Generative Modeling on ImageNet (val)
2.31FDvalDRL (λ*)
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| DRL (λ*)Model=RAE, Embedder=Incep., λ=1, R1=10−5, cfg=2.25, interval=[.1, 1]2026.06 | 2.31 | 0.38 | 71.4 | 65.2 | 1.042 | 95.2 | |
| BaseModel=RAE, Embedder=Incep., λ=−−, R1=−−, cfg=2, interval=[.1, 1]2026.06 | 2.32 | 0.35 | 71.4 | 65.1 | 1.043 | 95.3 | |
| DRL (λ=1, R1=0)Model=RAE, Embedder=Incep., λ=1, R1=0, cfg=2.5, interval=[.1, 1]2026.06 | 2.43 | 0.49 | 71.6 | 64.5 | 1.052 | 95.3 | |
| BaseModel=REPA, Embedder=Incep., λ=−−, R1=−−, cfg=2.25, interval=[.3, 1]2026.06 | 2.77 | 0.71 | 70.2 | 64.2 | 1.004 | 93.9 | |
| BaseModel=SiT, Embedder=Incep., λ=−−, R1=−−, cfg=2, interval=[.3, 1]2026.06 | 2.9 | 0.7 | 71.3 | 62.6 | 1.065 | 94.1 | |
| DRL (λ=1, R1=0)Model=JiT, Embedder=Incep., λ=1, R1=0, cfg=2.5, interval=[.3, 1]2026.06 | 3.05 | 0.74 | 69.8 | 62.7 | 0.976 | 93.2 | |
| DRL (λ*)Model=JiT, Embedder=Incep., λ=1, R1=0, cfg=2.5, interval=[.3, 1]2026.06 | 3.05 | 0.74 | 69.8 | 62.7 | 0.976 | 93.2 | |
| DRL (λ*)Model=REPA, Embedder=Incep., λ=1, R1=10−5, cfg=1.5, interval=[.3, 1]2026.06 | 3.05 | 0.94 | 71.8 | 62.1 | 1.067 | 94.5 | |
| BaseModel=JiT, Embedder=Incep., λ=−−, R1=−−, cfg=2.25, interval=[.1, 1]2026.06 | 3.16 | 0.74 | 68.4 | 62.9 | 0.95 | 92.8 | |
| DRL (λ*)Model=SiT, Embedder=Incep., λ=1, R1=10−5, cfg=1.5, interval=[.3, 1]2026.06 | 3.26 | 1.08 | 73.4 | 61 | 1.136 | 94.6 | |
| DRL (λ=1, R1=0)Model=REPA, Embedder=Incep., λ=1, R1=0, cfg=1, interval=[0, 1]2026.06 | 3.66 | 1.16 | 71.9 | 61.1 | 1.072 | 94.3 | |
| DRL (λ=1, R1=0)Model=SiT, Embedder=Incep., λ=1, R1=0, cfg=1.25, interval=[.3, 1]2026.06 | 4.17 | 1.77 | 74.2 | 58.2 | 1.169 | 94.7 | |
| DRL (λ*)Model=RAE, Embedder=DINOv3-L, λ=5, R1=10−5, cfg=1.25, interval=[0, 1]2026.06 | 5.97 | 2.71 | 80 | 65.1 | 0.954 | 89.5 | |
| DRL (λ=1, R1=0)Model=RAE, Embedder=DINOv3-L, λ=1, R1=0, cfg=1.5, interval=[0, 1]2026.06 | 6.44 | 2.91 | 82 | 60.5 | 1.013 | 89.5 | |
| BaseModel=RAE, Embedder=DINOv3-L, λ=−−, R1=−−, cfg=1.5, interval=[0, 1]2026.06 | 6.66 | 3.37 | 81 | 60.8 | 0.992 | 89.5 | |
| DRL (λ*)Model=JiT, Embedder=DINOv3-L, λ=20, R1=10−5, cfg=3.25, interval=[.1, 1]2026.06 | 10.14 | 5.94 | 81.8 | 45.6 | 1.029 | 86.2 | |
