Accurate detection of thin cracks in road surfaces remains a significant challenge with pixel-level precision. In this work, we evaluate the performance of U-Net variants including ThinCrack U-Net and a pretrained VGG-16 encoder. Models are trained using different loss functions and different optimizers like Stochastic Gradient Descent with learning rate reduction on plateau. Experiments on the CrackTree260 dataset (260 annotated road images) assess precision, recall, F1-score and mean Intersection over Union, using both exact-match and some pixel dilation tolerance. While ThinCrack U-Net was designed to enhance narrow crack detection, results show it underperforms the baseline models across several metrics. Notably, U-Net with a combination of Binary Cross Entropy and Dice loss functions yields the best F1-score for exact-match evaluation. We discuss how these results reflect on the limitations of current architectures in capturing true crack width, and propose directions for model refinement and loss design in future work.