Human Preference Optimization on REFL (300 random images)
1.9088ImageRewardGradient Ascent
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| Gradient Ascent2025.09 | 1.9088 | 0.2247 | 0.2877 | 5.5775 | 0.1474 | 0.9195 | |
| Trajectory Optimization Control2025.09 | 1.8914 | 0.2526 | 0.2904 | 6.1088 | 0.1717 | 0.9242 | |
| Inversion+TFG2025.09 | 1.7053 | 0.2362 | 0.2727 | 5.6331 | 0.2927 | 0.8401 | |
| Inversion+FreeDoM2025.09 | 1.5995 | 0.2226 | 0.2356 | 5.4951 | 0.5503 | 0.7225 | |
| Inversion+DPS2025.09 | 1.5988 | 0.2323 | 0.265 | 5.8276 | 0.2875 | 0.8505 | |
| None2025.09 | 0.1542 | 0.2385 | 0.2887 | 6.0516 | 0 | 1 |