What is the difference between this model and other SR models?

Hi, I have few questions for better understanding. 1- What is the difference between this model and other SR models? Is it mainly about using the mouth mask? 2- How to run it in real-time? It is too slow on GPU. 3- How to possibly use the reference frame during inference? Thanks