UniAIDet: A Unified and Universal Benchmark for AI-Generated Image Content Detection and Localization

TL;DR

This work presents UniAIDet, a large-scale benchmark that unifies detection and localization for AI-generated images across both photographic and artistic content. It covers holistic and partial synthesis across 20 generative models, including text-to-image, image-to-image, inpainting, image editing, and deepfakes, with ground-truth masks for partially generated regions and a total of 80k real and synthetic images. Comprehensive experiments reveal that existing detection and localization methods struggle to generalize across models and content domains, with detection-vs-localization trends generally aligned but overall performance remaining far from mature. The benchmark provides a robust, end-to-end evaluation framework to drive future research toward robust, generalizable detection and localization of AI-generated imagery. The results emphasize the need for methods that handle partial synthesis and artistic content to ensure reliable authenticity assessment in real-world scenarios.

Abstract

With the rapid proliferation of image generative models, the authenticity of digital images has become a significant concern. While existing studies have proposed various methods for detecting AI-generated content, current benchmarks are limited in their coverage of diverse generative models and image categories, often overlooking end-to-end image editing and artistic images. To address these limitations, we introduce UniAIDet, a unified and comprehensive benchmark that includes both photographic and artistic images. UniAIDet covers a wide range of generative models, including text-to-image, image-to-image, image inpainting, image editing, and deepfake models. Using UniAIDet, we conduct a comprehensive evaluation of various detection methods and answer three key research questions regarding generalization capability and the relation between detection and localization. Our benchmark and analysis provide a robust foundation for future research.

Figures (10)

• Figure 1: Data distribution of our benchmark.
• Figure 2: Overview of generative models used in the benchmark.
• Figure 3: Trends of corresponding f.Acc and mIoU.
• Figure 4: f.Acc of different methods. Darker blue indicates smaller values.
• Figure 5: mIoU of different methods. Darker blue indicates smaller values.