Text-to-Image Generation on DrawBench (PickScore, CLIPScore, HPSv3, DeQA, P(real), Avg)
23.17PickScoreFlow-GRPO
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| Flow-GRPOReward=ImageDoctor2026.06 | 23.17 | 93.7 | 12.23 | 85.7 | 20.1 | 0.3 | |
| BoxFlow-GRPOReward=UR2+SDG2026.06 | 22.91 | 91.5 | 12.14 | 87.7 | 22.8 | 2.4 | |
| DenseFlow-GRPOReward=ImageDoctor2026.06 | 22.88 | 94.5 | 11.57 | 86.4 | 19.9 | -1 | |
| Base2026.06 | 22.84 | 91.2 | 11.75 | 87.6 | 21.1 | — | |
| Flow-GRPOReward=UR22026.06 | 22.78 | 90.8 | 11.27 | 88.2 | 14.9 | -6.7 |