Statistical evidence suggests that underwater scenarios have object detection difficulties for models that fall behind terrestrial scenarios by an average of 20–30%, based on factors such as attenuated light, adverse water conditions, and backgrounds, among others. This paper attempts to explain some of the newer models that can be used to advance underwater object detection performance under various environmental settings. Five state-of-the-art models are tested: Gated Cross-Domain, Boosting RCNN, Dynamic YOLO, Learning Degradation, and DMCL. This study evaluates state-of-the-art models using key metrics such as precision, recall, F1-score, mean Average Precision (mAP), and IoU across three datasets: RUOD, URPC2023, and UODD. The Gated Cross-Domain Model outperforms all others, achieving the highest precision (88.7%, 87.2%, 86.5%), recall (80.1%, 83.6%, 82.3%), F1-score (84.3%, 85.4%, 84.4%), mAP (83.5%, 84.8%, 83.2%), and IoU (80.3%, 79.5%). Dynamic YOLO follows with competitive F1-scores (82.9%, 84.4%) and mAP (84.2%, 82.1%). Boosting RCNN excels in precision (85.2%) but lags in recall, while the Learning Degradation Model and DMCL show lower recall and mAP. The findings highlight the crucial role of hyperparameter tuning in improving both model performance and computational efficiency. The Gated Cross-Domain Model’s CFI module and gated feature fusion enhance detection in challenging underwater conditions, emphasizing the importance of robust algorithms and optimized parameters for achieving higher accuracy and efficient training.