Image Classification on Downstream Datasets Average (Robustness)
85.4Average AccuracyWAVE
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| WAVEMethod Category=Learngene, Backbone=DeiT-S, 11.3M, Para. (M)=4.02024.06 | 85.4 | — | — | — | — | |
| Direct FTMethod Category=PT, Backbone=DeiT-S, 11.3M, Para. (M)=11.02024.06 | 85.4 | — | — | — | — | |
| iBOTBackbone=ViT-L/16, Pre-training Epochs=250, Training Heuristics=w/ heuristics, Linear Probe Training Epochs=102026.06 | 84.5 | — | — | — | — | |
| Auto-LGMethod Category=Learngene, Backbone=DeiT-S, 11.3M, Para. (M)=7.52024.06 | 84.2 | — | — | — | — | |
| LIGOMethod Category=Trans., Backbone=DeiT-S, 11.3M, Para. (M)=7.52024.06 | 83.8 | — | — | — | — | |
| Direct FTMethod Category=PT, Backbone=DeiT-Ti, 3.0M, Para. (M)=2.92024.06 | 83.5 | — | — | — | — | |
| TLEGMethod Category=Learngene, Backbone=DeiT-S, 11.3M, Para. (M)=3.92024.06 | 83.2 | — | — | — | — | |
| iBOTBackbone=ViT-B/16, Pre-training Epochs=400, Training Heuristics=w/ heuristics, Linear Probe Training Epochs=102026.06 | 83.2 | — | — | — | — | |
| DINOBackbone=ViT-B/16, Pre-training Epochs=400, Training Heuristics=w/ heuristics, Linear Probe Training Epochs=102026.06 | 83.1 | — | — | — | — | |
| WAVEMethod Category=Learngene, Backbone=DeiT-Ti, 3.0M, Para. (M)=1.12024.06 | 82.9 | — | — | — | — | |
| Share InitMethod Category=Trans., Backbone=DeiT-S, 11.3M, Para. (M)=2.22024.06 | 82.1 | — | — | — | — | |
| LIGOMethod Category=Trans., Backbone=DeiT-Ti, 3.0M, Para. (M)=2.02024.06 | 81.1 | — | — | — | — | |
| Auto-LGMethod Category=Learngene, Backbone=DeiT-Ti, 3.0M, Para. (M)=2.02024.06 | 80.9 | — | — | — | — | |
| VISRegBackbone=ViT-L/14, Pre-training Epochs=100, Training Heuristics=w/o heuristics, Linear Probe Training Epochs=1002026.06 | 80.7 | — | — | — | — | |
| MoCoV3Backbone=ViT-B/16, Pre-training Epochs=300, Training Heuristics=w/ heuristics, Linear Probe Training Epochs=102026.06 | 80.5 | — | — | — | — | |
| Share InitMethod Category=Trans., Backbone=DeiT-Ti, 3.0M, Para. (M)=0.62024.06 | 80.3 | — | — | — | — | |
| TLEGMethod Category=Learngene, Backbone=DeiT-Ti, 3.0M, Para. (M)=1.12024.06 | 79.6 | — | — | — | — | |
| LeJEPA*Backbone=ViT-L/14, Pre-training Epochs=100, Training Heuristics=w/o heuristics, Linear Probe Training Epochs=1002026.06 | 79.5 | — | — | — | — | |
| VISRegBackbone=ViT-B/16, Pre-training Epochs=400, Training Heuristics=w/o heuristics, Linear Probe Training Epochs=102026.06 | 79.1 | — | — | — | — | |
| I-JEPABackbone=ViT-H/14, Pre-training Epochs=300, Training Heuristics=w/ heuristics, Linear Probe Training Epochs=102026.06 | 78.7 | — | — | — | — | |
| VISRegBackbone=ViT-L/14, Pre-training Epochs=400, Training Heuristics=w/o heuristics, Linear Probe Training Epochs=102026.06 | 78.5 | — | — | — | — | |
