Reproduce “Luminark: Training-free, Probabilistically-Certified Watermarking for General Vision 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
Jiayi Xu, Zhang Zhang, Yuanrui Zhang, Ruitao Chen, Yixian Xu, Tianyu He, Di He
Luminark, a training-free and probabilistically-certified watermarking method, provides formal statistical guarantees for detection while maintaining high perceptual quality across various generative models. It consistently achieved over 95% detection accuracy against diverse image transformations and significantly better FID scores (e.g., FID of 2.13 vs. 4.87 for certified baselines on EDM2-XXL) compared to previous certified approaches.
Key insights
- Defining watermarks through patch-level luminance statistics provides inherent imperceptibility and robustness as it aligns with human perception and is stable against minor image perturbations.
- Incorporating watermark constraints via a guidance mechanism allows for a model-agnostic, training-free injection process that adaptively integrates the watermark during image generation.
- The statistical properties of the patch-level luminance patterns enable probabilistic certification, offering strong theoretical guarantees on controlling false positive rates during detection.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.