AI-generated image detection on Chameleon (Real/Fake/Overall Accuracy)
82.18Overall AccuracyForeAgent
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| ForeAgentCategory=MLLM-based Methods2026.06 | 82.18 | 96.69 | 61.86 | |
| GPT-5Category=MLLM-based Methods2026.06 | 81.91 | 92.74 | 67.5 | |
| AIGI-HolmesCategory=MLLM-based Methods, Model size=7B2026.06 | 75.9 | — | — | |
| GPT-5-miniCategory=MLLM-based Methods2026.06 | 69.65 | 94.4 | 36.71 | |
| AIDECategory=Feature-based Forensic Methods2026.06 | 65.77 | 95.06 | 26.8 | |
| Qwen3-VL-32B-InstructCategory=MLLM-based Methods, Model size=32B2026.06 | 62.33 | 96.96 | 17.79 | |
| EffortCategory=Feature-based Forensic Methods2026.06 | 60.65 | 41.66 | 85.92 | |
| Qwen3-VL-8B-InstructCategory=MLLM-based Methods, Model size=8B2026.06 | 60.63 | 98.67 | 10.03 | |
| UnivFDCategory=Feature-based Forensic Methods2026.06 | 60.42 | 41.56 | 85.52 | |
| DireCategory=Feature-based Forensic Methods2026.06 | 57.83 | 99.73 | 2.09 | |
| NPRCategory=Feature-based Forensic Methods2026.06 | 57.81 | 100 | 1.68 | |
| PatchCraftCategory=Feature-based Forensic Methods2026.06 | 55.7 | 96.52 | 1.39 | |
| Llama-3.2-11B-Vision-InstructCategory=MLLM-based Methods, Model size=11B2026.06 | 55.59 | 76.23 | 28.15 |