I might argue that generating and decoding an actual NTSC signal, as the OP project does, would be true in ways that a generative model based on all of that would not be.