Object detection has become a key part in many modern computer vision systems, allowing applications such as autonomous steering, supervision, medical diagnostics and robotics. With the rapid development of deep learning techniques, many object detection models have appeared, each offering a compromise between speed and accuracy. This article presents a detailed comparative analysis of the three state-of-the-art models: Yolov8, Yolov8 integrated with RCNN and Jolov8 integrated with efficiency. This article presents a detailed comparative analysis of the three state-of-the-art object detection models: Yolov8, Yolov8 integrated with RCNN and Yolov8 integrated with effectively. Each model is evaluated on a standardized data set of actual and synthetic facial images across different visual conditions, by accuracy, inference time and average confidence as key metrics. We evaluate these models on a standardized data file containing real and synthetic images of face with different lighting, occlusion and background condition. Our analysis focuses on three basic metrics: detection accuracy, inference rate and average confidence score. Experimental results show that Yolov8 + efficient has achieved the highest accuracy of 90.1% and the average reliability score of 0.877, thanks to its robust fusion based on BIFPN. Yolov8 + RCNN followed the accuracy of 88.7%, while Yolov8 showed the fastest inference speed at just 8.6 ms per picture. The finding shows that while Yolov8 is ideal for real -time applications with limited computational resources, hybrid models such as Yolov8 + EfectionDet are more suitable for accurate critical tasks where latency can be tolerated. Through this study, we provide valuable knowledge about the model’s compromise and help lead the selection of object detection architectures based on specific applications.