The Bottleneck Was Never the Detector

In detection problems, the binding constraint usually hides one layer below the model. When many cheap per-signal classifiers stack behind a learned aggregator, the model is rarely the problem; the aggregation step is. Deepfake video detection just produced the cleanest recent example.

CryptoAILab's FakeI2V-Bench, a 97,548-video public benchmark, ran a head-to-head with no aggregation. The best image-level detector posted 80.16% AUC, the best video-level detector 79.99% AUC, a 0.17-point edge either way.

Stack a random forest over statistical features of each frame's score distribution. The IV-Bridge framework from CryptoAILab inverts the result. Eleven off-the-shelf image-level detectors rise above the video-level state of the art; the best variant reaches 93.80% AUC on the FakeI2V-Bench test set.

For IV-Bridge on FakeI2V-Bench, the gain came not from better detectors but from how frame predictions were combined. A learned aggregator over cheap signals, from temporal variance to agreement, extracts structure that no single frame and no monolithic video model used. The ceiling is real but narrow: the test set is self-curated, the aggregator bespoke, 93.80% a number on the CryptoAILab benchmark, not a general claim. The next gain, when many weak per-frame signals are available, is rarely a bigger backbone. It is a better way to listen to all of them at once.