Conditional diffusion model with residual decomposition generation strategy for SAR to optical image translation

Synthetic aperture radar (SAR)-to-optical image translation aims to recover visually interpretable optical images from SAR images acquired under all-weather and all-day conditions, yet remains challenging due to the substantial modality gap between SAR and optical image. Existing learning-based approaches exhibit complementary limitations: GAN-based methods tend to preserve global structure but often produce over-smoothed textures and suffer from training instability, while diffusion-based models excel at detail synthesis but struggle when directly modeling the full SAR-to-optical mapping, leading to structural distortions and inefficient optimization. To address these problems, this paper proposes a two-stage SAR-to-optical translation framework based on conditional residual diffusion refinement strategy (CRDRS). In the first stage, a conditional GAN generates a coarse pseudo-optical image that captures the overall scene structure, providing a stable and geometry-consistent prior. In the second stage, CRDRS is used to predict the residual between the pseudo-optical image and the real optical target, rather than directly synthesizing the optical image. CRDRS can simplify the learning objective by concentrating the diffusion process on high-frequency details and localized discrepancies. Furthermore, a heuristic theoretical analysis is provided to show that CRDRS can reduce target variance and lower intrinsic denoising error compared with direct optical generation, leading to a better conditioned optimization problem. Experiments on two representative SAR-optical datasets demonstrate that the proposed method consistently improves both pixel-level and perceptual performance, while exhibiting enhanced robustness to degraded coarse guidance. The source code is publicly available at https://github.com/Mizar29/Residual-Decomposition-Generation-Strategy-for-S2O .