I'm hoping to use a simpler version of this neural net, which should reduce computation time by half. There are 8 weight matrices that are dynamically blended every frame ( the bulk of the computation ). This number can supposedly be reduced to 4 (according to original research).
I'm getting ~ 50fps with 2.2GHz Intel Core i7 and Iris Pro 1536MB, MacBook Pro 2015. The matrix multiplication is using numjs, and runs on a single core (no web workers).