Colorization of grayscale images has found widespread applications across artistic, historical, scientific, medical, and industrial domains. Traditional manual colorization methods, however, are labor-intensive, time-consuming, and prone to subjective interpretation. In recent years, various deep learning (DL) approaches have been explored to automate the colorization process. Despite this progress, the efficacy and accuracy of these DL methods remain largely unexamined. One of the primary challenges is the accurate handling of color boundaries, which significantly affects the realism and visual clarity of the colorized output. In this study, we investigated the impact of different DL algorithms and training epochs on the quality of colorized images, comparing them to their corresponding ground truths. To explore these effects, a semantic-guided adversarial generative framework was employed, and five commonly used DL algorithms were compared.