A lightweight dual-domain constrained generative adversarial network (LDDC-GAN) is proposed to address the limited availability of butyl rubber samples for industrial visual inspection, as well as issues of mode collapse and attribute distortion in existing generative models under small-sample conditions. This model introduces a frequency-domain energy alignment mechanism to suppress mode collapse and incorporates cosine loss to constrain texture features. By jointly optimizing the frequency and spatial domains, the model ensures structural continuity and realistic details in the generated images. In addition, a progressive channel compression strategy is designed to reduce the model size to 4.42 M parameters and optimize memory usage to 6.4 GB. Combined with an improved StyleGAN2-based generator, discriminator architecture, and a dynamic gradient clipping strategy, the model efficiently converged on a dataset of 386 industrial samples of resolution 256×256. The experimental results demonstrate that the proposed model achieved better performance than the current mainstream baseline models, as measured by the FID and SSIM metrics.