Image Classification on VTAB 1k
79Overall Mean Accuracy+PACE
Evaluation results
| Method | Links | ||||||||||||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| +PACEBackbone=ViT-B/16, Resolution=224x2242024.09 | 79 | 79 | 94.2 | 73.6 | 99.4 | 92.4 | 93.7 | 58 | — | 87.4 | 96.4 | 89.3 | 77.1 | — | 84.9 | 70.9 | 54.9 | 84.3 | 84.7 | 57.3 | 39.3 | 44.8 | — | — | |
| DyTr=0.5, Param. (M)=0.44, FLOPS (G)=43.522024.03 | 78.5 | 79.1 | 95.6 | 74.5 | 99.5 | 92.6 | 90.8 | 59.3 | — | 86.9 | 96.6 | 87.2 | 76.5 | — | 84.5 | 62.9 | 53.3 | 83.5 | 88.4 | 57.3 | 38.7 | 44.6 | — | — | |
| LoSAParam (10^6)=0.19, Rank (r)=16, Backbone=ViT-B2024.02 | 78.4 | 82.5 | 92.8 | 76.1 | 99.7 | 90.5 | 82 | 55.8 | 82.8 | 86.6 | 97.1 | 87 | 76.7 | 86.9 | 81.5 | 62.3 | 48.6 | 82.1 | 94.2 | 61.7 | 47.9 | 45.6 | 65.5 | — | |
| LoSAParam (10^6)=0.10, Rank (r)=8, Backbone=ViT-B2024.02 | 78.3 | 82.2 | 92.7 | 76.7 | 99.7 | 90.7 | 81 | 55.4 | 82.6 | 86.9 | 97.1 | 87.4 | 76.5 | 87 | 79.9 | 61.8 | 48.6 | 82.4 | 92.3 | 61.1 | 48.7 | 47.3 | 65.3 | — | |
| OursParams. (M)=5.50, Memory (G)=2.1, FPS=88.892026.04 | 78.3 | 79.8 | 93.6 | 75.7 | 99.5 | 91.8 | 91.5 | 56.1 | — | 86.9 | 96.8 | 87.5 | 75.9 | — | 84.6 | 64.6 | 52.4 | 81.3 | 87.2 | 57.1 | 37.7 | 47.9 | — | — | |
| HSTParam (10^6)=0.78, Backbone=ViT-B2024.02 | 78.1 | 76.7 | 94.1 | 74.8 | 99.6 | 91.1 | 91.2 | 52.3 | 82.8 | 87.1 | 96.3 | 88.6 | 76.5 | 87.1 | 85.4 | 63.7 | 52.9 | 81.7 | 87.2 | 56.8 | 35.8 | 52.1 | 64.5 | — | |
| HSTParams. (M)=0.78, Memory (G)=2.8, FPS=44.172026.04 | 78.1 | 76.7 | 94.1 | 74.8 | 99.6 | 91.1 | 91.2 | 52.3 | — | 87.1 | 96.3 | 88.6 | 76.5 | — | 85.4 | 63.7 | 52.9 | 81.7 | 87.2 | 56.8 | 35.8 | 52.1 | — | — | |
| GLORABackbone=ViT-B/16, Resolution=224x2242024.09 | 78 | 76.4 | 92.9 | 74.6 | 99.6 | 92.5 | 91.5 | 57.8 | — | 87.3 | 96.8 | 88 | 76 | — | 83.1 | 67.3 | 54.5 | 86.2 | 83.8 | 52.9 | 37 | 41.4 | — | — | |
| SHERLParams. (M)=5.08, Memory (G)=2.0, FPS=48.702026.04 | 77.9 | 78.8 | 93.2 | 75.4 | 99.5 | 91.6 | 90.4 | 55.6 | — | 86.8 | 96.7 | 87.2 | 75.3 | — | 84.2 | 63.5 | 51.8 | 81.7 | 86.9 | 57.5 | 36.9 | 48.2 | — | — | |
| LoSAParam (10^6)=0.05, Rank (r)=4, Backbone=ViT-B2024.02 | 77.8 | 82.7 | 93 | 76.2 | 99.7 | 89.8 | 80 | 56.1 | 82.5 | 86.3 | 96.7 | 86.7 | 76.3 | 86.5 | 78.8 | 61.4 | 48 | 82.6 | 91.7 | 58.4 | 46.9 | 47.6 | 64.4 | — | |
