Image Classification on ImageNet1k
88.4AccuracyDINOv3
Evaluation Results
| Method | Links | |
|---|---|---|
| DINOv3Param.=7B2026.03 | 88.4 | |
| SigLIP2Param.=1.2B, Resolution=256x2562026.03 | 88 | |
| DINOv3 ViT-H+Param.=0.8B2026.03 | 87.9 | |
| PEcore GParam.=1.9B, Resolution=256x2562026.03 | 87.6 | |
| DINOv2Param.=1.1B, Resolution=256x2562026.03 | 86.1 | |
| InternVideo2s2-1BParam.=1B, Resolution=256x2562026.03 | 85.8 | |
| V-JEPA 2.1 ViT-GParam.=2B, Resolution=384x3842026.03 | 85.5 | |
| V-JEPA 2 ViT-gParam.=1B, Resolution=384x3842026.03 | 85.1 | |
| V-JEPA 2.1 ViT-gParam.=1B, Resolution=384x3842026.03 | 84.8 | |
| FP32Backbone=DINOv2-B/14, Precision=FP322025.10 | 83.26 | |
| PTQ4ViTBackbone=DINOv2-B/14, Bit-width=W8A82025.10 | 82.97 | |
| PTQ4ViT + RegCacheBackbone=DINOv2-B/14, Bit-width=W8A82025.10 | 82.92 | |
| ERQBackbone=DINOv2-B/14, Bit-width=W8A82025.10 | 82.59 | |
| ERQ + RegCacheBackbone=DINOv2-B/14, Bit-width=W8A82025.10 | 82.54 | |
| RepQ-ViTBackbone=DINOv2-B/14, Bit-width=W8A82025.10 | 82.27 | |
| RepQ-ViT + RegCacheBackbone=DINOv2-B/14, Bit-width=W8A82025.10 | 81.55 | |
| V-JEPA ViT-HParam.=600M, Resolution=256x2562026.03 | 80 | |
| PTQ4ViT + RegCacheBackbone=DINOv2-B/14, Bit-width=W6A62025.10 | 78.55 | |
| FIMA-QBackbone=SigLIP2-B/16, Bit-width=W4A42025.10 | 78.48 | |
| FIMA-QBackbone=SigLIP2-B/16, Bit-width=W8A82025.10 | 78.48 | |
| PTQ4ViTBackbone=DINOv2-B/14, Bit-width=W6A62025.10 | 78.48 | |
| FP32Backbone=SigLIP2-B/16, Precision=FP322025.10 | 78.47 | |
| FIMA-QBackbone=SigLIP2-B/16, Bit-width=W6A62025.10 | 78.47 | |
| ERQ + RegCacheBackbone=SigLIP2-B/16, Bit-width=W8A82025.10 | 78.45 | |
| FIMA-Q + RegCacheBackbone=SigLIP2-B/16, Bit-width=W4A42025.10 | 78.38 | |
| FIMA-Q + RegCacheBackbone=SigLIP2-B/16, Bit-width=W6A62025.10 | 78.37 | |
| FIMA-Q + RegCacheBackbone=SigLIP2-B/16, Bit-width=W8A82025.10 | 78.37 | |
| ERQBackbone=SigLIP2-B/16, Bit-width=W8A82025.10 | 78.05 | |
| ERQBackbone=DINOv2-B/14, Bit-width=W6A62025.10 | 78.03 | |
| ERQ + RegCacheBackbone=DINOv2-B/14, Bit-width=W6A62025.10 | 77.88 | |
| CapPa L/14LiT pairs=12B, Zero-shot=true, Frozen transfer=true2023.06 | 77.5 | |
| RepQ-ViT + RegCacheBackbone=SigLIP2-B/16, Bit-width=W8A82025.10 | 77.13 | |
| PTQ4ViTBackbone=SigLIP2-B/16, Bit-width=W8A82025.10 | 76.92 | |
| PTQ4ViT + RegCacheBackbone=SigLIP2-B/16, Bit-width=W8A82025.10 | 76.88 | |
