Atmospheric haze significantly degrades image quality, impacting critical applications like surveillance, autonomous driving, and aerial photography by reducing visibility, altering contrast, and changing color fidelity. Traditional dehazing techniques, which rely on paired training datasets of hazy and clear images, face limitations due to the often impractical task of collecting such varied and specific data. We introduce a novel use of CycleGAN, a Generative Adversarial Network that efficiently trains on unpaired datasets to address these limitations. This method employs cyclic consistency loss to transform hazy to clear images while preserving key attributes, eliminating the need for exact paired examples. To further enhance this model, we integrate attention mechanisms that target regions most affected by haze, ensuring effective dehazing that improves visibility without compromising image detail integrity. The objectives of this study are to demonstrate the feasibility of using unpaired datasets for dehazing with a CycleGAN, enhancing the model’s focus on haze-affected areas through attention mechanisms, and evaluating the dehazing quality against existing methods to deliver more precise, more detailed outputs for real-world applications. This method aims to push forward the technological capabilities in image dehazing and offers a scalable and flexible solution adaptable to varying environmental conditions and application requirements.