Reproduce “UniAIDet: A Unified and Universal Benchmark for AI-Generated Image Content Detection and Localization” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Huixuan Zhang, Xiaojun Wan
Researchers at Peking University introduced UniAIDet, a comprehensive benchmark comprising 80,000 images from 20 diverse generative models across photographic and artistic categories, offering ground truth masks for both detection and localization of AI-generated content. This benchmark demonstrates that current methods struggle with generalization across diverse generative models and image types.
Key insights
- For partially generated images, detection accuracy and localization Intersection over Union (IoU) generally show consistent trends, suggesting that improving one aspect often benefits the other.
- Current methods demonstrate significant generalization failures across different generative models, particularly struggling with partial synthesis methods and newly released closed-source models.
- Existing approaches show limited generalization across image categories, with performance degrading when applied to artistic images compared to photographic images, especially for partial synthesis tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.