| CLIP* L/14LiT pairs=12B, Zero-shot=true, Frozen transfer=true2023.06 | 76.6 | |
| NoisyQuant + RegCacheBackbone=SigLIP2-B/16, Bit-width=W8A82025.10 | 76.44 | |
| RepQ-ViTBackbone=SigLIP2-B/16, Bit-width=W8A82025.10 | 76.43 | |
| CapPa L/14LiT pairs=3B, Zero-shot=true, Frozen transfer=true2023.06 | 76.4 | |
| FIMA-Q + RegCacheBackbone=SigLIP-B/16, Bit-width=W4A42025.10 | 76.13 | |
| FIMA-Q + RegCacheBackbone=SigLIP-B/16, Bit-width=W6A62025.10 | 76.12 | |
| FIMA-Q + RegCacheBackbone=SigLIP-B/16, Bit-width=W8A82025.10 | 76.12 | |
| FIMA-QBackbone=SigLIP-B/16, Bit-width=W6A62025.10 | 76.06 | |
| FIMA-QBackbone=SigLIP-B/16, Bit-width=W8A82025.10 | 76.06 | |
| FP32Backbone=SigLIP-B/16, Precision=FP322025.10 | 76.05 | |
| FIMA-QBackbone=SigLIP-B/16, Bit-width=W4A42025.10 | 76.05 | |
| ERQ + RegCacheBackbone=SigLIP-B/16, Bit-width=W8A82025.10 | 75.98 | |
| ERQBackbone=SigLIP-B/16, Bit-width=W8A82025.10 | 75.95 | |
| RepQ-ViT + RegCacheBackbone=SigLIP-B/16, Bit-width=W8A82025.10 | 75.94 | |
| CLIP* L/14LiT pairs=3B, Zero-shot=true, Frozen transfer=true2023.06 | 75.8 | |
| PTQ4ViT + RegCacheBackbone=SigLIP-B/16, Bit-width=W8A82025.10 | 75.75 | |
| NoisyQuant + RegCacheBackbone=SigLIP-B/16, Bit-width=W8A82025.10 | 75.64 | |
| PTQ4ViTBackbone=SigLIP-B/16, Bit-width=W8A82025.10 | 75.57 | |
| NoisyQuantBackbone=SigLIP-B/16, Bit-width=W8A82025.10 | 75.5 | |
| RepQ-ViT + RegCacheBackbone=SigLIP2-B/16, Bit-width=W6A62025.10 | 75.4 | |
| ERQ + RegCacheBackbone=SigLIP2-B/16, Bit-width=W6A62025.10 | 75.32 | |
| ERQ + RegCacheBackbone=SigLIP-B/16, Bit-width=W6A62025.10 | 75.25 | |
| RepQ-ViTBackbone=SigLIP-B/16, Bit-width=W8A82025.10 | 75.23 | |
| CLIP L/14LiT pairs=3B, Zero-shot=true, Frozen transfer=true2023.06 | 75.1 | |
| ERQBackbone=SigLIP2-B/16, Bit-width=W6A62025.10 | 74.75 | |
| RepQ-ViT + RegCacheBackbone=SigLIP-B/16, Bit-width=W6A62025.10 | 74.52 | |
| ERQBackbone=SigLIP-B/16, Bit-width=W6A62025.10 | 74.44 | |
| Naïve + RegCacheBackbone=SigLIP-B/16, Bit-width=W8A82025.10 | 74.38 | |
| RepQ-ViTBackbone=SigLIP-B/16, Bit-width=W6A62025.10 | 73.32 | |
| PTQ4ViT + RegCacheBackbone=SigLIP-B/16, Bit-width=W6A62025.10 | 72.38 | |
| Naïve + RegCacheBackbone=SigLIP2-B/16, Bit-width=W8A82025.10 | 72.35 | |
| NoisyQuant + RegCacheBackbone=SigLIP-B/16, Bit-width=W6A62025.10 | 72.28 | |
| NoisyQuantBackbone=DINOv2-B/14, Bit-width=W8A82025.10 | 71.46 | |
