The development of diffusion models has significantly enhanced the generation of high-quality images. However, it has also introduced notable challenges in the copyright protection of generated images. Existing research typically embeds watermarks invisibly during the diffusion process to increase their robustness. However, current watermarking methods based on diffusion processes focus on embedding a fixed watermark, failing to meet the diverse needs of users. In addition, malicious users may bypass watermarking by replacing the decoder, which presents a potential risk. To address these challenges, a two-stage watermarking scheme for diffusion models based on LoRA is proposed. In the first stage, the watermark codec is trained during pre-training to ensure stable watermark embedding. In the subsequent U-Net fine-tuning stage, LoRA and an adaptive attention mechanism are incorporated. This enables U-Net to learn watermark patterns from the first stage while maintaining image quality and supporting multi-user customization. Experimental results demonstrate that the proposed scheme outperforms existing methods in terms of image consistency and watermark robustness. Under image attacks, the FID distance of the watermarked images increases by only 0.61%, while the average extraction accuracy improves by 4.9%.