| LoSAParams. (M)=0.05, Memory (G)=2.1, FPS=68.922026.04 | 77.8 | 82.7 | 93 | 76.2 | 99.7 | 89.8 | 80 | 56.1 | — | 86.3 | 96.7 | 86.7 | 76.3 | — | 78.8 | 61.4 | 48 | 82.6 | 91.7 | 58.4 | 46.9 | 47.6 | — | — | |
| HyperAdapterBackbone=ViT-L/16, Pre-training=ImageNet-21K, Params (M)=1.182026.06 | 77.7 | — | — | — | — | — | — | — | 83.7 | — | — | — | — | 85.6 | — | — | — | — | — | — | — | — | 63.8 | — | |
| HyperAdapterBackbone=Swin-Base, Pre-trained=ImageNet-21K, Params (M)=0.602026.06 | 77.6 | — | — | — | — | — | — | — | 83.5 | — | — | — | — | 86.2 | — | — | — | — | — | — | — | — | 63 | — | |
| HyperAdapterBackbone=ViT-L2026.06 | 77.6 | — | — | — | — | — | — | — | 82.3 | — | — | — | — | 86.6 | — | — | — | — | — | — | — | — | 63.8 | — | |
| DyTTuning Protocol=Dynamic Tuning, Loss Configuration=with distillation, Sampling Ratio (r)=0.7, Params. (M)=0.162024.03 | 77.57 | 74.4 | 95.5 | 73.6 | 99.2 | 91.7 | 87.5 | 57.4 | — | 88.3 | 96.1 | 86.7 | 76.7 | — | 83.5 | 63.8 | 52.9 | 83.1 | 85.7 | 54.9 | 34.3 | 45.9 | — | — | |
| SA2VP + PAEBackbone=ViT-Base/16, Pre-trained=supervised ImageNet-21k, Speedup(x)=1.602026.01 | 77.49 | — | — | — | — | — | — | — | 82.86 | — | — | — | — | 86.58 | — | — | — | — | — | — | — | — | 63.05 | — | |
| UniPTParams. (M)=5.19, Memory (G)=2.1, FPS=45.722026.04 | 77.4 | 77.4 | 92.9 | 74.8 | 99.6 | 91.6 | 89.7 | 54.9 | — | 87 | 96.5 | 86.9 | 73.8 | — | 83.8 | 63.7 | 50.9 | 81.4 | 86.5 | 56.4 | 35.9 | 46.8 | — | — | |
| DyTTuning Protocol=Dynamic Tuning, Loss Configuration=with distillation, Sampling Ratio (r)=0.9, Params. (M)=0.162024.03 | 77.3 | 74.3 | 94.9 | 73.1 | 99.2 | 91.4 | 87.8 | 57.1 | — | 87.9 | 96.1 | 85.9 | 76 | — | 83.3 | 64.8 | 51.5 | 83.4 | 84 | 54.8 | 35.1 | 46.4 | — | — | |
| DyTr=0.5, FLOPs (G)=12.54, V100 Throughput (img/s)=912.3, T4 Throughput (img/s)=524.93, Xeon(R) 8163 Throughput (img/s)=3.892024.03 | 77.14 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| DyTr=0.5, FLOPs (G)=12.54, Param. (M)=0.16, Throughput (img/s)=912.392024.03 | 77.14 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| DyTTuning Protocol=Dynamic Tuning, Loss Configuration=with distillation, Sampling Ratio (r)=0.5, Params. (M)=0.162024.03 | 77.14 | 73.6 | 94.8 | 73 | 99.1 | 91.4 | 87 | 56.4 | — | 87.3 | 96.1 | 85.6 | 76.7 | — | 82.8 | 63.8 | 52.7 | 83.7 | 83.6 | 57.3 | 34.6 | 44.3 | — | — | |
