The researchers identify the watermarks to be removed by finding pixel patterns that persist across a large number of images, as shown on this animation: https://1.bp.blogspot.com/-cJwNoUxIBzM/WZTDpw3ru6I/AAAAAAAAB... . Their proposed solution is to randomly warp the watermarks.
Unfortunately, their solution could be quickly defeated with image-to-image generative adversarial convnets trained to... remove watermarks from image pairs. (That is, instead of training a model to change, say, image style or resolution, train it to remove artificially added watermarks.)
These seems likely to me too, but until someone demonstrates it, I'm less certain. GANs are notoriously unstable and it is still quite hard to produce useful models with them.
Unfortunately, their solution could be quickly defeated with image-to-image generative adversarial convnets trained to... remove watermarks from image pairs. (That is, instead of training a model to change, say, image style or resolution, train it to remove artificially added watermarks.)