| MAEBackbone=ViT-L/16, Pre-training Epochs=1600, Training Heuristics=w/o heuristics, Linear Probe Training Epochs=102026.06 | 77.8 | — | — | — | — | |
| BiFTAModel Architecture=ViT-L/14, Evaluation Mode=Zero-shot2026.01 | 72.98 | — | — | 0.48 | — | |
| WCAModel Architecture=ViT-L/14, Evaluation Mode=Zero-shot2026.01 | 72.5 | — | — | — | — | |
| data2vecBackbone=ViT-L/14, Pre-training Epochs=1600, Training Heuristics=w/ heuristics, Linear Probe Training Epochs=102026.06 | 72.1 | — | — | — | — | |
| CuPLModel Architecture=ViT-L/14, Evaluation Mode=Zero-shot2026.01 | 71.31 | — | — | — | — | |
| CLIP-EModel Architecture=ViT-L/14, Evaluation Mode=Zero-shot2026.01 | 70.12 | — | — | — | — | |
| CLIP-DModel Architecture=ViT-L/14, Evaluation Mode=Zero-shot2026.01 | 69.87 | — | — | — | — | |
| WaffleModel Architecture=ViT-L/14, Evaluation Mode=Zero-shot2026.01 | 69.61 | — | — | — | — | |
| CLIPModel Architecture=ViT-L/14, Evaluation Mode=Zero-shot2026.01 | 68.94 | — | — | — | — | |
| BiFTAModel Architecture=ViT-B/16, Evaluation Mode=Zero-shot2026.01 | 68.29 | — | — | 0.42 | — | |
| WCAModel Architecture=ViT-B/16, Evaluation Mode=Zero-shot2026.01 | 67.87 | — | — | — | — | |
| Heur-LGMethod Category=Learngene, Backbone=DeiT-S, 11.3M, Para. (M)=5.72024.06 | 67.5 | — | — | — | — | |
| CuPLModel Architecture=ViT-B/16, Evaluation Mode=Zero-shot2026.01 | 66.09 | — | — | — | — | |
| BiFTAModel Architecture=ViT-B/32, Evaluation Mode=Zero-shot2026.01 | 66 | — | — | 1.19 | — | |
| WCAModel Architecture=ViT-B/32, Evaluation Mode=Zero-shot2026.01 | 64.81 | — | — | — | — | |
| CLIP-DModel Architecture=ViT-B/16, Evaluation Mode=Zero-shot2026.01 | 64.67 | — | — | — | — | |
| CLIP-EModel Architecture=ViT-B/16, Evaluation Mode=Zero-shot2026.01 | 64.51 | — | — | — | — | |
| WaffleModel Architecture=ViT-B/16, Evaluation Mode=Zero-shot2026.01 | 64.34 | — | — | — | — | |
| CLIPModel Architecture=ViT-B/16, Evaluation Mode=Zero-shot2026.01 | 63.59 | — | — | — | — | |
| Heur-LGMethod Category=Learngene, Backbone=DeiT-Ti, 3.0M, Para. (M)=1.52024.06 | 63.5 | — | — | — | — | |
| BiFTAModel Architecture=RN-50, Evaluation Mode=Zero-shot2026.01 | 62.54 | — | — | 0.54 | — | |
| CuPLModel Architecture=ViT-B/32, Evaluation Mode=Zero-shot2026.01 | 62.16 | — | — | — | — | |
| BiFTAModel Architecture=RN-101, Evaluation Mode=Zero-shot2026.01 | 62.03 | — | — | 0.89 | — | |
| WCAModel Architecture=RN-50, Evaluation Mode=Zero-shot2026.01 | 62 | — | — | — | — | |
| CLIP-EModel Architecture=ViT-B/32, Evaluation Mode=Zero-shot2026.01 | 61.76 | — | — | — | — | |
| CLIP-DModel Architecture=ViT-B/32, Evaluation Mode=Zero-shot2026.01 | 61.4 | — | — | — | — | |
