<p>Image distribution shifts and natural corruptions severely degrade the robustness of computer vision models. Existing methods either discard noise-sensitive high-frequency components or leverage attention mechanisms to dynamically select reliable high-frequency cues. However, they overlook the coupled learning of frequency sub-bands under diverse spatial variations, lacking effective constraints on high-frequency expression via multi-level robustness prompt. To address this limitation, this work presents a novel Frequency Sub-bands Coupling Learning via Spatial-Branching Framework (FSCL-SBF), which optimizes discriminative yet perturbation-vulnerable high-frequency features. Specifically, each branch’s input features are first mapped to the spectral domain using learnable frequency filters, which are decomposed into attribute-distinct sub-bands. Spatial sampling perturbs high-frequency distributions to build diverse spatial branching structures. Then following global–local robust perception, cross-branch coupled learning of adjacent frequency sub-bands stabilizes high-frequency representations and improves model generalization. Extensive evaluations on three standard clean datasets and their corresponding corrupted counterparts reveal that the proposed FSCL-SBF achieves superior performance against state-of-the-art approaches, with prominent gains on clean images and corrupted samples with noise, blur, weather, and digital degradations.</p>