Text-to-Image Generation on DrawBench (HPSv2, ImageReward, PickScore, Aesthetics, Vendi Score, CLIP Score)
28.71HPSv2Oracle Noise
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| Oracle NoiseBase Model=SDXL-Turbo, Time=2s2026.04 | 28.71 | 0.76 | 19.98 | 5.64 | 17.72 | 69.57 | |
| ReNOBase Model=SDXL-Turbo, Time=30s2026.04 | 28.55 | 0.71 | 19.95 | 5.55 | 17.4 | 69.4 | |
| GaussianBase Model=SDXL-Turbo2026.04 | 28.41 | 0.67 | 19.99 | 5.49 | 17.11 | 69.25 | |
| Oracle NoiseModel=SD3.5-M, Time (s)=2.02026.04 | 25.97 | -0.4 | 19.42 | 5.21 | 14.6 | 59.18 | |
| Stable NoiseModel=SD3.5-M, Time (s)=> 6002026.04 | 25.82 | -0.48 | 19.39 | 5.23 | 13.2 | 58.85 | |
| InitNOModel=SD3.5-M, Time (s)=35.02026.04 | 25.8 | -0.5 | 19.39 | 5.2 | 13 | 58.8 | |
| GaussianModel=SD3.5-M2026.04 | 25.76 | -0.55 | 19.39 | 5.19 | 12.5 | 58.76 | |
| Oracle NoiseModel=SDXL, Time (s)=2.02026.04 | 25.69 | -0.68 | 19.44 | 5.48 | 13.02 | 54.23 | |
| Stable NoiseModel=SDXL, Time (s)=> 6002026.04 | 25.35 | -0.85 | 19.32 | 5.59 | 11.2 | 52 | |
| InitNOModel=SDXL, Time (s)=35.02026.04 | 25.1 | -0.9 | 19.3 | 5.55 | 10.5 | 51.8 | |
| GaussianModel=SDXL2026.04 | 24.67 | -1.16 | 19.26 | 5.56 | 9.76 | 49.68 |