AI-generated Image Detection on So-Fake OOD
79.9Flux.1 pro AccuracyFGTS
Evaluation Results
| Method | Links | |||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| FGTSEvaluation protocol=Linear probe, Backbone=DINOv3-ViT-7B2025.11 | 79.9 | 96.04 | 96.02 | 63.8 | 95.91 | 92.83 | 90.53 | 86.57 | 77.87 | 95.88 | 87.53 | |
| FGTSEvaluation protocol=Training free, Backbone=DINOv3-ViT-7B2025.11 | 68.9 | 84.2 | 83.33 | 56.96 | 83.55 | 79.59 | 69.9 | 71 | 67.42 | 85.77 | 75.06 | |
| CLIP-LDadapted using=LDM-generated images2025.11 | 66.8 | 69.67 | 71.13 | 53.15 | 66.69 | 77.42 | 67.06 | 73.71 | 68.38 | 74.14 | 68.82 | |
| Community-Forensics2025.11 | 59.37 | 86.14 | 80.48 | 66.53 | 77.68 | 75.36 | 77.74 | 75.65 | 58.59 | 81.87 | 75.61 | |
| AIDEfine-tuned on=SD v1.42025.11 | 51.15 | 88.55 | 63.21 | 53.15 | 90.31 | 86.22 | 72.81 | 68.29 | 70.71 | 92.61 | 73.7 | |
| AIDEfine-tuned on=ProGAN2025.11 | 49.91 | 97.42 | 56.46 | 51.31 | 95.71 | 94.71 | 77.89 | 71.91 | 59.83 | 96.12 | 75.13 | |
| UniverFD2025.11 | 49.58 | 49.63 | 65.43 | 45.29 | 46.95 | 45.96 | 47.15 | 47.93 | 49.66 | 48.18 | 49.57 | |
| CoDE2025.11 | 48.71 | 42.36 | 41.03 | 43.62 | 45.18 | 46.03 | 45.88 | 49.66 | 46.74 | 44.39 | 45.37 |