DINOv2 REPA baseline
PixelDiT-XL on ImageNet-256. Online patch block 8 is projected and aligned patch-wise to frozen DINOv2 clean-image tokens with cosine distance.
50k-sample ADM metrics
| Step | FID ↓ | sFID ↓ | IS ↑ |
|---|---|---|---|
| 10,000 | 389.2824 | 414.7020 | 1.0710 |
| 20,000 | 363.6215 | 304.7034 | 1.4038 |
| 30,000 | 106.0903 | 78.1710 | 11.7050 |
| 40,000 | 66.7541 | 45.9615 | 17.5108 |
| 50,000 | 31.1164 | 20.3315 | 26.6462 |
| 60,000 | 21.3407 | 14.7369 | 32.9395 |
Raw evaluator outputs are stored under metrics/.