minimaxir

A couple months ago, I found that a) visual image watermarking is trivially defeated by AI image editing so invisible image watermarks are likely the future and b) the only steganographic image watermarking tools are hard-to-use open-source tools or proprietary tools like SynthID. So as an experiment I tried using agents to create a novel image watermarking approach…and it unexpectedly worked: mostly imperceptible, tamper resistant, higher capacity, doesn't use a neural network, and real time encode/decode.

I want to open-source it as there are very many legitimate uses for image watermarking. However, as of late there has been a lot of discourse about image watermarking on both sides, namely a) invisible image watermarks can be used to faciliate dystopian user tracking (https://news.ycombinator.com/item?id=48198291 ) and b) tools to strip AI image watermarks are unethical/antisocial (https://news.ycombinator.com/item?id=48200569), so overall I'm confused on the ethics of working with watermarking tools and I am concerned there could be a negative externality I am missing by making it more causally accessible. At the least, having only closed-sourced options for image watermarking isn't a good outcome.