Text-to-Image Semantic Consistency on UniGenBench
69.46Overall Semantic ConsistencyFLUX.1-dev w/ Pref-GRPO
Evaluation results
| Method | Links | |||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| FLUX.1-dev w/ Pref-GRPOBase Model=FLUX.1-dev, Fine-tuning Strategy=Pref-GRPO2025.08 | 69.46 | 88.4 | 90.35 | 75 | 69.77 | 76.52 | 44.09 | 63.27 | 62.43 | 77.61 | 47.13 | |
| FLUX.1-dev w/ UnifiedRewardBase Model=FLUX.1-dev, Fine-tuning Strategy=Flow-GRPO, Reward Model=UnifiedReward2025.08 | 63.62 | 86.1 | 89.72 | 71.55 | 63.69 | 70.42 | 32.05 | 62.43 | 52.32 | 73.51 | 34.44 | |
| FLUX.1-dev w/ HPSv2&CLIPBase Model=FLUX.1-dev, Fine-tuning Strategy=Flow-GRPO, Reward Model=HPSv2&CLIP2025.08 | 61.81 | 84.92 | 88.98 | 68.44 | 62.54 | 68.1 | 31.01 | 59.36 | 50.6 | 71.07 | 33.07 | |
| FLUX.1-devBase Model=FLUX.1-dev2025.08 | 61.3 | 83.9 | 88.92 | 67.84 | 62.17 | 67.26 | 30.91 | 60.96 | 47.04 | 71.83 | 32.18 | |
| FLUX.1-dev w/ HPSv2Base Model=FLUX.1-dev, Fine-tuning Strategy=Flow-GRPO, Reward Model=HPSv22025.08 | 58.77 | 75.2 | 88.77 | 66.56 | 58.94 | 66.88 | 28.18 | 58.02 | 45.88 | 67.91 | 31.32 |