There is something you may not realize about neural network methods: the NNs used are huge, composed of billions of coefficients, and usually take several hundred mega or tens of giga in memory to work.
This is not something that can be considered for a software like G’MIC, where filters run on the user’s computer (and not online).
Even if we were to add tensorflow features to G’MIC one day, this would not solve the problem: the neural networks currently described in scientific articles are memory monsters, not at all suitable for integration into mainstream software.
The study of light neural networks with equivalent performance is an open research subject, and few things exist. I think we have to wait a little while until the size of the networks becomes more reasonable before we consider integrating these things into G’MIC.