| DRL (λ*)Model=RAE, Embedder=SigLIP-L, λ=10, R1=10−5, cfg=1.75, interval=[.1, 1]2026.06 | 11.03 | 19.18 | 83.9 | 52.4 | 1.185 | 91.7 | |
| BaseModel=REPA, Embedder=SigLIP-L, λ=−−, R1=−−, cfg=3.75, interval=[.3, 1]2026.06 | 11.18 | 20.65 | 83 | 48.8 | 1.141 | 91 | |
| DRL (λ*)Model=REPA, Embedder=SigLIP-L, λ=1, R1=10−5, cfg=3, interval=[.3, 1]2026.06 | 11.34 | 20.71 | 83.3 | 48.2 | 1.165 | 90.7 | |
| DRL (λ=1, R1=0)Model=RAE, Embedder=SigLIP-L, λ=1, R1=0, cfg=1.25, interval=[0, 1]2026.06 | 11.77 | 23.36 | 86 | 48.8 | 1.28 | 92.7 | |
| BaseModel=RAE, Embedder=SigLIP-L, λ=−−, R1=−−, cfg=1.25, interval=[0, 1]2026.06 | 11.81 | 22.86 | 85.5 | 49.4 | 1.252 | 92.5 | |
| DRL (λ=1, R1=0)Model=JiT, Embedder=DINOv3-L, λ=1, R1=0, cfg=2.25, interval=[0, 1]2026.06 | 12.43 | 7.74 | 81.6 | 39.7 | 1.039 | 84.7 | |
| DRL (λ=1, R1=0)Model=REPA, Embedder=SigLIP-L, λ=1, R1=0, cfg=2.25, interval=[.3, 1]2026.06 | 12.77 | 23.03 | 83.2 | 47.5 | 1.172 | 89.9 | |
| BaseModel=JiT, Embedder=DINOv3-L, λ=−−, R1=−−, cfg=2.75, interval=[0, 1]2026.06 | 13.16 | 8.56 | 81.3 | 37.6 | 1.038 | 84.1 | |
| BaseModel=SiT, Embedder=SigLIP-L, λ=−−, R1=−−, cfg=3.5, interval=[.3, 1]2026.06 | 13.48 | 25.3 | 84.2 | 40 | 1.241 | 89.8 | |
| DRL (λ*)Model=SiT, Embedder=SigLIP-L, λ=1, R1=10−5, cfg=2.75, interval=[.3, 1]2026.06 | 13.64 | 24.76 | 84.5 | 40.1 | 1.253 | 89.6 | |
| DRL (λ=1, R1=0)Model=SiT, Embedder=SigLIP-L, λ=1, R1=0, cfg=2.25, interval=[.3, 1]2026.06 | 14.95 | 27.46 | 84.2 | 40.5 | 1.238 | 88.7 | |
| DRL (λ*)Model=REPA, Embedder=DINOv3-L, λ=10, R1=10−5, cfg=2.25, interval=[.3, 1]2026.06 | 15 | 9.02 | 85.2 | 35.4 | 1.243 | 86.2 | |
| DRL (λ=1, R1=0)Model=REPA, Embedder=DINOv3-L, λ=1, R1=0, cfg=2.75, interval=[.1, 1]2026.06 | 16.8 | 11.71 | 84.1 | 32.4 | 1.235 | 86.6 | |
| DRL (λ*)Model=JiT, Embedder=SigLIP-L, λ=5, R1=10−5, cfg=3, interval=[.1, 1]2026.06 | 18.43 | 42.03 | 85.4 | 33.9 | 1.253 | 89.9 | |
| DRL (λ*)Model=SiT, Embedder=DINOv3-L, λ=10, R1=10−5, cfg=2.25, interval=[.3, 1]2026.06 | 18.54 | 10.99 | 85.1 | 23.4 | 1.354 | 83.1 | |
| DRL (λ=1, R1=0)Model=JiT, Embedder=SigLIP-L, λ=1, R1=0, cfg=3, interval=[.1, 1]2026.06 | 18.94 | 42.95 | 84.6 | 34.8 | 1.206 | 89.6 | |