| OursBackbone=ViT-Base/16, Tuned/Total (%)=0.75, Extra Params=true2026.06 | 77.01 | — | — | — | — | — | — | — | 82.88 | — | — | — | — | 85.61 | — | — | — | — | — | — | — | — | 62.55 | — | |
| RepAdapterBackbone=Swin-Base, Pre-trained=ImageNet-21K, Params (M)=0.392026.06 | 77 | — | — | — | — | — | — | — | 82.8 | — | — | — | — | 87.2 | — | — | — | — | — | — | — | — | 61.2 | — | |
| BPT + PAEBackbone=ViT-Base/16, Pre-trained=supervised ImageNet-21k, Speedup(x)=1.372026.01 | 76.94 | — | — | — | — | — | — | — | 82.46 | — | — | — | — | 86.33 | — | — | — | — | — | — | — | — | 62.05 | — | |
| SNELL-32Backbone=ViT-B/16, Pre-training=ImageNet-21k supervisedly2024.11 | 76.9 | — | — | — | — | — | — | — | 82.7 | — | — | — | — | 86.1 | — | — | — | — | — | — | — | — | 61.8 | — | |
| SNELL-32Backbone=ViT-B/16, Pre-trained=ImageNet-21K2024.11 | 76.9 | — | — | — | — | — | — | — | 82.7 | — | — | — | — | 86.1 | — | — | — | — | — | — | — | — | 61.8 | — | |
| RepAdapterBackbone=ViT-L/16, Pre-training=ImageNet-21K, Params (M)=0.792026.06 | 76.8 | — | — | — | — | — | — | — | 84 | — | — | — | — | 86.3 | — | — | — | — | — | — | — | — | 60.1 | — | |
| DyTTuning Protocol=Dynamic Tuning, Loss Configuration=without L'cls + Ldistill, Sampling Ratio (r)=0.9, Params. (M)=0.162024.03 | 76.74 | 74 | 95.1 | 72.9 | 99.3 | 91.7 | 87.6 | 56.9 | — | 87.7 | 95.7 | 85.4 | 76.1 | — | 81.6 | 63.2 | 50.1 | 83 | 83.3 | 52 | 34.5 | 44.5 | — | — | |
| SNELL-16Backbone=ViT-B/16, Pre-training=ImageNet-21k supervisedly2024.11 | 76.7 | — | — | — | — | — | — | — | 82.4 | — | — | — | — | 86.1 | — | — | — | — | — | — | — | — | 61.7 | — | |
| SNELL-16Backbone=ViT-B/16, Pre-trained=ImageNet-21K2024.11 | 76.7 | — | — | — | — | — | — | — | 82.4 | — | — | — | — | 86.1 | — | — | — | — | — | — | — | — | 61.7 | — | |
| RLRRBackbone=ViT-B/16, Resolution=224x2242024.09 | 76.7 | 75.6 | 92.4 | 72.9 | 99.3 | 91.5 | 89.8 | 57 | — | 86.8 | 95.2 | 85.3 | 75.9 | — | 79.7 | 64.2 | 53.9 | 82.1 | 83.9 | 53.7 | 33.4 | 43.6 | — | — | |
| DyTTuning Protocol=Dynamic Tuning, Loss Configuration=without L'cls + Ldistill, Sampling Ratio (r)=0.7, Params. (M)=0.162024.03 | 76.69 | 73.9 | 94.9 | 72.1 | 99.4 | 91.8 | 88.4 | 55.5 | — | 87.2 | 95.6 | 86.2 | 75.9 | — | 80.3 | 61.8 | 51.7 | 83.1 | 81.6 | 53.7 | 35.3 | 45.2 | — | — | |
| ConvpassParam (10^6)=0.33, Backbone=ViT-B2024.02 | 76.6 | 72.3 | 91.2 | 72.2 | 99.2 | 90.9 | 91.3 | 54.9 | 81.7 | 84.2 | 96.1 | 85.3 | 75.6 | 85.3 | 82.3 | 67.9 | 51.3 | 80 | 85.9 | 53.1 | 36.4 | 44.4 | 62.7 | — | |
