Image Reconstruction on COCO 512p
0.38FIDViTok-v2 5B
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| ViTok-v2 5BComp.=f16×64, r=12, high DINO loss weight (λ=1000)=true2026.05 | 0.38 | 0.72 | 32.1 | 92.5 | |
| ViTok-v2 5BComp.=f16×64, r=122026.05 | 0.41 | 1.02 | 35.9 | 93.8 | |
| FLUX.2Comp.=f8×16, r=122026.05 | 0.45 | 0.45 | 33.2 | 91.4 | |
| QwenComp.=f8×16, r=122026.05 | 0.79 | 2.24 | 30.8 | 86.3 | |
| ViTok-v2 5BComp.=f16×32, r=242026.05 | 1.02 | 1.44 | 32.9 | 88.8 | |
| SDXL-VAEComp.=f8×8, r=242026.05 | 2.2 | 4.77 | 27.5 | 76.2 | |
| ViTok-v2 5BComp.=f16×16, r=482026.05 | 2.27 | 2.24 | 30.1 | 82 | |
| SD-VAEComp.=f8×4, r=482026.05 | 2.28 | 4.78 | 27.1 | 74.2 |