Google announced the roll-out of night sight recently. It seems like it is a mixture of classical computer vision algorithms and deep learning and is basically based on the HDR+ project.
It seems like it is quite easy to come up with an DNN that generates improve night images. However, it seems like they come at the costs of speed and still are not as good as a coupled approach. I am asking myself how much better a pure ML solution could be (given a big enough NN) for even worse conditions than covered in the paper (they included some failure cases as well).