Synthetic Image Detection on ForenSynths, DIRE, and UniversalFake suites Setting 2 (test)
100ProGAN Detection RateDIRE
Evaluation Results
| Method | Links | |||||||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| DIRETraining Dataset=ProGAN, Backbone=CLIP ViT-L/14, Resolution=224x2242026.04 | 100 | 67.73 | 64.78 | 83.08 | 65.3 | 100 | 80.16 | 86.31 | 86.36 | 72.37 | 83.2 | 82.7 | 84.05 | 84.25 | 87.1 | 90.8 | 90.25 | 58.75 | 81.51 | |
| RINETraining Dataset=ProGAN, Backbone=CLIP ViT-L/14, Resolution=224x2242026.04 | 100 | 99.3 | 99.6 | 88.9 | 99.8 | 99.5 | 97.82 | 78.9 | 96.08 | 89.2 | 76.1 | 98.3 | 88.2 | 98.6 | 88.9 | 92.6 | 90.7 | 85 | 92.64 | |
| EffortTraining Dataset=ProGAN, Backbone=CLIP ViT-L/14, Resolution=224x2242026.04 | 100 | 99.85 | 99.6 | 95.05 | 99.6 | 100 | 96.9 | 67.62 | 93.04 | 87.83 | 69.15 | 99.3 | 96.8 | 99.45 | 97.45 | 97.8 | 97.15 | 98.05 | 94.15 | |
| I2PTraining Dataset=ProGAN, Training Sample Size=1600 images per set, Backbone=CLIP ViT-L/14, Resolution=224x2242026.04 | 99.99 | 99.94 | 99.6 | 97.32 | 99.6 | 99.84 | 96.22 | 86.96 | 96.33 | 89.4 | 85.24 | 98.78 | 96.39 | 99.75 | 98.2 | 98.51 | 98.43 | 98.2 | 96.59 | |
| ForgeLensTraining Dataset=ProGAN, Backbone=CLIP ViT-L/14, Resolution=224x2242026.04 | 99.95 | 99.24 | 97.67 | 96.64 | 98.84 | 95.24 | 96.2 | 81.23 | 98.44 | 89.42 | 73.34 | 98.72 | 96.98 | 98.86 | 96.07 | 96.17 | 95.43 | 98.29 | 94.82 | |
| FerretNetTraining Dataset=ProGAN, Backbone=CLIP ViT-L/14, Resolution=224x2242026.04 | 99.9 | 98.8 | 92.6 | 98 | 91.4 | 99.1 | 98.5 | 88.82 | 94.9 | 86 | 86.4 | 98.8 | 98.5 | 98.8 | 97.3 | 97.2 | 97.9 | 91.4 | 95.24 | |
| UniFDTraining Dataset=ProGAN, Backbone=CLIP ViT-L/14, Resolution=224x2242026.04 | 99.9 | 98.4 | 94.65 | 84.95 | 99.4 | 96.75 | 83.35 | 71.5 | 86.15 | 73.35 | 69.65 | 94.4 | 74 | 95 | 78.5 | 79.05 | 77.9 | 87.3 | 85.79 | |
| FatFormerTraining Dataset=ProGAN, Backbone=CLIP ViT-L/14, Resolution=224x2242026.04 | 99.89 | 99.32 | 99.5 | 97.15 | 99.41 | 99.75 | 99.8 | 76.82 | 99.3 | 84.53 | 76 | 98.6 | 94.9 | 98.65 | 94.35 | 94.65 | 94.2 | 98.75 | 94.75 | |
| LGardTraining Dataset=ProGAN, Backbone=CLIP ViT-L/14, Resolution=224x2242026.04 | 99.83 | 86.94 | 86.63 | 91.08 | 78.46 | 99.27 | 54.71 | 72.3 | 67.5 | 68.92 | 60.34 | 90.3 | 95.1 | 93.61 | 86.13 | 90.3 | 83.24 | 67.6 | 81.79 | |
| NPRTraining Dataset=ProGAN, Backbone=CLIP ViT-L/14, Resolution=224x2242026.04 | 99.75 | 71.9 | 90.1 | 94.3 | 77.7 | 100 | 87.8 | 88 | 82.3 | 83.8 | 68.35 | 89.8 | 90.1 | 89.05 | 76.55 | 80.15 | 79.5 | 85.4 | 85.25 | |
| FreqNetTraining Dataset=ProGAN, Backbone=CLIP ViT-L/14, Resolution=224x2242026.04 | 97.9 | 95.84 | 90.45 | 97.55 | 90.24 | 93.41 | 92.03 | 94.92 | 93.22 | 87.76 | 86.7 | 84.55 | 99.58 | 65.56 | 85.69 | 97.4 | 88.15 | 59.06 | 88.89 |