Text-to-Image Generation on UniGenBench 600 prompts (test)
36.4HPS-v2 ScorePave-GRPO
Evaluation results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| Pave-GRPOReward Model=HPS-v2, Backbone=Flux.1-dev, SDE sampling strategy=marginal-preserving (F^mar_sde), Denoising steps=502026.06 | 36.4 | 15.12 | 36.72 | 1.3331 | 3.5396 | 3.4181 | 23.25 | |
| Pave-GRPOReward Model=HPS-v2, Backbone=Flux.1-dev, SDE sampling strategy=SNR-preserving (F^snr_sde), Denoising steps=502026.06 | 36.24 | 15.1 | 36.88 | 1.3003 | 3.5467 | 3.4362 | 23.14 | |
| Flow-CPSReward Model=HPS-v2, Backbone=Flux.1-dev, Denoising steps=502026.06 | 34.92 | 14.8 | 37.45 | 1.2561 | 3.5276 | 3.4242 | 23.11 | |
| Pave-GRPOReward Model=HPS-v2+CLIP, Backbone=Flux.1-dev, SDE sampling strategy=marginal-preserving (F^mar_sde), Denoising steps=502026.06 | 34.84 | 14.53 | 39.57 | 1.3841 | 3.702 | 3.3978 | 23.27 | |
| Flow-GRPOReward Model=HPS-v2, Backbone=Flux.1-dev, Denoising steps=502026.06 | 34.63 | 14.44 | 37.54 | 1.2435 | 3.5167 | 3.3975 | 22.84 | |
| Pave-GRPOReward Model=HPS-v2+CLIP, Backbone=Flux.1-dev, SDE sampling strategy=SNR-preserving (F^snr_sde), Denoising steps=502026.06 | 34.36 | 14.45 | 39.1 | 1.3115 | 3.6668 | 3.3955 | 23.15 | |
| Dance-GRPOReward Model=HPS-v2, Backbone=Flux.1-dev, Denoising steps=502026.06 | 34.3 | 14.47 | 37.24 | 1.2178 | 3.5129 | 3.3878 | 23.01 | |
| Flow-CPSReward Model=HPS-v2+CLIP, Backbone=Flux.1-dev, Denoising steps=502026.06 | 33.64 | 14.28 | 39.37 | 1.2889 | 3.6482 | 3.3877 | 23.14 | |
| Flow-GRPOReward Model=HPS-v2+CLIP, Backbone=Flux.1-dev, Denoising steps=502026.06 | 33.54 | 14.27 | 39.22 | 1.2496 | 3.6181 | 3.3919 | 23.03 | |
| Pave-GRPOReward Model=HPS-v3, Backbone=Flux.1-dev, SDE sampling strategy=marginal-preserving (F^mar_sde), Denoising steps=502026.06 | 33.44 | 15.18 | 37.22 | 1.1658 | 3.5431 | 3.389 | 23.14 | |
| Dance-GRPOReward Model=HPS-v2+CLIP, Backbone=Flux.1-dev, Denoising steps=502026.06 | 33.18 | 14.18 | 39.32 | 1.2307 | 3.5848 | 3.3699 | 22.99 | |
| Pave-GRPOReward Model=HPS-v3+CLIP, Backbone=Flux.1-dev, SDE sampling strategy=marginal-preserving (F^mar_sde), Denoising steps=502026.06 | 33.15 | 14.62 | 39.58 | 1.2749 | 3.6591 | 3.3834 | 23.19 | |
| Pave-GRPOReward Model=HPS-v3, Backbone=Flux.1-dev, SDE sampling strategy=SNR-preserving (F^snr_sde), Denoising steps=502026.06 | 33.06 | 15.05 | 37.26 | 1.1421 | 3.4896 | 3.3705 | 22.85 | |
| Flow-CPSReward Model=HPS-v3, Backbone=Flux.1-dev, Denoising steps=502026.06 | 32.91 | 14.76 | 37.83 | 1.1598 | 3.5375 | 3.3883 | 23.02 | |
| Dance-GRPOReward Model=HPS-v3, Backbone=Flux.1-dev, Denoising steps=502026.06 | 32.75 | 14.67 | 38.13 | 1.1431 | 3.546 | 3.3879 | 23 | |
| Pave-GRPOReward Model=HPS-v3+CLIP, Backbone=Flux.1-dev, SDE sampling strategy=SNR-preserving (F^snr_sde), Denoising steps=502026.06 | 32.74 | 14.37 | 39.46 | 1.2318 | 3.6123 | 3.3741 | 23.08 | |
| Flow-GRPOReward Model=HPS-v3, Backbone=Flux.1-dev, Denoising steps=502026.06 | 32.72 | 14.71 | 37.87 | 1.1475 | 3.5821 | 3.3955 | 23.07 | |
| Flow-CPSReward Model=HPS-v3+CLIP, Backbone=Flux.1-dev, Denoising steps=502026.06 | 32.52 | 14.32 | 39.26 | 1.2222 | 3.6339 | 3.3715 | 23.13 | |
| Flow-GRPOReward Model=HPS-v3+CLIP, Backbone=Flux.1-dev, Denoising steps=502026.06 | 32.5 | 14.25 | 39.38 | 1.1988 | 3.626 | 3.3774 | 23.02 | |
| Dance-GRPOReward Model=HPS-v3+CLIP, Backbone=Flux.1-dev, Denoising steps=502026.06 | 31.89 | 13.86 | 39.14 | 1.1406 | 3.5859 | 3.3759 | 22.73 | |
| Flux.1.devReward Model=None, Backbone=Flux.1-dev, Denoising steps=502026.06 | 30.65 | 13.28 | 39.01 | 1.0546 | 3.6006 | 3.381 | 22.68 |