Abstract: Active deception jamming for synthetic aperture radar (SAR) relies on high-fidelity SAR jamming templates. However, measured data collection and electromagnetic simulation suffer from high costs and heavy computational overhead. Generating pseudo-SAR templates using optical images can expand adversarial samples at a low cost. To address the problems of target geometric distortion, loss of radar scattering details, and gray histogram distortion in optical-to-SAR (OPT2SAR) cross-modal translation, this paper proposes the nested attention and multi-scale discriminative generative adversarial network (NAMD-GAN). The network adopts a VCB-UNet++ generator incorporating the convolutional block attention module (CBAM) to preserve the strong-scattering contours of ships. It designs a three-layer multi-scale PatchGAN to hierarchically constrain global contours and local scattering textures, and introduces a one-dimensional discrete Wasserstein loss to align the gray statistical distribution of images. Experimental results show that the Fréchet Inception Distance (FID) and Kernel Inception Distance (KID) of the proposed method are 57.27 and 0.01027, respectively. The target detection rate reaches 79.80% in the constant false alarm rate (CFAR) detection pipeline, and it achieves stable adaptability for multiple image recognition networks. This study can provide effective technical support for the construction of radar adversarial sample libraries.