Robustness
How the detector holds up across generation methods · 4,500 held-out samples
Accuracy
91.40%
Precision
93.57%
Recall
89.51%
AUC-ROC
0.9486
By generation method
StyleGAN294.8%
inswapper_12891.5%
FaceSwap89.3%
DeepFaceLab85.1%
Stable Diffusion78.4%
Accuracy drops on diffusion-generated faces because the training set was built from face-swap pipelines rather than full image synthesis.
Confusion matrix
Called real
Called fake
Is real
2,031
143
Is fake
244
2,082
143 real photographs were wrongly flagged — a false-positive rate of 6.6%. In a moderation setting that number matters more than headline accuracy, because every one of them is a real person being accused of posting a fake.