Researchers have developed a novel watermarking technique for generative models that embeds signals directly into the learned continuous dynamics, specifically the velocity field of flow matching models. This method formulates watermarking as random coding over a continuous channel, adding a key-dependent perturbation during training that does not alter the generated distribution. Experiments on standard datasets like MNIST and CIFAR-10 demonstrate reliable message recovery and preserved generation quality, with decoding accuracy dropping to chance levels without the secret key. AI
Summary written by gemini-2.5-flash-lite from 1 source. How we write summaries →
IMPACT Introduces a new method for watermarking generative models, potentially enhancing content provenance and security.
RANK_REASON The cluster contains an academic paper detailing a new technical method for generative models. [lever_c_demoted from research: ic=1 ai=1.0]