Deepfake Detection on OmniFake Cross-Model
99.4Midjourney AccuracyVIGIL
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| VIGILBackbone=Qwen3-VL-8B2026.03 | 99.4 | 96.1 | 96.5 | 95.2 | 98.8 | 98.4 | |
| DDA2026.03 | 97.3 | 95.4 | 93.8 | 93.6 | 96.8 | 80.7 | |
| VIGIL Cold-StartBackbone=Qwen3-VL-8B, mode=Cold-Start2026.03 | 96.2 | 93.7 | 92.7 | 91.2 | 95.4 | 87.5 | |
| Co-SPY2026.03 | 94.5 | 95.1 | 92.7 | 94 | 96.2 | 71.5 | |
| Gemini-3-Pro2026.03 | 91.6 | 72.3 | 91.6 | 66.8 | 79 | 78.9 | |
| AIDE2026.03 | 91.1 | 83.9 | 77.9 | 71.3 | 94.2 | 71.4 | |
| GPT-5.22026.03 | 90.8 | 62.6 | 75.5 | 66.4 | 79.1 | 68.1 | |
| Veritas2026.03 | 89.5 | 94.3 | 93 | 87.5 | 95.8 | 86.5 | |
| Kimi-K2.52026.03 | 87.3 | 61.8 | 79.1 | 62.8 | 77.5 | 62.2 | |
| D^32026.03 | 86.3 | 88.3 | 90 | 87.8 | 96.7 | 62.4 | |
| Qwen3-VL-8Bzero-shot=true2026.03 | 84.9 | 52.2 | 58.2 | 51.6 | 47.3 | 59.7 | |
| FakeVLM2026.03 | 77.5 | 81.5 | 76.8 | 75 | 79 | 74.8 | |
| GLM-4.6V2026.03 | 72.1 | 54.4 | 64.4 | 55.4 | 53 | 64.8 | |
| UnivFD2026.03 | 66.3 | 83 | 70.7 | 79 | 97 | 74.6 | |
| SAFE2026.03 | 65.2 | 78.8 | 54.8 | 59.9 | 74.4 | 59.5 | |
| C2P-CLIP2026.03 | 63.4 | 83.1 | 72.8 | 72.9 | 83.9 | 73.6 | |
| NPR2026.03 | 62.9 | 78.9 | 67 | 74.9 | 95.5 | 71.2 |