Text-to-Image Generation on Out-of-Domain T2I Dataset
6,707Laplacian VarianceDanceGRPO w/ LoRA Scale
Evaluation Results
| Method | Links | |||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| DanceGRPO w/ LoRA ScaleBase Model=DanceGRPO, Strategy=LoRA Scale2026.02 | 6,707 | 99 | 2 | 2.71 | 2.43 | 0.66 | 1.93 | 0.33 | 1.16 | 0.228 | 5.98 | 0.37 | 3.4 | 3.45 | 2.43 | |
| DanceGRPOBase Model=DanceGRPO2026.02 | 5,348 | 91 | 2.1 | 2.76 | 2.69 | 0.68 | 2.04 | 0.33 | 1.16 | 0.226 | 5.9 | 0.36 | 3.3 | 3.38 | 2.39 | |
| MixGRPOBase Model=MixGRPO2026.02 | 3,900 | 82 | 2.65 | 3.12 | 2.73 | 0.79 | 2.21 | 0.35 | 1.2 | 0.222 | 5.83 | 0.35 | 3.29 | 3.4 | 2.38 | |
| DDPOBase Model=DDPO2026.02 | 3,895 | 68 | 0.78 | 2.48 | 2.52 | 0.61 | 1.87 | 0.32 | 1.13 | 0.227 | 5.9 | 0.39 | 3.39 | 3.57 | 2.43 | |
| DanceGRPO w/ KLBase Model=DanceGRPO, Strategy=KL Divergence2026.02 | 3,699 | 79 | 2.22 | 2.37 | 2.03 | 0.55 | 1.65 | 0.31 | 1.11 | 0.227 | 5.92 | 0.37 | 3.31 | 3.46 | 2.4 | |
| DanceGRPO w/ Early StopBase Model=DanceGRPO, Strategy=Early Stop2026.02 | 3,604 | 77 | 2.29 | 3.16 | 3.17 | 0.83 | 2.39 | 0.32 | 1.09 | 0.225 | 5.92 | 0.36 | 3.29 | 3.38 | 2.38 | |
| OnlineDPOBase Model=OnlineDPO2026.02 | 2,752 | 44 | 0.61 | 2.43 | 2.2 | 0.58 | 1.74 | 0.32 | 1.19 | 0.227 | 5.85 | 0.38 | 3.4 | 3.51 | 2.43 | |
| DanceGRPO w/ PrefGRPOBase Model=DanceGRPO, Strategy=PrefGRPO2026.02 | 2,634 | 71 | 2.24 | 2.59 | 2.2 | 0.73 | 1.84 | 0.33 | 1.17 | 0.227 | 5.95 | 0.37 | 3.34 | 3.49 | 2.43 | |
| DanceGRPO w/ ConsistentRFTBase Model=DanceGRPO, Alignment Strategy=ConsistentRFT2026.02 | 1,421 | 45 | 0.93 | 1.81 | 1.59 | 0.37 | 1.26 | 0.3 | 1.3 | 0.23 | 6.2 | 0.38 | 3.41 | 3.62 | 2.52 | |
| MixGRPO w/ ConsistentRFTBase Model=MixGRPO, Alignment Strategy=ConsistentRFT2026.02 | 1,309 | 45 | 1.17 | 2.43 | 2.08 | 0.54 | 1.68 | 0.31 | 1.32 | 0.228 | 6.05 | 0.37 | 3.33 | 3.43 | 2.46 | |
| OnlineDPO w/ ConsistentRFTBase Model=OnlineDPO, Alignment Strategy=ConsistentRFT2026.02 | 1,131 | 37 | 0.59 | 1.5 | 1.34 | 0.29 | 1.04 | 0.3 | 1.25 | 0.232 | 6.08 | 0.39 | 3.5 | 3.57 | 2.5 | |
| DDPO w/ ConsistentRFTBase Model=DDPO, Alignment Strategy=ConsistentRFT2026.02 | 1,028 | 40 | 0.62 | 1.07 | 0.92 | 0.12 | 0.7 | 0.3 | 1.2 | 0.228 | 6 | 0.4 | 3.48 | 3.59 | 2.48 | |
| Flux DevBase Model=Flux Dev2026.02 | 977 | 26 | 0.27 | 1 | 0.87 | 0.12 | 0.66 | 0.3 | 1.09 | 0.226 | 5.84 | 0.39 | 3.37 | 3.52 | 2.4 |