A frozen vision model out-detects purpose-built AI image detectors
- Source
- x.com
- Date
Can a strong general-purpose vision model distinguish real from AI-generated images using only a simple decision rule on frozen representations? Our latest results, to be presented at #ECCV2026, show it can. Blog: https://t.co/EvTnCIqkB5 Paper: https://t.co/NzvieckZlh New image generators keep appearing. A trained AI-generated image detector that works well on familiar images can fail when the generator, prompt, style, or image domain changes. Our benchmark study introduced Percept-Lens, a common evaluation framework for these shifts, and showed how sharply released AI-generated image detectors can degrade beyond familiar data. That led us to a more basic question. When a detector fails, has its underlying vision model lost the distinction between real and AI-generated images, or is its decision rule failing to recover it? In our upcoming ECCV paper, we built a new detector by keeping a general-purpose vision model frozen and fitting a simple Gaussian decision rule to its representations. The method models how labeled real and AI-generated images are arranged in the vision model’s feature space, then classifies a new image by the group it most closely resembles. On the same…
Context
Sakana AI reports that an existing general-purpose vision model can distinguish real from generated images without retraining its image representation. The researchers pass labeled examples through the unchanged model, estimate how the two groups are distributed, and classify new images by their proximity to those groups. The decision rule accounts for variation and correlations among the image features.
In the representation study’s reported comparison, the strongest configuration reached 94.46% mean class accuracy across 39 datasets, versus 87.54% for the strongest released detector tested. That result supports extracting a better decision from existing visual features. The comparison changes both the encoder and decision rule, so it does not isolate either component’s contribution.
Checking sign-in…
Loading comments…