WaterFlow: Fast and Robust Watermarking in the Latent Fourier Domain

ICLR 2026 Conference Submission13951 Authors

Keywords: image watermarking, latent diffusion models, computer vision

TL;DR: Fast and Robust learned watermarking with Stable Diffusion Models

Abstract: The rise of high-fidelity generative imagery has amplified the need for practical and robust visual watermarking techniques. However, existing methods often suffer from high computational cost and fail to withstand modern generative and adversarial attacks, limiting their real-world applicability. In this work, we present WaterFlow, a fast, lightweight, and highly robust watermarking framework that embeds hidden signals with high fidelity. WaterFlow leverages pretrained latent diffusion models to insert watermarks directly in the latent space. Unlike prior approaches, it learns a watermark in the Fourier domain of the latent representation—enhancing robustness while preserving perceptual quality. This design enables efficient and accurate watermark detection, even under challenging compound perturbations. Additionally, WaterFlow supports real-time control over the quality–robustness trade-off without retraining, making it adaptable to diverse use cases. We evaluate WaterFlow on MS-COCO, DiffusionDB, and WikiArt, where it consistently outperforms prior methods in robustness while matching the image quality of top-performing approaches.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 13951