Text-to-Image Generation on Image Quality & Realism Evaluation Set
0.905Reward (Baseline)NormGuard
Evaluation results
| Method | Links | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| NormGuardReward Model=PickScore, Model=FLUX.2-4B, Fine-tuning Method=NFT2026.06 | 0.905 | 0.91 | 38 | 47 | 15 | 46 | 51 | 3 | 0.239 | 0.274 | |
| NormGuardReward Model=PickScore, Model=FLUX.2-4B, Fine-tuning Method=AWM2026.06 | 0.886 | 0.886 | 41 | 59 | 0 | 38 | 55 | 7 | 0.277 | 0.32 | |
| NormGuardReward Model=PickScore, Model=SD3.5-M, Fine-tuning Method=NFT2026.06 | 0.869 | 0.866 | 35 | 64 | 1 | 30 | 67 | 2 | 0.248 | 0.31 | |
| NormGuardReward Model=PickScore, Model=SD3.5-M, Fine-tuning Method=AWM2026.06 | 0.869 | 0.88 | 28 | 72 | 0 | 23 | 73 | 4 | 0.283 | 0.269 | |
| NormGuardReward Model=PickScore, Model=SD3.5-M, Fine-tuning Method=DPO2026.06 | 0.849 | 0.851 | 28 | 68 | 4 | 20 | 47 | 33 | 0.468 | 0.472 | |
| NormGuardReward Model=HPSv2, Model=FLUX.2-4B, Fine-tuning Method=NFT2026.06 | 0.315 | 0.311 | 36 | 63 | 1 | 36 | 60 | 4 | 0.24 | 0.274 | |
| NormGuardReward Model=HPSv2, Model=SD3.5-M, Fine-tuning Method=NFT2026.06 | 0.308 | 0.309 | 42 | 58 | 0 | 47 | 52 | 1 | 0.218 | 0.299 |