Topic: Removing noise and watermarks with AI

Watermarks don't have to be visible, FWIW. AnotherDay mentioned 'adding noise/grain for no discernible reason'. That is roughly the effect of Fourier watermarking -- the watermark is distributed throughout the image pixels, and applying an Inverse Fourier Transform shows the watermark. But the grain added is very slight, so you can't determine that there is a watermark (rather than just noise) by simply looking at it. And because it's distributed through the image, it's robust against mere cropping.

This post by the author of GMIC illustrates how it works (this post was written before the dedicated 'Fourier Watermark' filter was added to GMIC).

AFAICS this creates a watermark that is both a) 'huge' / difficult to remove and b) quite non-intrusive.

Updated by anonymous