Model-agnostic Watermarked Image Restoration without Additional Training

ICLR 2026 Conference Submission15855 Authors

Keywords: post-processing watermarking, image restoration, watermark removal, image quality

Abstract: Post-processing image watermarking technology, which can prove the authenticity of real images, causes quality degradation and information loss in the original image. Although various methods have been proposed to restore a watermarked image to the original image, these methods are model-dependent. In this study, we propose a model-agnostic watermarked image restoration method that requires no additional training. The proposed method first extracts a message from a watermarked image and embeds the same message into the watermarked image. Then, our method computes a watermark component as the subtraction between the watermarked image and the double watermarked image. Finally, the proposed method generates a restored image by subtracting the watermark component from the watermarked image because the watermark component has a high correlation with the subtraction between the watermarked image and the original image. Experimental results show that the proposed method obtains a restored image with higher image quality for 10 of 11 existing watermarking methods. Furthermore, we have extended the existing eight methods and added a re-watermarking function that updates an embedded watermark with another watermark.

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

Submission Number: 15855