StyleGAN facilitates the creation of high-resolution images and provides an ecosystem for a variety of applications. However, the capability poses challenges in distinguishing real and generated images. To address this issue, we propose a novel generative dual-feature watermarking scheme for StyleGAN, called DFWGAN. The proposed method addresses a new issue in the domain of generative watermarking. It is the first approach to embed watermarks into both the semantic and texture features of AI-generated images during the generation process. This method enables robust watermark embedding while maintaining the generation of high-quality images. Experimental results demonstrate that the watermarked image generated by the proposed method exhibits excellent visual quality and robustness. For example, the FID for the 128 × 128 CelebA dataset is 9.23 after embedding 128 bits, resulting in a gain of 2.27 compared to the best result in the literature. Additionally, the Acc is 99.39% under JPEG compression (Q = 50), surpassing the best result reported in the literature by 27.29%.