Diabetic Retinopathy (DR) is a major complication of diabetes, leading to vision loss if not detected early. Despite significant advancements in medical imaging, accurate and efficient diagnosis of DR remains a challenging task due to the subtle variations in retinal images. Addressing this issue is critical in developing reliable AI-driven tools for ophthalmology. Given the critical need for accurate diagnosis, this paper presents an ensemble learning approach utilizing EfficientNetB0 and EfficientNetB3 models for robust DR classification. The method achieves high accuracy by combining the strengths of these architectures, validated on comprehensive datasets such as Messidor, Eyepacs, and RFMiD, without relying on complex data augmentation techniques. A validation accuracy of 91% and a test accuracy of 96.19% are achieved, demonstrating the potential for enhancing AI-driven diagnostic tools in clinical ophthalmology.