Publication Date

Spring 2026

Degree Type

Master's Project

Degree Name

Master of Science in Computer Science (MSCS)

Department

Computer Science

First Advisor

Mark Stamp

Second Advisor

Katerina Potika

Third Advisor

Fabio Di Troia

Keywords

AI-generated artwork, deep learning, diffusion transformers, generator shift, out-of-distribution detection, Stable Diffusion 3.5

Abstract

Text-to-image generative models have advanced rapidly, with newer Diffusion Transformer architectures producing artwork that is difficult to distinguish from human-made images. Most AI-art detectors are trained and evaluated on the same generator family, leaving robustness to newer architectures underexplored. This project studies generator shift by constructing a 10,000-image prompt-aligned Stable Diffusion 3.5 Medium (SD3.5m) artwork dataset across ten art styles through reverse prompting of held-out human artworks. Five detectors are trained on U-Net-based latent diffusion artwork and evaluated in a zero-shot cross-generator setting on the SD3.5m dataset. Deep learning models perform strongly in-distribution but degrade under generator shift, misclassifying many SD3.5m images as human while keeping human false positives low. CLIP ViT-L/14 performs best overall and Grad-CAM analysis reveals weaker and more diffuse activation on false negatives. These findings highlightageneralizationgapincurrentAI-artdetectorsandmotivatethedevelopment of detection methods that remain reliable across evolving generative architectures.

Download

Available for download on Sunday, May 23, 2027

Share

COinS