Sakana AI says a frozen vision model can distinguish real from AI-generated images

Sakana AI says its Percept-Lens evaluation framework showed that released AI-generated image detectors can degrade sharply when generators, prompts, styles or image domains change. The team describes a new detector that keeps a general-purpose vision model unchanged and fits a Gaussian decision rule to its features, classifying images by whether they more closely resemble labeled real or AI-generated examples. On the same broad evaluation suite, Sakana AI says its detector outperformed the strongest released detector it tested, even though the underlying vision model had not been trained specifically for this task.