Text-to-Image Generation on DrawBench (HPS, IR, Pick, Aes., CLIP)
28.49HPSGolden Noise
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| Golden NoiseBackbone=SDXL2026.04 | 28.49 | 0.58 | 19.98 | 5.44 | 70.55 | |
| Oracle NoiseBackbone=SDXL2026.04 | 28.25 | 0.57 | 20.02 | 5.48 | 70.21 | |
| GaussianBackbone=SDXL2026.04 | 28.07 | 0.51 | 19.91 | 5.43 | 69.79 | |
| EAMOptimized Reward=PickScore + HPSv2.1 + Aesthetics2026.05 | 0.306 | 0.98 | 22.93 | 6.56 | 0.261 | |
| AMOptimized Reward=PickScore + HPSv2.1 + Aesthetics2026.05 | 0.303 | 0.83 | 23.06 | 6.21 | 0.253 | |
| EAMOptimized Reward=PickScore2026.05 | 0.297 | 0.95 | 23.16 | 6.27 | 0.252 | |
| AMOptimized Reward=PickScore2026.05 | 0.292 | 0.97 | 23 | 6.25 | 0.259 | |
| SD3.5-M + CFGBase Model=SD3.5-M, Classifier-Free Guidance=true2026.05 | 0.265 | 0.65 | 21.82 | 5.41 | 0.275 | |
| SD3.5-MBase Model=SD3.5-M2026.05 | 0.214 | -0.58 | 20.34 | 5.29 | 0.226 |