Reproduce “Distortion-Free Semantic Watermarks for Black-Box Generative Models” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Aviv Yossef, Or Patashnik, Or Zamir

The authors introduce a black-box image watermark that selects one of several completed generator outputs using a secret key and releases it unchanged. The key-averaged method preserves the generator’s single-output distribution under an ideal pseudorandom-function assumption, though this does not guarantee preservation for a fixed key. On Stable Diffusion 2.1, ten-image batches achieve 1.000 TPR at 20% FPR, for clean detection.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.