Class-conditioned Image Generation on ImageNet 1K (FID50K)
9.85FIDPAFM
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| PAFMModel=SiT-XL/2, K=16, NFE=50, CFG=None, Training iterations=400K, Loss=REPA2026.05 | 9.85 | 5.24 | 68 | 65 | |
| FMModel=SiT-XL/2, K=–, NFE=50, CFG=None, Training iterations=400K, Loss=REPA2026.05 | 11.14 | 8.25 | 67 | 66 | |
| PAFMModel=SiT-B/2, K=16, NFE=50, CFG=None, Training iterations=400K, Loss=REPA2026.05 | 24.88 | 6.81 | 58 | 64 | |
| PAFMModel=SiT-B/2, K=8, NFE=50, CFG=None, Training iterations=400K, Loss=REPA2026.05 | 25.25 | 6.9 | 58 | 65 | |
| PAFMModel=SiT-B/2, K=64, NFE=50, CFG=None, Training iterations=400K, Loss=REPA2026.05 | 25.45 | 6.91 | 58 | 64 | |
| PAFMModel=SiT-B/2, K=4, NFE=50, CFG=None, Training iterations=400K, Loss=REPA2026.05 | 25.47 | 6.97 | 58 | 64 | |
| FMModel=SiT-B/2, K=–, NFE=50, CFG=None, Training iterations=400K, Loss=REPA2026.05 | 27.57 | 11.44 | 57 | 63 |