No, you're misunderstanding my proposal (w.r.t. Generative Models). This is different from deniable encryption (where an eavesdropper will detect you may be encrypting something), this is traffic literally indistinguishable from normal traffic.

It only fails if the government knows a priori you are a criminal and inspects your hardware.

I thought about making a few proof of concept models for this, but I don't think there's real demand for it (and I'm lazy) -- steganography is probably sufficient for what most people use.