Comparative Analysis of Vision Transformer (DeiT) and Transfer Learning-Based CNN (ResNet50) Performance in Identifying AI-Generated 2D Animation Illustrations

Abstract

The development of generative artificial intelligence has introduced new challenges in the digital creative industry, particularly in distinguishing human-created 2D anime illustrations from AI-generated images due to their increasingly similar visual characteristics. This study aims to compare the performance of a Vision Transformer model, the Data-efficient Image Transformer (DeiT), and a transfer-learning-based Convolutional Neural Network model, ResNet50, for detecting AI-generated 2D anime illustrations. The study employed the SEMMA methodology, consisting of the Sample, Explore, Modify, Model, and Assess stages. An initial dataset of 2,000 images was collected from Safebooru and Civitai, of which 1,736 images were validated and used in the experiments. The images underwent preprocessing, including resizing, normalization, data augmentation, and splitting the dataset into training, validation, and test sets. Model performance was evaluated using accuracy, precision, recall, F1-score, confusion matrix, and ROC-AUC metrics. The experimental results showed that ResNet50 achieved 94% accuracy with an AUC of 0.993, while DeiT achieved a slightly higher accuracy of 95% with an AUC of 0.990. Although both models demonstrated excellent discriminative power, DeiT achieved slightly better overall classification performance under the same experimental setting. These findings indicate that Vision Transformer-based models have strong potential for detecting AI-generated anime illustrations and can contribute to the development of more reliable digital artwork authenticity detection systems.