AFAD-Net: scene-adaptive diffusion model with Frequency-spatial logic for Remote Sensing Image Super-Resolution

High-resolution (HR) remote sensing imagery is vital for precision tasks like urban planning, yet its acquisition is often limited by sensor costs and noise. Although probabilistic diffusion models have set new benchmarks in Remote Sensing Image Super-Resolution (RSISR), practical deployment is hindered by static sampling schedules and constrained local attention. Such rigid designs ignore inherent scene priors, causing unnecessary computational overhead and compromised structural fidelity. To address these limitations, we propose the Adaptive Frequency-Domain Aware Diffusion Network (AFAD-Net) to balance generative quality with inference efficiency. Specifically, a lightweight scene-complexity estimator dynamically modulates hybrid noise-scheduling weights for an adaptive trade-off between detail reconstruction and speed. Furthermore, we devise a wavelet-based Scene Frequency Domain Perception Condition Guidance Module (SFPCGM) for high-frequency texture consistency and an Overlapping Cross-Attention Block (OCAB) for long-range spatial dependencies. Extensive experiments on Potsdam, Toronto, and AID datasets demonstrate that AFAD-Net consistently outperforms state-of-the-art baselines. Notably, relative to representative diffusion-based baselines (DDPM, GDP_ x0, and TESR) under matched hardware, our model achieves 15 × to 33 × inference acceleration while using fewer parameters, effectively alleviating the inherent tension between computational efficiency and reconstruction performance.