Defective images generated by generative adversarial networks (GANs) often exhibit insufficiently constructed defects. The discriminator’s dominance leads the generator to prioritize generating color blocks favored by the discriminator, disregarding the original information. In this paper, we propose a combined GAN that retains feature information, which comprises a defect generating GAN and a mask generating GAN. The two trained GANs are synergistically combined to generate defective images. Additionally, the structuration loss introduced in this paper guides and constrains the GAN model, aiming to preserve texture trends, grayscale distribution, and narrow defect regions. Experimental results show that our model produces high-quality images, with texture information closer to the original sample, and without the additional discriminators. This approach is evaluated against the latest detection model, demonstrating 4% improvement in effectiveness over the standard enhancement method.