This paper addresses the problem of predicting stress/strain/damage fields at the microscale in fiber-reinforced composite materials. It proposes a conditional generative adversarial network （RP-cGAN） that integrates residual connections and PixelShuffle up-sampling to solve issues with complex finite element method modeling， low computational efficiency， and insufficient accuracy of existing machine learning models in predicting interface transition zones. Based on the real material distribution characteristics and parameters， a representative volume element model is established to generate multi-field datasets for T300 materials under tensile/shear loads. The RP-cGAN enhances the extraction of interface features through residual connections， reducing the prediction error of the transition zone by 30%. Additionally， it combines PixelShuffle up-sampling to increase the peak signal-to-noise ratio by 7%， effectively suppressing the checkerboard effect of traditional transposed convolutions. Experiment results show that the model can complete a single cloud map prediction in 26 ms and maintain stable performance under multiple load conditions （SSIM > 0.983 9）. When predicting damage clouds， the RP-cGAN results based on the Mises stress criterion and predictions under stiffness degradation are highly consistent with finite element calculations， providing an efficient and precise computational tool for multiscale failure analysis of composite materials.