Text-to-image alignment on Pick-a-Pic V1 (test)
23.14PickScoreTraining-free alignment for flow matching
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| Training-free alignment for flow matchingType=Finetuning-free, Base Model=SD3.5 Large Turbo2026.01 | 23.14 | 32.31 | 110.25 | 6.528 | |
| Finetuning-free alignment for diffusion modelsType=Finetuning-free, Base Model=SDXL, Training GPU Hour=922026.01 | 23.08 | 32.12 | 106.25 | 6.452 | |
| SPOType=Finetuning-based, Base Model=SDXL, Training GPU Hour=2342026.01 | 23.06 | 31.8 | 108.03 | 6.364 | |
| Diff.-DPOType=Finetuning-based, Base Model=SDXL, Training GPU Hour=48002026.01 | 22.64 | 29.31 | 94.36 | 6.015 | |
| Tweedie's formulaType=Training-free, Base Model=SDXL2026.01 | 22.34 | 28.76 | 95.01 | 6.002 | |
| SD3.5 Large TurboType=Baseline, Base Model=SD3.5 Large Turbo2026.01 | 22.3 | 30.29 | 101.59 | 6.519 | |
| SDXLType=Baseline, Base Model=SDXL2026.01 | 21.95 | 26.95 | 53.8 | 5.95 | |
| Direct backpropagateType=Training-free, Base Model=SDXL2026.01 | 21.84 | 27.53 | 58.7 | 5.922 |