Forgery Detection on CASIA V2.0
81AccuracyGPT-5
Evaluation results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| GPT-5Scenario=Oracle-assisted2026.06 | 81 | — | — | — | — | |
| FKGScenario=Real-world2026.06 | 71 | — | — | — | — | |
| GPT-5Scenario=Real-world2026.06 | 39 | — | — | — | — | |
| Gemma 3Scenario=Oracle-assisted2026.06 | 38 | — | — | — | — | |
| Gemini 2.0Scenario=Oracle-assisted2026.06 | 34 | — | — | — | — | |
| Sonnet 4Scenario=Oracle-assisted2026.06 | 32 | — | — | — | — | |
| Nemotron N.Scenario=Oracle-assisted2026.06 | 31 | — | — | — | — | |
| Nemotron N.Scenario=Real-world2026.06 | 30 | — | — | — | — | |
| Gemma 3Scenario=Real-world2026.06 | 25 | — | — | — | — | |
| Qwen3 VLScenario=Oracle-assisted, VLM Type=Thinking2026.06 | 23 | — | — | — | — | |
| Gemini 2.0Scenario=Real-world2026.06 | 22 | — | — | — | — | |
| Llama 4 S.Scenario=Oracle-assisted2026.06 | 20 | — | — | — | — | |
| Qwen3 VLScenario=Real-world, VLM Type=Thinking2026.06 | 6 | — | — | — | — | |
| Llama 4 S.Scenario=Real-world2026.06 | 5 | — | — | — | — | |
| Sonnet 4Scenario=Real-world2026.06 | 3 | — | — | — | — | |
| Qwen3 VLScenario=Real-world, VLM Type=Instruct2026.06 | 1 | — | — | — | — | |
| Qwen3 VLScenario=Oracle-assisted, VLM Type=Instruct2026.06 | 0 | — | — | — | — | |
| DeepSignatureEvaluation Subset (Authentic vs Forged)=Authentic, Number of Images=852, Optimal threshold tau_ts for F1=0.632026.04 | — | 100 | 0.27 | 100 | 99.6 | |
| DeepSignatureEvaluation Subset (Authentic vs Forged)=Forged, Number of Images=1619, Optimal threshold tau_ts for F1=0.632026.04 | — | 51 | 0.97 | 100 | 0.7 |