| DyT + ToMer=0.5, FLOPs (G)=9.85, Param. (M)=0.16, Throughput (img/s)=1114.72024.03 | 76.6 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| ConvpassParams. (M)=0.33, Memory (G)=2.6, FPS=67.022026.04 | 76.6 | 72.3 | 91.2 | 72.2 | 99.2 | 90.9 | 91.3 | 54.9 | — | 84.2 | 96.1 | 85.3 | 75.6 | — | 82.3 | 67.9 | 51.3 | 80 | 85.9 | 53.1 | 36.4 | 44.4 | — | — | |
| ConvPassTuning Protocol=Parameter-efficient tuning, Params. (M)=0.332024.03 | 76.56 | 72.3 | 91.2 | 72.2 | 99.2 | 90.9 | 91.3 | 54.9 | — | 84.2 | 96.1 | 85.3 | 75.6 | — | 82.3 | 67.9 | 51.3 | 80 | 85.9 | 53.1 | 36.4 | 44.4 | — | — | |
| ProVP + PAEBackbone=ViT-Base/16, Pre-trained=supervised ImageNet-21k, Speedup(x)=1.192026.01 | 76.48 | — | — | — | — | — | — | — | 82.33 | — | — | — | — | 85.41 | — | — | — | — | — | — | — | — | 61.72 | — | |
| SPT-AdapterBackbone=ViT-B/16, Pre-training=ImageNet-21k supervisedly2024.11 | 76.4 | — | — | — | — | — | — | — | 82 | — | — | — | — | 85.8 | — | — | — | — | — | — | — | — | 61.4 | — | |
| SPT-LoRABackbone=ViT-B/16, Pre-training=ImageNet-21k supervisedly2024.11 | 76.4 | — | — | — | — | — | — | — | 81.9 | — | — | — | — | 85.9 | — | — | — | — | — | — | — | — | 61.3 | — | |
| SNELL-8Backbone=ViT-B/16, Pre-training=ImageNet-21k supervisedly2024.11 | 76.4 | — | — | — | — | — | — | — | 82 | — | — | — | — | 85.7 | — | — | — | — | — | — | — | — | 61.6 | — | |
| SNELL-8Backbone=ViT-B/16, Pre-trained=ImageNet-21K2024.11 | 76.4 | — | — | — | — | — | — | — | 82 | — | — | — | — | 85.7 | — | — | — | — | — | — | — | — | 61.6 | — | |
| BaselineBackbone=ViT-B/16, Resolution=224x2242024.09 | 76.4 | 74.9 | 93.3 | 72 | 99.4 | 91 | 91.5 | 54.8 | — | 83.2 | 95.7 | 86.9 | 74.2 | — | 83 | 70.5 | 51.9 | 81.4 | 77.9 | 51.7 | 33.6 | 44.4 | — | — | |
| ViaPTBackbone=ViT-L2026.06 | 76.4 | — | — | — | — | — | — | — | 82.6 | — | — | — | — | 85.2 | — | — | — | — | — | — | — | — | 61.3 | — | |
| VFPT + PAEBackbone=ViT-Base/16, Pre-trained=supervised ImageNet-21k, Speedup(x)=1.272026.01 | 76.33 | — | — | — | — | — | — | — | 82.07 | — | — | — | — | 85.96 | — | — | — | — | — | — | — | — | 60.96 | — | |
| Res-TuningTuning Protocol=Parameter-efficient tuning, Params. (M)=0.512024.03 | 76.32 | 75.2 | 92.7 | 71.9 | 99.3 | 91.9 | 86.7 | 58.5 | — | 86.7 | 95.6 | 85 | 74.6 | — | 80.2 | 63.6 | 50.6 | 80.2 | 85.4 | 55.7 | 31.9 | 42 | — | — | |
