A preprint on arXiv, "On the Intrinsic Limited Robustness of Latent-Based Watermarking," challenges the reliability of watermarking techniques used in diffusion models. The authors contend that prior evaluations have overstated how well these watermarks withstand common image distortions, especially geometric ones.
The paper specifically calls out rotation, scaling, and translation (RST) as transformations that latent-based watermarks fail to survive as robustly as claimed. This points to a broader issue: the latent-based paradigm itself may have inherent limitations that no amount of fine-tuning can fully overcome.
Because this is a single preprint, the findings have not yet been peer-reviewed or independently replicated. Still, the claim that latent-based watermarking is intrinsically fragile under RST distortions could have implications for content provenance and copyright protection in AI-generated media.