FADE: Frequency-aware desnowing with explicit gradient structural prompts for robust foundation model perception

Visual perception under adverse weather, particularly in snowy environments, remains challenging for human viewing and downstream visual tasks because snow veiling and opaque occlusions severely degrade image quality. Although recent foundation models have shown strong generalization across vision tasks, their performance often degrades under severe snow corruption because critical structural details are weakened, occluded, or distorted. To address this problem, we propose FADE, a frequency-aware desnowing network that restores snowy images and serves as a reliable pre-processing interface for downstream vision foundation models. Rather than relying only on spatial-domain restoration, FADE integrates explicit structural prompting with frequency-domain modulation to recover local structural details and global spectral consistency. Specifically, we design a Structural Prompt Generator (SPG) that progressively encodes gradient features into explicit structural prompts. These prompts provide structural anchors for the restoration process and reduce the representation discrepancy between snowy inputs and clean images. We further introduce a Frequency Guided Module (FGM) that uses the structural prompts to regularize amplitude and phase reconstruction, enabling more effective separation of snow degradation from background structures. Extensive experiments on multiple benchmark datasets demonstrate that FADE achieves state-of-the-art restoration performance and improves the robustness of downstream visual applications in snowy conditions.