Enhancing SAR Image Generation Quality via Laplacian Pyramid-Enhanced VAE with Frequency Domain Loss

We address a persistent limitation in diffusion-based SAR image synthesis: the loss of fine spatial detail and texture inconsistency that undermines realism and downstream utility. To mitigate these issues, we propose an enhanced Variational Autoencoder (VAE) tailored for SAR imagery and engineered to serve as a drop-in replacement for standard VAE modules in diffusion pipelines. The design embeds a multi-scale Laplacian feature pyramid into the encoder–decoder bottleneck, enabling explicit modeling and reconstruction of high-frequency spatial details across multiple resolutions. Complementing this, we introduce a Fourier spectral discrepancy loss that penalizes mismatches in the frequency domain, promoting spectral fidelity and preserving SAR-specific textural statistics. Together, these components reduce texture artifacts while retaining structural information critical for SAR interpretation. We integrate the improved VAE into several mainstream diffusion frameworks, including GeoGDiffusion and AeroGen, and perform extensive experiments on diverse SAR datasets. Quantitative evaluations with FID, SSIM, and LPIPS metrics demonstrate substantial improvements over baseline approaches, while qualitative comparisons reveal sharper edges, more coherent speckle behavior, and improved object delineation. Ablation studies confirm the complementary contributions of the Laplacian pyramid and spectral loss. Our method provides a practical and effective enhancement for diffusion-based SAR synthesis, improving visual quality and the preservation of SAR image characteristics for use in training, analysis, and downstream sensing tasks.