Percept-Lens

Percept-Lens is like an exam paper that tests how well a method can tell AI-generated images apart from real photos. It combines 39 different public datasets into a single collection of 7.1 million images, and pits various detection methods against each other under the same conditions.

Just as grading many students fairly requires giving them the same test questions, this benchmark standardizes both the way images are converted into numbers and the tendencies of the training data used for comparison. Without matching these conditions, there's no way to tell whether a detection method is genuinely better or just scored well because of favorable setup.

Sakana AI's researchers found something interesting using this benchmark: methods that separate real from fake using nothing but statistical formulas like mean and variance, without any dedicated training, performed on par with or even better than existing detectors that had been carefully trained.

It shows up in articles when describing scale, as in "verified on 7.1 million images across 39 datasets." Percept-Lens itself isn't an image-generating AI — it's a testing tool that measures the performance of image-detection methods, so it's worth not confusing the two.