Text-to-Image Generation on T2I-CompBench out-of-domain
51.37Semantic ConsistencyFLUX.1-dev + UnifiedReward-Flex
Evaluation results
| Method | Links | |
|---|---|---|
| FLUX.1-dev + UnifiedReward-FlexBackbone=FLUX.1-dev, Optimization=GRPO, Reward Model=UnifiedReward-Flex, Sampling steps=30, CFG scale=3.52026.02 | 51.37 | |
| FLUX.1-dev + UnifiedReward-ThinkBackbone=FLUX.1-dev, Optimization=GRPO, Reward Model=UnifiedReward-Think, Sampling steps=30, CFG scale=3.52026.02 | 50.1 | |
| FLUX.1-dev + UnifiedRewardBackbone=FLUX.1-dev, Optimization=GRPO, Reward Model=UnifiedReward, Sampling steps=30, CFG scale=3.52026.02 | 49.13 | |
| FLUX.1-devSampling steps=30, CFG scale=3.52026.02 | 48.57 | |
| FLUX.1-dev + HPSv3Backbone=FLUX.1-dev, Optimization=GRPO, Reward Model=HPSv3, Sampling steps=30, CFG scale=3.52026.02 | 46.46 | |
| FLUX.1-dev + PickScoreBackbone=FLUX.1-dev, Optimization=GRPO, Reward Model=PickScore, Sampling steps=30, CFG scale=3.52026.02 | 45.92 | |
| FLUX.1-dev + HPSv2Backbone=FLUX.1-dev, Optimization=GRPO, Reward Model=HPSv2, Sampling steps=30, CFG scale=3.52026.02 | 44.84 |