Fine-grained Detection on MMFakeBench 1000 (val)
38.9F1 ScoreREFORM
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| REFORMLanguage Model=Florence2-0.3B, Prompt Method=Ours (Sec.3.2), Zero-shot=true2026.03 | 38.9 | 34.6 | 48.9 | 57.3 | |
| LLaVA-1.6Language Model=Vicuna-13B, Prompt Method=MMD-Agent, Zero-shot=true2026.03 | 38 | 44.5 | 41 | 40.6 | |
| BLIP2Language Model=FlanT5-XXL, Prompt Method=MMD-Agent, Zero-shot=true2026.03 | 31.6 | 39.8 | 32.2 | 34.4 | |
| InstructBLIPLanguage Model=Vicuna-13B, Prompt Method=MMD-Agent, Zero-shot=true2026.03 | 26 | 33.3 | 30.1 | 29.5 | |
| VILALanguage Model=LLaMA2-13B, Prompt Method=MMD-Agent, Zero-shot=true2026.03 | 22.7 | 27.3 | 24.4 | 28.7 | |
| LLaVA-1.6Language Model=Vicuna-7B, Prompt Method=Standard, Zero-shot=true2026.03 | 17.4 | 14.8 | 25.7 | 30.8 | |
| BLIP2Language Model=FlanT5-XXL, Prompt Method=Standard, Zero-shot=true2026.03 | 16.7 | 34.9 | 27.3 | 32.8 | |
| BLIP2Language Model=FlanT5-XL, Prompt Method=Standard, Zero-shot=true2026.03 | 16.4 | 20.1 | 27.5 | 33 | |
| mPLUG-Owl2Language Model=LLaMA2-7B, Prompt Method=Standard, Zero-shot=true2026.03 | 14.5 | 22.2 | 25.9 | 31.1 | |
| InstructBLIPLanguage Model=Vicuna-13B, Prompt Method=Standard, Zero-shot=true2026.03 | 13.7 | 13.2 | 24 | 28.8 | |
| LLaVA-1.6Language Model=Vicuna-13B, Prompt Method=Standard, Zero-shot=true2026.03 | 12 | 22.5 | 25 | 30 | |
| PandaGPTLanguage Model=Vicuna-7B, Prompt Method=Standard, Zero-shot=true2026.03 | 11.8 | 9.8 | 25 | 30 | |
| VILALanguage Model=LLaMA2-7B, Prompt Method=Standard, Zero-shot=true2026.03 | 11.5 | 7.5 | 25 | 30 | |
| VILALanguage Model=LLaMA2-13B, Prompt Method=Standard, Zero-shot=true2026.03 | 11.5 | 7.5 | 25 | 30 | |
| Qwen-VLLanguage Model=Qwen-7B, Prompt Method=Standard, Zero-shot=true2026.03 | 7.5 | 10.3 | 24.3 | 11 | |
| InstructBLIPLanguage Model=Vicuna-7B, Prompt Method=Standard, Zero-shot=true2026.03 | 7.1 | 7.9 | 6.5 | 7.8 | |
| Otter-ImageLanguage Model=MPT-7B, Prompt Method=Standard, Zero-shot=true2026.03 | 5.2 | 10.5 | 3.4 | 4.1 | |
| MiniGPT4Language Model=Vicuna-7B, Prompt Method=Standard, Zero-shot=true2026.03 | 5.2 | 5.2 | 21.2 | 9 |