A new detection method catches AI-generated images by zeroing in on the parts diffusion models render laziest.
Researchers describe EIB-Net, an "entropy-guided information bottleneck network," in a paper posted to arXiv on September 25, 2026 (arXiv:2609.29064, https://arxiv.org/abs/2609.29064). The method first scans an image for its lowest-entropy patch - the flattest, least-detailed region, like a stretch of sky or a plain wall - because the researchers found generative models nail overall composition but skimp on fine texture there. That patch gets fed through a variational information bottleneck, a technique that compresses it into a compact set of features while discarding anything that does not help tell real from fake. On the DIFF, DiffusionForensics, and GenImage benchmarks, EIB-Net hit 85.7% accuracy while training on just 2% of the usual data, beat full-image baseline detectors by more than 15 points, and held up on generators it was not trained on, averaging 83.5% accuracy on GenImage.
That data efficiency is the real story. Most AI-image detectors need retraining every time a new model version shifts the artifacts it leaves behind, which means constant retraining on huge labeled datasets. A detector that generalizes across generators using a sliver of that data is what would actually make detection deployable at platform scale, rather than a benchmark trophy that goes stale in a few months.
Texture-based tells like this have a track record of working right up until generator makers patch them - so call this a good diagnostic for today's diffusion models, not a permanent fix.