The Semantic Imprinting Hypothesis: How Semantic Watermarks Survive Prompt-based Editing

The emergence of text-to-image diffusion models necessitates robust intellectual property protections. The threat landscape has evolved from simple perturbations to generative editing, which introduces significant distributional shifts by extensively re-synthesizing pixel and latent statistics. Most existing benchmarks have focused on post-hoc methods. As a result, limited research has addressed the durability of semantic watermarks embedded during generation, especially under prompt-based editing. This study presents the first comprehensive evaluation of semantic watermarks across a range of editing approaches, including inversion-based methods that preserve generation trajectories and inversion-free editors, such as InfEdit, that reinitialize the generation process. The results reveal a marked contrast: traditional watermarking methods that diverge from the Gaussian prior exhibit substantial degradation, whereas manifold-aligned watermarks, such as HSQR and G.Shading, remain almost perfectly detectable even in seed-independent editing scenarios. This resilience is attributed to Semantic Imprinting, indicating that well-aligned watermarks are internalized as stable visual features rather than fragile noise patterns, enabling them to withstand distributional shifts introduced by generative re-synthesis.