| RSParam (10^6)=0.55, Backbone=ViT-B2024.02 | 76.3 | 75.2 | 92.7 | 71.9 | 99.3 | 91.9 | 86.7 | 58.5 | 82.3 | 86.7 | 95.6 | 85 | 74.6 | 85.5 | 80.2 | 63.6 | 50.6 | 80.2 | 85.4 | 55.7 | 31.9 | 42 | 61.2 | — | |
| Res-TuningParams. (M)=0.55, Memory (G)=4.9, FPS=57.702026.04 | 76.3 | 75.2 | 92.7 | 71.9 | 99.3 | 91.9 | 86.7 | 58.5 | — | 86.7 | 95.6 | 85 | 74.6 | — | 80.2 | 63.6 | 50.6 | 80.2 | 85.4 | 55.7 | 31.9 | 42 | — | — | |
| RepAdapterParam (10^6)=0.22, Backbone=ViT-B2024.02 | 76.1 | 72.4 | 91.6 | 71 | 99.2 | 91.4 | 90.7 | 55.1 | 81.6 | 85.3 | 95.9 | 84.6 | 75.9 | 85.4 | 82.3 | 68 | 50.4 | 79.9 | 80.4 | 49.2 | 38.6 | 41 | 61.2 | — | |
| RepAdapterBackbone=ViT-B/16, Resolution=224x2242024.09 | 76.1 | 69 | 92.6 | 75.1 | 99.4 | 91.8 | 90.2 | 52.9 | — | 87.4 | 95.9 | 87.4 | 75.5 | — | 75.9 | 62.3 | 53.3 | 80.6 | 77.3 | 54.9 | 29.5 | 37.9 | — | — | |
| LoCA (r4)Backbone=ConvNeXt-B, Param. (M)=0.972026.07 | 75.98 | — | — | — | — | — | — | — | 82.22 | — | — | — | — | 87.54 | — | — | — | — | — | — | — | — | 64.74 | — | |
| LoCA (r16)Backbone=ConvNeXt-B, Param. (M)=5.012026.07 | 75.93 | — | — | — | — | — | — | — | 81.81 | — | — | — | — | 87.51 | — | — | — | — | — | — | — | — | 65.01 | — | |
| MAE-RefinerBackbone=ViT-H/14, Protocol=Fine-tuning2024.02 | 75.9 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| VQT + PAEBackbone=ViT-Base/16, Pre-trained=supervised ImageNet-21k, Speedup(x)=1.522026.01 | 75.85 | — | — | — | — | — | — | — | 82.39 | — | — | — | — | 85.34 | — | — | — | — | — | — | — | — | 59.83 | — | |
| SA2VPBackbone=ViT-Base/16, Pre-trained=supervised ImageNet-21k2026.01 | 75.83 | — | — | — | — | — | — | — | 80.97 | — | — | — | — | 85.73 | — | — | — | — | — | — | — | — | 60.8 | — | |
| SA2VPBackbone=ViT-Base/16, Tuned/Total (%)=0.81, Extra Params=true2026.06 | 75.83 | — | — | — | — | — | — | — | 80.97 | — | — | — | — | 85.73 | — | — | — | — | — | — | — | — | 60.8 | — | |
| ConvpassattnParam (10^6)=0.16, Backbone=ViT-B2024.02 | 75.8 | 71.8 | 90.7 | 72 | 99.1 | 91 | 89.9 | 54.2 | 81.2 | 85.2 | 95.6 | 83.4 | 74.8 | 84.8 | 79.9 | 67 | 50.3 | 79.9 | 84.3 | 53.2 | 34.8 | 43 | 61.6 | — | |
| DyTTuning Protocol=Dynamic Tuning, Loss Configuration=without L'cls + Ldistill, Sampling Ratio (r)=0.5, Params. (M)=0.162024.03 | 75.73 | 70.4 | 94.2 | 71.1 | 99.1 | 91.7 | 88 | 51.5 | — | 87.1 | 95.3 | 84.2 | 75.8 | — | 79.2 | 61.8 | 51 | 82.4 | 79.7 | 52.3 | 35.3 | 44.5 | — | — | |
