Predicting seizures using electroencephalogram (EEG) data is the focus of this research, which looks at the use of Vision Transformers (ViTs). In order to make two ViT models work with multi-channel time-series EEG data, one using the Swin architecture and the other using SegFormer, we implement new data pretreatment methods to convert the EEG segments to a format that ViT can read. We use important parameters including accuracy, sensitivity, specificity, and AUC-ROC to evaluate the models using MLSPred-Bench, a benchmark dataset with 12 configurations. The ViT-1 greatest overall AUC (87.57%) on longer horizons demonstrates its superior trade-off between sensitivity and specificity, whereas ViT-2 outperforms other models on benchmarks with shorter seizure prediction horizons. The results show that transformer-based designs may use EEG data well with individualised adjustments, which opens up new possibilities for seizure prediction.