Subjective Verification on Expert-annotated prompt-image pairs 360 stratified
0.29MAEHuman
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| HumanProtocol=leave-one-out cross-validation2026.05 | 0.29 | 0.97 | 0.83 | |
| GPT-5.4Framework=AtelierJudge (Ours)2026.05 | 0.33 | 0.95 | 0.81 | |
| GPT-5.2Framework=AtelierJudge (Ours)2026.05 | 0.34 | 0.93 | 0.79 | |
| Gem-3Framework=AtelierJudge (Ours)2026.05 | 0.35 | 0.93 | 0.77 | |
| Cl-4.5Framework=AtelierJudge (Ours)2026.05 | 0.37 | 0.91 | 0.73 | |
| Gem-3Framework=Base2026.05 | 0.65 | 0.68 | 0.51 | |
| GPT-5.2Framework=Base2026.05 | 0.72 | 0.64 | 0.56 | |
| GPT-5.4Framework=Base2026.05 | 0.74 | 0.67 | 0.55 | |
| Cl-4.5Framework=Base2026.05 | 0.78 | 0.61 | 0.48 |