DualBranch-LoRA: An Image watermarking Method for Balancing high Fidelity and Robustness

Existing image watermarking methods face a significant trade-off between visual fidelity and watermark robustness, particularly in regions rich in detail, where visible artifacts are oftenly introduced, which severely degrading image quality. To address this issue, this paper proposes a high-fidelity image watermarking method based on a dynamically decoupled latent space structure and a low-rank adaptation mechanism, enabling efficient modeling and fine-grained control of the watermark embedding process by introducing a decoupled dualbranch architecture, Specifically, the content branch is kept frozen to preserve the image’s original reconstruction capability to the greatest extent, while the watermark branch incorporates specialized modules for efficient parameter embedding and learning of watermark information. Moreover, we propose a highfrequency perceptual loss function that combines Laplacian gradient constraints with YUV color space differences to guide the model in enhancing watermark robustness while preserving structural details and color consistency. Experimental results on the public CLIC dataset demonstrate that the proposed method achieves significant improvements over mainstream baseline models in key evaluation metrics.