| BaseModel=JiT, Embedder=SigLIP-L, λ=−−, R1=−−, cfg=3.5, interval=[.1, 1]2026.06 | 19.93 | 46.8 | 83.7 | 34.8 | 1.168 | 88.8 | |
| DRL (λ=1, R1=0)Model=SiT, Embedder=DINOv3-L, λ=1, R1=0, cfg=3, interval=[.1, 1]2026.06 | 20.61 | 15.16 | 83.7 | 22 | 1.303 | 84 | |
| BaseModel=REPA, Embedder=DINOv3-L, λ=−−, R1=−−, cfg=3.25, interval=[0, 1]2026.06 | 20.97 | 14.02 | 83.1 | 23.8 | 1.233 | 81.8 | |
| BaseModel=SiT, Embedder=DINOv3-L, λ=−−, R1=−−, cfg=3.5, interval=[0, 1]2026.06 | 24.1 | 16.7 | 83.5 | 15.6 | 1.337 | 80 | |
| DRL (λ*)Model=RAE, Embedder=DINOv2-L, λ=5, R1=10−5, cfg=1.25, interval=[0, 1]2026.06 | 34.68 | 10.51 | 76.2 | 71.9 | 0.915 | 89.7 | |
| DRL (λ=1, R1=0)Model=RAE, Embedder=DINOv2-L, λ=1, R1=0, cfg=1.25, interval=[0, 1]2026.06 | 37.74 | 15.22 | 74.3 | 72.3 | 0.87 | 88.4 | |
| BaseModel=RAE, Embedder=DINOv2-L, λ=−−, R1=−−, cfg=1.5, interval=[0, 1]2026.06 | 40.78 | 16.59 | 76.4 | 67.7 | 0.924 | 88.7 | |
| DRL (λ*)Model=JiT, Embedder=DINOv2-L, λ=10, R1=10−5, cfg=2.75, interval=[.1, 1]2026.06 | 45.59 | 15.39 | 78.8 | 57.4 | 1.054 | 88.4 | |
| DRL (λ=1, R1=0)Model=REPA, Embedder=DINOv2-L, λ=1, R1=0, cfg=2.75, interval=[.3, 1]2026.06 | 49.11 | 20.44 | 79.7 | 56.7 | 1.058 | 88.2 | |
| DRL (λ*)Model=REPA, Embedder=DINOv2-L, λ=1, R1=0, cfg=2.75, interval=[.3, 1]2026.06 | 49.11 | 20.44 | 79.7 | 56.7 | 1.058 | 88.2 | |
| DRL (λ=1, R1=0)Model=JiT, Embedder=DINOv2-L, λ=1, R1=0, cfg=3.5, interval=[.1, 1]2026.06 | 49.96 | 24.87 | 75.3 | 58.5 | 0.943 | 86.7 | |
| BaseModel=REPA, Embedder=DINOv2-L, λ=−−, R1=−−, cfg=2.75, interval=[.1, 1]2026.06 | 52.75 | 22.95 | 80.9 | 51.1 | 1.111 | 88.4 | |
| BaseModel=JiT, Embedder=DINOv2-L, λ=−−, R1=−−, cfg=4, interval=[.1, 1]2026.06 | 54.96 | 31.69 | 73.7 | 57.6 | 0.896 | 85.1 | |
| DRL (λ=1, R1=0)Model=SiT, Embedder=DINOv2-L, λ=1, R1=0, cfg=2.75, interval=[.3, 1]2026.06 | 58.94 | 23.46 | 79.6 | 46.5 | 1.148 | 86 | |
| DRL (λ*)Model=SiT, Embedder=DINOv2-L, λ=1, R1=0, cfg=2.75, interval=[.3, 1]2026.06 | 58.94 | 23.46 | 79.6 | 46.5 | 1.148 | 86 | |
| BaseModel=SiT, Embedder=DINOv2-L, λ=−−, R1=−−, cfg=2.75, interval=[.1, 1]2026.06 | 62.68 | 28.17 | 80.1 | 43.6 | 1.171 | 86.2 |