Vision Transformer-Based Framework for AI-Generated Image Detection in Interior Design

Abstract

Increasingly, images generated by artificial intelligence (AI) are being used within interior design as a source of authenticity and ethical use. Based on limited Convolutional Neural Network (CNN) capabilities in data descriptive processes, including long-range dependencies and global patterns, this study examines how Vision Transformer (ViT) can be utilized in detecting AI-generated interior design images. We finetuned and evaluated four ViT models, ViT-B16, ViT-B32, ViT-L16, and ViT-L32, on 1,000 samples per class dataset. Accuracy, precision, recall, F1-score, and computational efficiency were used to assess performance. Results show that models with smaller patch sizes (i.e., 16×16) perform better than larger ones (i.e., 32×32). It was found that ViT-B16 and ViT-L16 had the highest accuracy (96.25%) and F1- score (0.9625) in identifying minor inconsistencies in the AI-generated images. ViT-B32 and ViT-L32 enjoy better computational efficiency based on lower classification performance (80.00% and 81.25% accuracy, respectively, for ViT-B32 and ViT-L32). The best tradeoff between accuracy and resource efficiency turns out to be ViT-B16. However, computational costs were higher with ViT — ViT-L16, although just as accurate. Computationally, ViT-B32 and ViT-L32 were also efficient, which was more appropriate for realtime applications with lower accuracy than speed. Through this work, we contribute a domain-specific deep learning framework for AI-generated image detection in interior design to increase authenticity verification. Future work will address improving computational efficiency and generalizing the model across all (or most) generative models and design styles.