A novel approach is presented in this paper that contributes to haze removal in images, departing from traditional methods reliant on estimating transmission maps. Single image scenarios often pose challenges as the images lack in-depth information, rendering the traditional task ill-posed. The paper introduces a complete learning-based methodology employing an improved and conditional Generative Adversarial Network (GAN) to address haze removal directly. The Tiramisu model is adopted as the generator instead of the classic U-Net model, chosen for its superior performance and parameter efficiency. Additionally, a patch-based discriminator is implemented to lighten the artifacts in the unhazed image. To enhance the perceptual quality of results, a hybrid weighted loss function is employed during model training. Experimental evaluation conducted on both real-world hazy and synthetic images demonstrates the competitiveness of our model against traditional methods.