Interesting. I suppose if it takes minutes on a GPU, it might take ages to generate one on a mobile ML/NN chip? If it were able to be optimized to somewhat-less-than-ages, it would be a different business model w/o the need to manage credits and worry about abuse.