Robustness under common transformations

We evaluate Seer on 13,843 held-out images from the TechJam evaluation set: 5,000 real COCO photographs and 8,843 WildFake DALL·E 3 images. The sweep measures clean performance and 15 common transformations, including recompression, blur, resizing, noise, color jitter, cropping, and the pangram submission protocol.

For comparison, the table below is the NTIRE 2026 public-test leaderboard from Table 3 of Gushchin et al.. The published entries are 7 billion parameter models. Our model, Seer, at 302 million parameters sits third on robust ROC AUC, state of the art at this scale.