PromptFFTNet: Prompt-Guided Hierarchical Transformer with FFT Attention for Single frame Atmospheric Turbulence removal

Abstract Atmospheric turbulence distorts images captured over long distances, blurring details and warping structures in unpredictable ways. Restoring a sharp, faithful scene from a single degraded frame remains a complex task caused by the absence of temporal information as well as the complex, non-uniform nature of the distortions. A novel single image atmospheric turbulence mitigation architecture is proposed, incorporating FFT-based window attention with prompt-conditioned hierarchical transformers for adaptive and fast restoration of degraded images. The proposed method computes frequency correlations through FFT to effectively capture long-range dependencies, unlike self-attention methods that scale poorly in terms of computation and offer limited global context. Extensive experiments on standard image restoration benchmarks demonstrate that Prompt-guided hierarchical transformer with FFT Attention Network (PromptFFTNet) achieves higher-quality restoration with less computation than existing methods. PromptFFTNet shows better performance than leading methods when evaluated by PSNR and SSIM, demonstrating high-quality image restoration and structural fidelity. This also implies that the restored images are more authentic and computationally practical, thus suitable for real-time applications. This makes PromptFFTNet a strong benchmark for removing atmospheric turbulence in images using FFT attention-based learning and prompt-conditioned restoration.