Deformable mesh models like the one you see in the video below are cool and very useful too. Each index has it’s own semantic meaning. In this case, semantic is related to features of the face: tip of the nose, chin, line of the eyebrows, etc. And in parallel a pupil detector is running too.
But one thing is to run these DL models on dedicated hardware, another is to run them on web based platform potentially targeting any platform, any hardware, any os. What you see here is running on a mid-level laptop exploiting the CPU and the internal GPU, by means of Tensorflow JS, Wasm and webGL. And on a mid-level mobile smartphone runs just as smooth as this.