| SSFParam (10^6)=0.24, Backbone=ViT-B2024.02 | 75.7 | 69 | 92.6 | 75.1 | 99.4 | 91.8 | 90.2 | 52.9 | 81.6 | 87.4 | 95.9 | 87.4 | 75.5 | 86.6 | 75.9 | 62.3 | 53.3 | 80.6 | 77.3 | 54.9 | 29.5 | 37.9 | 59 | — | |
| Full tuningParam. (M)=303.3, FLOPS (G)=61.62024.03 | 75.7 | 69.5 | 96.2 | 73.8 | 98.8 | 90.7 | 91.6 | 44.8 | — | 85.8 | 96.2 | 87.8 | 75.3 | — | 83 | 62 | 50.8 | 80 | 85.8 | 54.6 | 29.7 | 35.4 | — | — | |
| SSFBackbone=ViT-B/16, Pre-training=ImageNet-21k supervisedly2024.11 | 75.7 | — | — | — | — | — | — | — | 81.6 | — | — | — | — | 86.6 | — | — | — | — | — | — | — | — | 59 | — | |
| MAE-CTBackbone=ViT-H/14, Protocol=Fine-tuning2024.02 | 75.7 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| SSFTuning Protocol=Parameter-efficient tuning, Params. (M)=0.22024.03 | 75.69 | 69 | 92.6 | 75.1 | 99.4 | 91.8 | 90.2 | 52.9 | — | 87.4 | 95.9 | 87.4 | 75.5 | — | 75.9 | 62.3 | 53.3 | 80.6 | 77.3 | 54.9 | 29.5 | 37.9 | — | — | |
| E2VPT + PAEBackbone=ViT-Base/16, Pre-trained=supervised ImageNet-21k, Speedup(x)=1.652026.01 | 75.62 | — | — | — | — | — | — | — | 81.39 | — | — | — | — | 85.76 | — | — | — | — | — | — | — | — | 59.73 | — | |
| FacT-TKParam (10^6)=0.06, Backbone=ViT-B2024.02 | 75.6 | 70.6 | 90.6 | 70.8 | 99.1 | 90.7 | 88.6 | 54.1 | 80.6 | 84.8 | 96.2 | 84.5 | 75.7 | 85.3 | 82.6 | 68.2 | 49.8 | 80.7 | 80.8 | 47.4 | 33.2 | 43 | 60.7 | — | |
| MAE-RefinerBackbone=ViT-2B/14, Protocol=Fine-tuning2024.02 | 75.6 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| FacTParams. (M)=0.06, Memory (G)=4.82026.04 | 75.6 | 70.6 | 90.6 | 70.8 | 99.1 | 90.7 | 88.6 | 54.1 | — | 84.8 | 96.2 | 84.5 | 75.7 | — | 82.6 | 68.2 | 49.8 | 80.7 | 80.8 | 47.4 | 33.2 | 43 | — | — | |
| ARCBackbone=Swin-Base, Pre-trained=ImageNet-21K, Params (M)=0.162026.06 | 75.6 | — | — | — | — | — | — | — | 79 | — | — | — | — | 86.6 | — | — | — | — | — | — | — | — | 59.9 | — | |
| NOAHParam (10^6)=0.36, Backbone=ViT-B2024.02 | 75.5 | 69.6 | 92.7 | 70.2 | 99.1 | 90.4 | 86.1 | 53.7 | 80.3 | 84.4 | 95.4 | 83.9 | 75.8 | 84.9 | 82.8 | 68.9 | 49.9 | 81.7 | 81.8 | 48.3 | 32.8 | 44.2 | 61.3 | — | |
| NOAHParams. (M)=0.36, Memory (G)=5.2, FPS=56.782026.04 | 75.5 | 69.6 | 92.7 | 70.2 | 99.1 | 90.4 | 86.1 | 53.7 | — | 84.4 | 95.4 | 83.9 | 75.8 | — | 82.8 | 68.9 | 49.9 | 81.7 | 81.8 | 48.3 | 32.8 | 44.2 | — | — | |
