Conditional Generation on MNIST (Acc, FID, sFID, IS)
0.004FIDMoEMA
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| MoEMACondition representation=CLIP text encoder2026.07 | 0.004 | — | — | — | 0.053 | — | |
| SimpleAvg2026.07 | 0.011 | — | — | — | 0.471 | 60.1 | |
| StaticMA2026.07 | 0.016 | — | — | — | 0.783 | 71.1 | |
| BestSingle2026.07 | 0.018 | — | — | — | 1.001 | 75.3 | |
| TJS-0.8NFE=25, gamma=0.82026.07 | 2.65 | — | — | — | — | — | |
| FullNFE=302026.07 | 3.2 | — | — | — | — | — | |
| FMwCinference cost ratio=1.003×2026.05 | 4.22 | — | — | — | — | — | |
| MC-Dropoutk=5, inference cost ratio=5×2026.05 | 4.25 | — | — | — | — | — | |
| Ensemblek=5, inference cost ratio=5×2026.05 | 4.35 | — | — | — | — | — | |
| TJS-0.7NFE=22, gamma=0.72026.07 | 5.15 | — | — | — | — | — | |
| FMinference cost ratio=1×2026.05 | 5.17 | — | — | — | — | — | |
| CFMBackbone=LightningDiT [46], Auxiliary Path=None2026.05 | 5.9 | 95.5 | 0.006 | 2.099 | — | — | |
| CFM + numeric ηBackbone=LightningDiT [46], Auxiliary Variable (η)=numeric2026.05 | 7.33 | 98.1 | 0.007 | 2.096 | — | — | |
| CFM + learned ηBackbone=LightningDiT [46], Auxiliary Variable (η)=learned2026.05 | 7.57 | 97.1 | 0.008 | 2.112 | — | — | |
| TJS-0.6NFE=19, gamma=0.62026.07 | 8.51 | — | — | — | — | — | |
| numeric ηBackbone=LightningDiT [46], Auxiliary Variable (η)=numeric, Guidance=auxiliary only2026.05 | 10.04 | 94.3 | 0.009 | 2.097 | — | — | |
| TJS-0.5NFE=16, gamma=0.52026.07 | 12.22 | — | — | — | — | — | |
| learned ηBackbone=LightningDiT [46], Auxiliary Variable (η)=learned, Guidance=auxiliary only2026.05 | 13.82 | 90.6 | 0.015 | 2.128 | — | — | |
| TJS-0.3NFE=10, gamma=0.32026.07 | 27.36 | — | — | — | — | — |