Atmospheric scattering model-based deep learning methods for single image dehazing have faced challenges due to the difficulty of obtaining real-world ground truth priors for training. To address this issue, we explore a novel approach using the Retinex model, which, while commonly applied to low-light enhancement, has been underutilized for dehazing. Our method first estimates the residual illumination prior and compares it with real-world ground truth residual illumination during training to generate an intermediate dehazed image. To address the lack of detail in this intermediate image, we employ a coarse-to-fine strategy. A parallel unsupervised Detail Refinement branch estimates a detail map, which is then fused with the coarse intermediate image using a U-Net architecture to produce a refined final dehazed output. We introduce a novel AutoNeXt block that incorporates ConNeXt convolution, SimAM Attention, and a Feature Attention Block, which is effective for both Detail Refinement and Illumination Estimation tasks. Our experiments and analysis confirm the effectiveness of our coarse-to-fine strategy. Results on real-world datasets show that our approach significantly surpasses state-of-the-art dehazing methods, delivering enhanced image clarity and detail while effectively recovering information from even the challenging thick haze.