| VFPTBackbone=ViT-L2026.06 | 75.5 | — | — | — | — | — | — | — | 81.4 | — | — | — | — | 84.9 | — | — | — | — | — | — | — | — | 60.2 | — | |
| VFPTBackbone=ViT-Base/16, Tuned/Total (%)=0.66, Extra Params=true2026.06 | 75.49 | — | — | — | — | — | — | — | 81.35 | — | — | — | — | 84.93 | — | — | — | — | — | — | — | — | 60.19 | — | |
| DynamicViT + AdaptFormerFLOPs (G)=14.23, Param. (M)=3.1, Throughput (img/s)=954.822024.03 | 75.48 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| NOAHTuning Protocol=Parameter-efficient tuning, Params. (M)=0.362024.03 | 75.48 | 69.6 | 92.7 | 70.2 | 99.1 | 90.4 | 86.1 | 53.7 | — | 84.4 | 95.4 | 83.9 | 75.8 | — | 82.8 | 68.9 | 49.9 | 81.7 | 81.8 | 48.3 | 32.8 | 44.2 | — | — | |
| HTABackbone=ViT-Large, Pre-training=ImageNet-21k, Params.(M)=0.542024.10 | 75.4 | — | — | — | — | — | — | — | 84.1 | — | — | — | — | 86.6 | — | — | — | — | — | — | — | — | 62.3 | — | |
| VFPTBackbone=ViT-Base/16, Pre-trained=supervised ImageNet-21k2026.01 | 75.39 | — | — | — | — | — | — | — | 81.35 | — | — | — | — | 84.93 | — | — | — | — | — | — | — | — | 60.19 | — | |
| FacT-TTTuning Protocol=Parameter-efficient tuning, Params. (M)=0.042024.03 | 75.3 | 71.3 | 89.6 | 70.7 | 98.9 | 91 | 87.8 | 54.6 | — | 85.2 | 95.5 | 83.4 | 75.7 | — | 82 | 69 | 49.8 | 80 | 79.2 | 48.4 | 34.2 | 41.4 | — | — | |
| RLRRBackbone=ViT-Large, Pre-training=ImageNet-21k, Params.(M)=0.822024.10 | 75.2 | — | — | — | — | — | — | — | 83.9 | — | — | — | — | 86.4 | — | — | — | — | — | — | — | — | 61.9 | — | |
| MAE-RefinerBackbone=ViT-L/16, Protocol=Fine-tuning2024.02 | 75.2 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| CABackbone=ConvNeXt-B, Param. (M)=6.832026.07 | 75.18 | — | — | — | — | — | — | — | 80.62 | — | — | — | — | 86.29 | — | — | — | — | — | — | — | — | 64.88 | — | |
| LPT + PAEBackbone=ViT-Base/16, Pre-trained=supervised ImageNet-21k, Speedup(x)=1.442026.01 | 75.15 | — | — | — | — | — | — | — | 81.01 | — | — | — | — | 85.02 | — | — | — | — | — | — | — | — | 59.41 | — | |
| RLRR*Backbone=ViT-B/16, Pre-training=ImageNet-21k, AugReg=true, Params.(M)=0.332024.03 | 75.1 | 76.7 | 92.7 | 76.3 | 99.6 | 92.6 | 91.8 | 56 | 83.7 | 87.8 | 96.2 | 89.1 | 76.3 | 87.3 | 80.4 | 63.3 | 54.5 | 83.3 | 83 | 53.7 | 32 | 41.7 | 61.5 | — | |
| ARCBackbone=ViT-L/16, Pre-training=ImageNet-21K, Params (M)=0.742026.06 | 75.1 | — | — | — | — | — | — | — | 82.3 | — | — | — | — | 85.6 | — | — | — | — | — | — | — | — | 57.3 | — | |
