<p>Image dehazing, a fundamental task in low-level computer vision, aims to recover haze-free images from their degraded counterparts. Recently, Denoising Diffusion Probabilistic Models (DDPMs) have demonstrated considerable potential for image dehazing owing to their powerful generative capabilities. However, existing DDPM-based methods typically suffer from inadequate preservation of fine-grained details and color fidelity, necessitating further investigation. These approaches are conventionally trained to learn comprehensive visual dehazing information from scratch, which exhibits deficiencies in contextual reasoning—particularly in capturing inter-patch relationships within images—and incurs substantial computational overhead. To address these limitations, we propose DehazeDiff, a mask-guided diffusion model that incorporates a masking mechanism to explicitly enhance the model’s capacity for learning contextual relationships among image patches. Specifically, we first partition the input image into masked and visible regions through strategic patch masking. Subsequently, an encoder block and mask-interpolator block collaborate to predict content in masked regions based on visible information while simultaneously maintaining the reverse denoising process for visible regions, thereby enabling the model to learn intrinsic spatial dependencies among image patches. Finally, a decoder block reconstructs the complete haze-free image from the generated intermediate representation. DehazeDiff effectively reduces the complexity of direct end-to-end learning from complete hazy images through progressive masked learning. Extensive experimental evaluations demonstrate that our proposed diffusion-based dehazing model achieves superior performance compared to state-of-the-art methods across three widely-adopted benchmark datasets.</p>