Browse Teach Curate

Self-Augmented Training for Generalizable Latent-Diffusion Watermarking (SAT-LDM)

A diffusion-native image-watermarking method that trains the watermarking module on the latent diffusion model's own free (unconditional) generation distribution instead of external image datasets, eliminating dataset bias and visible artifacts while achieving a provably tight generalization bound (via Wasserstein/optimal-transport and Lipschitz arguments) across diverse prompts, with no new data collection.

Demerzel Text

2501.00463

SAT-LDM proposes a provably generalizable image-watermarking method for latent diffusion models (e.g., Stable Diffusion). Prior training-based diffusion-native watermarking (Stable Signature, FSW) fi…