Yeah their approach of using two different models to detect and then inpaint is very subliminal given that many watermarks are semi-transparent. They could have just trained a UNet with adversarial loss + LPIPS to do all the work and it would have worked much better already.