A new detection framework keeps AI image detectors current on new generators without retraining them from scratch.
Researchers built EvoKnow, a system that treats catching AI-generated images as an ongoing problem rather than a one-time training job. Instead of replaying old training images or overwriting what a model already knows, EvoKnow keeps a shared base of forensic knowledge and bolts on small 'expert' modules for each new image generator it encounters. A routing mechanism, called the Analytical Incremental Router, decides which expert to consult and updates itself using statistics from new images rather than full retraining. In tests, the system hit 96.70% average accuracy on GenImage generators it had not been trained on and 94.48% on a benchmark called Chameleon, using just ten sample images per new generator.
This matters because image generators multiply faster than detectors can keep up. A detector trained on last year's generators can miss images from whatever shipped last month, and retraining on old and new data together is expensive and risks degrading what already works. EvoKnow's approach of adding isolated modules instead of overwriting shared parameters produced only 4.32% average forgetting across a long sequence of new generators, a sign the method does not quietly erode old detection skills to learn new ones.
It is still a lab benchmark, not a deployed moderation tool, so treat the accuracy figures as a ceiling rather than a guarantee against tomorrow's generator.