AquaLoRA is a watermarking method for Stable Diffusion models that provides white-box protection by embedding watermarks into the U-Net structure. The method uses a two-stage approach: latent watermark pre-training and prior preserving fine-tuning. In the first stage, a watermark scheme is trained to be learned by the U-Net. In the second stage, the watermark is integrated into the model using a scaling matrix, allowing for flexible secret modifications. The method ensures high fidelity, robustness against image distortions, and flexibility for large-scale deployment. AquaLoRA is evaluated on various models and experimental settings, demonstrating strong performance in terms of fidelity, robustness, and flexibility. The method is effective in protecting the copyright of customized Stable Diffusion models and can be applied to various scenarios, including checkpoint-sharing. The results show that AquaLoRA outperforms existing watermarking methods in terms of accuracy and robustness. The method is also efficient and can be adapted to different model types, making it a practical solution for protecting AI-generated content.AquaLoRA is a watermarking method for Stable Diffusion models that provides white-box protection by embedding watermarks into the U-Net structure. The method uses a two-stage approach: latent watermark pre-training and prior preserving fine-tuning. In the first stage, a watermark scheme is trained to be learned by the U-Net. In the second stage, the watermark is integrated into the model using a scaling matrix, allowing for flexible secret modifications. The method ensures high fidelity, robustness against image distortions, and flexibility for large-scale deployment. AquaLoRA is evaluated on various models and experimental settings, demonstrating strong performance in terms of fidelity, robustness, and flexibility. The method is effective in protecting the copyright of customized Stable Diffusion models and can be applied to various scenarios, including checkpoint-sharing. The results show that AquaLoRA outperforms existing watermarking methods in terms of accuracy and robustness. The method is also efficient and can be adapted to different model types, making it a practical solution for protecting AI-generated content.