Text-to-Image Generation on Parti-Prompts
24.34PickScoreDRaFT
Evaluation results
| Method | Links | ||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| DRaFTInference Steps=1, Reward Grad=Yes2026.06 | 24.34 | — | — | — | — | — | — | — | — | 12.66 | 6.485 | — | — | — | |
| VGG-Flow1-stepInference Steps=1, Reward Grad=Yes2026.06 | 23.98 | — | — | — | — | — | — | — | — | 11.4 | 6.2 | — | — | — | |
| DrPOInference Steps=1, Reward Grad=No2026.06 | 23.71 | — | — | — | — | — | — | — | — | 12.6 | 6.665 | — | — | — | |
| Diff.-NPOBase Model=SDXL2026.05 | 23.26 | 82.66 | 84.88 | 45.81 | 86.73 | 85.04 | 77.02 | 29.18 | 35.34 | 1.2505 | 6.2496 | — | — | — | |
| Online-DPOBase Model=SDXL2026.05 | 23.07 | 78.14 | 77.98 | 47.32 | 79.88 | 76.74 | 72.01 | 28.85 | 35.26 | 1.0427 | 6.0486 | — | — | — | |
| DRaFTInference Steps=1, Reward Grad=Yes2026.06 | 23.07 | — | — | — | — | — | — | — | — | 7.72 | 6.516 | — | — | — | |
| VGG-Flow1-stepInference Steps=1, Reward Grad=Yes2026.06 | 22.99 | — | — | — | — | — | — | — | — | 6.5 | 6.027 | — | — | — | |
| DrPOInference Steps=1, Reward Grad=No2026.06 | 22.99 | — | — | — | — | — | — | — | — | 7.46 | 6.284 | — | — | — | |
| SDXL-DPOInference Steps=502026.06 | 22.95 | — | — | — | — | — | — | — | — | 10.66 | 5.811 | — | — | — | |
| RealAlignBase Model=SD-3.5-M, Train Data=512 pairs2026.05 | 22.9 | — | — | — | — | — | — | — | — | 1.27 | 5.73 | 3.99 | 10.66 | 4.2 | |
| PSOInference Steps=1, Reward Grad=No2026.06 | 22.87 | — | — | — | — | — | — | — | — | 9.17 | 5.744 | — | — | — | |
| DPO1-stepInference Steps=1, Reward Grad=No2026.06 | 22.85 | — | — | — | — | — | — | — | — | 11.25 | 6.019 | — | — | — | |
| GRPO1-stepInference Steps=1, Reward Grad=No2026.06 | 22.8 | — | — | — | — | — | — | — | — | 9.27 | 5.71 | — | — | — | |
| SDXL-Turbo (base)Inference Steps=12026.06 | 22.77 | — | — | — | — | — | — | — | — | 9.13 | 5.693 | — | — | — | |
| Diffusion-DPOBase Model=SD-3.5-M, Train Data=851k pairs, re-implementation=true2026.05 | 22.75 | — | — | — | — | — | — | — | — | 1.22 | 5.62 | 3.98 | 9.41 | 3.88 | |
| SDXL baseInference Steps=502026.06 | 22.64 | — | — | — | — | — | — | — | — | 7.24 | 5.761 | — | — | — | |
| SD-3.5-MBase Model=SD-3.5-M, Train Data=/2026.05 | 22.54 | — | — | — | — | — | — | — | — | 1.11 | 5.6 | 3.88 | 8.97 | 4 | |
| Original (SDXL)Base Model=SDXL2026.05 | 22.51 | — | — | — | — | — | — | 28 | 35.23 | 0.6003 | 5.7438 | — | — | — | |
| DPO1-stepInference Steps=1, Reward Grad=No2026.06 | 22.39 | — | — | — | — | — | — | — | — | 5.21 | 5.793 | — | — | — | |
| GRPO1-stepInference Steps=1, Reward Grad=No2026.06 | 22.35 | — | — | — | — | — | — | — | — | 5.65 | 5.779 | — | — | — | |
| SD-Turbo (base)Inference Steps=1, Reward Grad=–2026.06 | 22.29 | — | — | — | — | — | — | — | — | 5.37 | 5.758 | — | — | — | |
| PSOInference Steps=1, Reward Grad=No2026.06 | 22.29 | — | — | — | — | — | — | — | — | 5.42 | 5.763 | — | — | — | |
| Online-SFTBase Model=SDXL2026.05 | 22.21 | 30.21 | 36.02 | 54.42 | 51.38 | 44.13 | 43.23 | 27.68 | 35.86 | 0.6976 | 5.6925 | — | — | — | |
| Diff.-NPOBackbone=SD1.52026.05 | 22.07 | 74.76 | 79.7 | 54.74 | 77.04 | 76.96 | 72.64 | 28.49 | 34.07 | 0.7966 | 5.6231 | — | — | — | |
| SPINBackbone=SD1.52026.05 | 21.98 | 72.73 | 72.34 | 39.01 | 62.34 | 81.92 | 65.67 | 28.15 | 31.98 | 0.4014 | 5.7324 | — | — | — | |
| SeppoBackbone=SD1.52026.05 | 21.97 | 73.97 | 77.59 | 50.44 | 68.58 | 75.56 | 69.23 | 28.24 | 33.43 | 0.5143 | 5.5904 | — | — | — | |
| Online-DPOBackbone=SD1.52026.05 | 21.95 | 71.43 | 78.42 | 56.38 | 74.73 | 74.73 | 71.14 | 28.31 | 34.09 | 0.6699 | 5.5266 | — | — | — | |
| SD2.1Inference Steps=50, Reward Grad=–2026.06 | 21.77 | — | — | — | — | — | — | — | — | 3.97 | 5.547 | — | — | — | |
| RealAlignBase Model=SD-1.5, Train Data=512 pairs2026.05 | 21.64 | — | — | — | — | — | — | — | — | 0.41 | 5.49 | 3.28 | 7.06 | 3.96 | |
| Diffusion-DPOBase Model=SD-1.5, Train Data=851k pairs2026.05 | 21.58 | — | — | — | — | — | — | — | — | 0.4 | 5.47 | 3.25 | 6.48 | 3.78 | |
| OriginalBackbone=SD1.52026.05 | 21.53 | — | — | — | — | — | — | 27.44 | 33.15 | 0.1926 | 5.3592 | — | — | — | |
| SD1.5Inference Steps=50, Reward Grad=–2026.06 | 21.49 | — | — | — | — | — | — | — | — | 2.25 | 5.358 | — | — | — | |
| Online-SFTBackbone=SD1.52026.05 | 21.35 | 38.88 | 45.24 | 55.1 | 54.18 | 59.83 | 50.65 | 27.37 | 33.84 | 0.3366 | 5.4508 | — | — | — | |
| SD-1.5Base Model=SD-1.5, Train Data=/2026.05 | 21.34 | — | — | — | — | — | — | — | — | 0.25 | 5.39 | 3.15 | 5.69 | 3.7 | |
| LCM-SD1.5Inference Steps=4, Reward Grad=–2026.06 | 21.15 | — | — | — | — | — | — | — | — | -1.94 | 5.396 | — | — | — |