The rapid advancement of next-generation intelligent solutions in image synthesis has raised concerns about misinformation and deepfake manipulation. Existing deepfake detection models struggle to generalize across diverse datasets, limiting their effectiveness against evolving synthesis techniques. This paper proposes a diversity-driven training to enhance both accuracy and real-time detection of deepfake images. We construct a dataset, integrating deepfake images from unconditional and conditional generative adversarial networks (GAN) with real images from public datasets. We employ Ultralytics YOLOv11, a real-time object detection model, repurposed for deepfake classification by modifying its output layers and retraining it on deepfake datasets. The model was fine-tuned using a multi-stage training strategy to improve generalization across unseen deepfakes. Experimental results show that our experiment achieves superior result making it an efficient and scalable deepfake detection solution.