IMPLEMENTASI MODEL DeiT UNTUK MEMBEDAKAN GAMBAR BUATAN AI DAN MANUSIA PADA ILUSTRASI ANIMASI 2D

The development of artificial intelligence (AI) has influenced various fields, including art and visual design. AI Generative Art, which mimics human styles, has sparked debates on originality, artistic value, as well as legal and ethical challenges. Therefore, methods are needed to distinguish between AI-generated and human-made images, particularly in 2D animation illustrations. This study proposes the use of Data-efficient Image Transformers (DeiT) for image classification. Two models tested are DeiT Base and DeiT Tiny, using a dataset of 6,000 images equally divided between AI and human categories. The dataset is split into training (70%), validation (15%), and testing (15%). Experimental results show that DeiT Base achieves over 95% accuracy with fast convergence and optimal loss function stability. Meanwhile, DeiT Tiny attains around 93% accuracy, being more computationally efficient despite requiring more epochs for stability. Compared to previous models using a larger dataset (11,000 images per category) but achieving only 80% accuracy, DeiT performs better in both accuracy and computational efficiency, even with a smaller dataset. In conclusion, DeiT is effective for classifying 2D animation images. DeiT Base excels in accuracy and convergence speed, while DeiT Tiny is more resource-efficient, making it an ideal choice for environments with computational constraints.