| WaffleModel Architecture=ViT-B/32, Evaluation Mode=Zero-shot2026.01 | 61.21 | — | — | — | — | |
| WCAModel Architecture=RN-101, Evaluation Mode=Zero-shot2026.01 | 61.14 | — | — | — | — | |
| GHN-3Method Category=Direct, Backbone=DeiT-S, 11.3M, Para. (M)=02024.06 | 61 | — | — | — | — | |
| CLIP-EModel Architecture=RN-101, Evaluation Mode=Zero-shot2026.01 | 60.5 | — | — | — | — | |
| CLIPModel Architecture=ViT-B/32, Evaluation Mode=Zero-shot2026.01 | 60.39 | — | — | — | — | |
| CuPLModel Architecture=RN-50, Evaluation Mode=Zero-shot2026.01 | 60.01 | — | — | — | — | |
| CLIP-DModel Architecture=RN-101, Evaluation Mode=Zero-shot2026.01 | 59.22 | — | — | — | — | |
| CLIPModel Architecture=RN-101, Evaluation Mode=Zero-shot2026.01 | 59.14 | — | — | — | — | |
| CuPLModel Architecture=RN-101, Evaluation Mode=Zero-shot2026.01 | 59.04 | — | — | — | — | |
| WaffleModel Architecture=RN-101, Evaluation Mode=Zero-shot2026.01 | 58.89 | — | — | — | — | |
| CLIP-EModel Architecture=RN-50, Evaluation Mode=Zero-shot2026.01 | 58.64 | — | — | — | — | |
| CLIP-DModel Architecture=RN-50, Evaluation Mode=Zero-shot2026.01 | 58.39 | — | — | — | — | |
| MimeticMethod Category=Direct, Backbone=DeiT-S, 11.3M, Para. (M)=02024.06 | 58.3 | — | — | — | — | |
| WaffleModel Architecture=RN-50, Evaluation Mode=Zero-shot2026.01 | 57.92 | — | — | — | — | |
| GHN-3Method Category=Direct, Backbone=DeiT-Ti, 3.0M, Para. (M)=02024.06 | 57.8 | — | — | — | — | |
| Wt SelectMethod Category=Trans., Backbone=DeiT-S, 11.3M, Para. (M)=11.02024.06 | 57.8 | — | — | — | — | |
| CLIPModel Architecture=RN-50, Evaluation Mode=Zero-shot2026.01 | 56.97 | — | — | — | — | |
| He-InitMethod Category=Direct, Backbone=DeiT-S, 11.3M, Para. (M)=02024.06 | 56.2 | — | — | — | — | |
| Wt SelectMethod Category=Trans., Backbone=DeiT-Ti, 3.0M, Para. (M)=2.92024.06 | 55.6 | — | — | — | — | |
| He-InitMethod Category=Direct, Backbone=DeiT-Ti, 3.0M, Para. (M)=02024.06 | 54.5 | — | — | — | — | |
| MimeticMethod Category=Direct, Backbone=DeiT-Ti, 3.0M, Para. (M)=02024.06 | 53.7 | — | — | — | — | |
| AdvFLYP_fullBackbone=CLIP ViT-B/16, Adversarial Epsilon (epsilon)=1/2552026.04 | — | 56.07 | 35.13 | — | 33.62 | |
| AdvVPTraining Shots=16-shot (1.25%), Params (/M)=0.07, Time (/Day)=0.652024.03 | — | 41.96 | 12.97 | — | — | |
| AdvVPTraining Shots=Entire (100%), Params (/M)=0.24, Time (/Day)=49.92024.03 | — | 46.58 | 25.21 | — | — | |
| FAPTraining Shots=16-shot (1.25%), Params (/M)=0.42, Time (/Day)=0.712024.03 | — | 48.18 | 25.06 | — | — | |
| FAPTraining Shots=32-shot (2.49%), Params (/M)=0.43, Time (/Day)=1.432024.03 | — | 49.93 | 25.39 | — | — | |
| PMG-AFTBackbone=CLIP ViT-B/16, Adversarial Epsilon (epsilon)=1/2552026.04 | — | 54.46 | 31.59 | — | 28.92 |