Generative image detection on FakeClue (test)
90.3Overall AccuracyLTE-32B
Evaluation Results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| LTE-32BSetting=Zero-shot, Parameters=32B2025.10 | 90.3 | — | — | — | — | — | 0.172 | |
| LTE-7BSetting=Zero-shot, Parameters=7B2025.10 | 87.1 | — | — | — | — | — | 0.157 | |
| Fake-HR1Training Stage=HRL, Reasoning Mode=Non-reasoning2026.02 | 86.92 | 91.65 | 83.52 | 84.24 | 89.16 | 87.12 | — | |
| Fake-HR1Training Stage=HRL, Reasoning Mode=Reasoning2026.02 | 85.96 | 94.41 | 84.15 | 81.17 | 88.16 | 86.71 | — | |
| FakeShieldSetting=Zero-shot2025.10 | 73.3 | — | — | — | — | — | 0.106 | |
| Fake-HR1Training Stage=SFT, Reasoning Mode=Reasoning2026.02 | 72.56 | 99.83 | 72.94 | 57.11 | 72.7 | 72.79 | — | |
| Fake-HR1Training Stage=SFT, Reasoning Mode=Non-reasoning2026.02 | 72.5 | 100 | 72.45 | 56.92 | 72.55 | 72.51 | — | |
| InternVL3-8B2026.02 | 55.84 | 69.14 | 57.04 | 48.31 | 60.49 | 59.24 | — | |
| GPT-4o2026.02 | 47.4 | — | — | — | — | 42 | — | |
| Qwen2.5-VL-7B2026.02 | 35.4 | 81.91 | 52.44 | 9.05 | 16.5 | 29.5 | — | |
| LEGIONSetting=Zero-shot2025.10 | 25.4 | — | — | — | — | — | 0.028 |