| VideoMAEv2Param.=1B2026.03 | 71.4 | |
| NoisyQuantBackbone=SigLIP-B/16, Bit-width=W6A62025.10 | 71.1 | |
| NoisyQuantBackbone=SigLIP2-B/16, Bit-width=W8A82025.10 | 70.83 | |
| NoisyQuant + RegCacheBackbone=DINOv2-B/14, Bit-width=W8A82025.10 | 70.38 | |
| FP32Backbone=OpenCLIP-B/16, Precision=FP322025.10 | 70.22 | |
| CapPaLiT pairs=12B, Zero-shot=true, Frozen transfer=true2023.06 | 70.2 | |
| ERQ + RegCacheBackbone=OpenCLIP-B/16, Bit-width=W8A82025.10 | 70.12 | |
| RepQ-ViT + RegCacheBackbone=OpenCLIP-B/16, Bit-width=W8A82025.10 | 70.06 | |
| CLIP*LiT pairs=12B, Zero-shot=true, Frozen transfer=true2023.06 | 70 | |
| NaïveBackbone=SigLIP-B/16, Bit-width=W8A82025.10 | 69.71 | |
| PTQ4ViT + RegCacheBackbone=OpenCLIP-B/16, Bit-width=W8A82025.10 | 69.6 | |
| PTQ4ViTBackbone=OpenCLIP-B/16, Bit-width=W8A82025.10 | 69.39 | |
| NoisyQuant + RegCacheBackbone=OpenCLIP-B/16, Bit-width=W8A82025.10 | 69.36 | |
| ERQBackbone=OpenCLIP-B/16, Bit-width=W8A82025.10 | 69.25 | |
| FIMA-Q + RegCacheBackbone=OpenCLIP-B/16, Bit-width=W8A82025.10 | 69.13 | |
| ERQ + RegCacheBackbone=OpenCLIP-B/16, Bit-width=W6A62025.10 | 69.09 | |
| PTQ4ViT + RegCacheBackbone=SigLIP2-B/16, Bit-width=W6A62025.10 | 69.02 | |
| CapLiT pairs=12B, Zero-shot=true, Frozen transfer=true2023.06 | 69 | |
| CLIP*LiT pairs=3B, Zero-shot=true, Frozen transfer=true2023.06 | 69 | |
| CapPaLiT pairs=3B, Zero-shot=true, Frozen transfer=true2023.06 | 68.8 | |
| RepQ-ViTBackbone=OpenCLIP-B/16, Bit-width=W8A82025.10 | 68.7 | |
| PTQ4ViTBackbone=SigLIP-B/16, Bit-width=W6A62025.10 | 68.68 | |
| FIMA-Q + RegCacheBackbone=OpenCLIP-B/16, Bit-width=W6A62025.10 | 68.62 | |
| FIMA-QBackbone=OpenCLIP-B/16, Bit-width=W8A82025.10 | 68.53 | |
| FP32Backbone=CLIP-B/16, Precision=FP322025.10 | 68.32 | |
| CLIPLiT pairs=3B, Zero-shot=true, Frozen transfer=true2023.06 | 68.3 | |
| RepQ-ViT + RegCacheBackbone=OpenCLIP-B/16, Bit-width=W6A62025.10 | 68.22 | |
| RepQ-ViT + RegCacheBackbone=CLIP-B/16, Bit-width=W8A82025.10 | 68.1 | |
| ERQ + RegCacheBackbone=CLIP-B/16, Bit-width=W8A82025.10 | 68.09 | |
| ERQBackbone=CLIP-B/16, Bit-width=W8A82025.10 | 67.99 | |
| PTQ4ViT + RegCacheBackbone=CLIP-B/16, Bit-width=W8A82025.10 | 67.86 | |
| FIMA-Q + RegCacheBackbone=CLIP-B/16, Bit-width=W8A82025.10 | 67.86 | |
| CapLiT pairs=3B, Zero-shot=true, Frozen transfer=true2023.06 | 67.8 |