation module, and decoder. Only the concatenation module and redirection module are optimized.