Detecting AI-Generated Images: Feature Fusion Observations

Generative image models evolve quickly, so single-feature detectors degrade over time. The stronger strategy is combining feature families that fail differently.

What Helped Most

• Mix texture and frequency-domain signals for broader coverage.
• Evaluate by source model diversity, not single-model splits.
• Track calibration drift over time, not just top-line accuracy.

The practical goal is not theoretical perfection. It is robust detection behavior under real-world distribution shift.