Art or Artifact? Segmenting AI-Generated Images for Deeper Detection

Abstract

In recent years, generative models have made remarkable progress by producing images with highly realistic visual effects that are often indistinguishable from real photographs. These advances have driven innovation in creative fields, but have also introduced new risks and social concerns, making the detection of AI-generated images an increasingly important research topic. Among the various applications, the detection of AI-generated artworks has attracted particular attention. However, most existing approaches focus solely on image-level classification, overlooking fine-grained pixel-level analyses. This limitation reduces the effectiveness of partially detecting manipulated images. To address this issue, we propose a multitask learning framework for detecting AI-generated artworks. Our method jointly performs image-level classification and pixel-level segmentation, enabling the model to assess the overall authenticity of an image while precisely localizing the AI-generated regions. The experimental results demonstrate that this method enhances the classification accuracy of backbone models, yielding improvements of 1.97% for ResNet50 and 1.29% for the Vision Transformer. Moreover, the proposed approach improves generalization to out-of-distribution (OOD) data, with performance gains of 4.56% and 3.73% on ResNet50 and ViT, respectively. These results highlight the effectiveness and practical potential of incorporating pixel-wise supervision in the detection of AI-generated content.