This research addresses the critical challenge of identifying AI-generated images in open-world settings where diverse image manipulation techniques are constantly evolving.

• Introduces OpenSDID, a new benchmark dataset that simulates real-world diffusion-based manipulations using various vision-language models
• Supports both detection (identifying if an image is AI-generated) and localization (pinpointing manipulated areas) tasks
• Proposes novel techniques to identify both globally and locally manipulated images
• Establishes a foundation for security measures against increasingly sophisticated AI-generated visual content

Significance: As AI image generation becomes more widespread, reliable detection methods are essential for maintaining media integrity, preventing misinformation, and establishing trust in digital content.

OpenSDI: Spotting Diffusion-Generated Images in the Open World