Recognizing handwritten Odia characters is a complex challenge due to the vast number of possible combinations resulting from the large variety of vowels, consonants, conjuncts, and matraas, coupled with limited research and dataset availability. In this research, we offer a comparative examination of four deep learning models—EfficientNet-B0, ResNet-50, DenseNet-121, and Xception-71—in terms of their accuracy in Odia handwritten character recognition. We further investigate the improvements observed in these models when leveraging data augmentation techniques using Generative Adversarial Networks (GANs). Our experiments are conducted on the NITROHCSv1.0 dataset, which comprises 15,040 images of Odia characters. The results demonstrate that ResNet-50 and EfficientNet-B0 achieve the highest accuracy among the four models, with ResNet-50 attaining 100% accuracy and EfficientNet-B0 reaching 99.8% accuracy. Interestingly, the incorporation of GANs enhanced the performance of these models, with both ResNet-50 with GAN and EfficientNet-B0 with GAN achieving 100% accuracy. DenseNet-121 reported a respectable accuracy of 99.1%, while Xception-71 exhibited relatively lower performance, reaching 58.3% accuracy in 30 epochs. When augmented with GANs, DenseNet-121 and Xception-71 achieved 99.6% and 61.2% accuracy, respectively, in 30 epochs.