Abstract <p>Real-world video dehazing in driving scenarios presents significant challenges due to the inherent difficulty of acquiring accurately aligned hazy/clear video pairs for training, especially under dynamic conditions with unpredictable weather. In this paper, we propose the Energy-Guided Adaptive Alignment (EGAA) framework, which introduces several innovative components to enhance both video dehazing quality and scene alignment. We present a multi-scale energy-guided alignment mechanism that effectively aligns brightness and texture information across adjacent frames, improving image consistency. The framework also integrates the Multi-Scale and Mean Self-Attention (MMSA) block, which utilizes multi-head and mean self-attention mechanisms to better capture scene features and estimate the infinite airlight in hazy environments. Extensive experiments on real-world hazy datasets demonstrate that our approach significantly outperforms existing state-of-the-art methods, achieving superior dehazing performance and video quality.</p> Graphical abstract <p></p>