| BPTBackbone=ViT-Base/16, Pre-trained=supervised ImageNet-21k2026.01 | 75.02 | — | — | — | — | — | — | — | 80.24 | — | — | — | — | 84.45 | — | — | — | — | — | — | — | — | 60.39 | — | |
| LoRA-16Backbone=ViT-B/16, Pre-training=ImageNet-21k supervisedly2024.11 | 75 | — | — | — | — | — | — | — | 79.8 | — | — | — | — | 84.9 | — | — | — | — | — | — | — | — | 60.2 | — | |
| Full fine-tuningBackbone=Swin-Base, Pre-trained=ImageNet-21K, Params (M)=86.902026.06 | 75 | — | — | — | — | — | — | — | 79.1 | — | — | — | — | 86.2 | — | — | — | — | — | — | — | — | 59.7 | — | |
| FFTBackbone=ConvNeXt-B, Param. (M)=87.622026.07 | 74.98 | — | — | — | — | — | — | — | 80.52 | — | — | — | — | 87.54 | — | — | — | — | — | — | — | — | 63.85 | — | |
| ProVPBackbone=ViT-Base/16, Pre-trained=supervised ImageNet-21k2026.01 | 74.91 | — | — | — | — | — | — | — | 80.35 | — | — | — | — | 84.07 | — | — | — | — | — | — | — | — | 60.31 | — | |
| LoRA-8Backbone=ViT-B/16, Pre-training=ImageNet-21k supervisedly2024.11 | 74.9 | — | — | — | — | — | — | — | 79.5 | — | — | — | — | 84.6 | — | — | — | — | — | — | — | — | 60.5 | — | |
| VPT + PAEBackbone=ViT-Base/16, Pre-trained=supervised ImageNet-21k, Speedup(x)=1.782026.01 | 74.84 | — | — | — | — | — | — | — | 81.73 | — | — | — | — | 84.52 | — | — | — | — | — | — | — | — | 58.28 | — | |
| AdaptFormerFLOPs (G)=17.61, V100 Throughput (img/s)=767.3, T4 Throughput (img/s)=400.42, Xeon(R) 8163 Throughput (img/s)=1.972024.03 | 74.75 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| AdaptFormerTuning Protocol=Parameter-efficient tuning, Params. (M)=0.162024.03 | 74.75 | 70.8 | 91.2 | 70.5 | 99.1 | 90.9 | 86.6 | 54.8 | — | 83 | 95.8 | 84.4 | 76.3 | — | 81.9 | 64.3 | 49.3 | 80.3 | 76.3 | 45.7 | 31.7 | 41.1 | — | — | |
| AdaptFormerParam (10^6)=0.16, Backbone=ViT-B2024.02 | 74.7 | 70.8 | 91.2 | 70.5 | 99.1 | 90.9 | 86.6 | 54.8 | 80.6 | 83 | 95.8 | 84.4 | 76.3 | 84.9 | 81.9 | 64.3 | 49.3 | 80.3 | 76.3 | 45.7 | 31.7 | 41.1 | 58.8 | — | |
| AdaptFormerBackbone=ViT-B/16, Resolution=224x2242024.09 | 74.7 | 70.8 | 91.2 | 70.5 | 99.1 | 90.9 | 86.6 | 54.8 | — | 83 | 95.8 | 84.4 | 76.3 | — | 81.9 | 64.3 | 49.3 | 80.3 | 76.3 | 45.7 | 31.7 | 41.1 | — | — | |
| AdaptFormerParams. (M)=0.16, Memory (G)=4.1, FPS=80.842026.04 | 74.7 | 70.8 | 91.2 | 70.5 | 99.1 | 90.9 | 86.6 | 54.8 | — | 83 | 95.8 | 84.4 | 76.3 | — | 81.9 | 64.3 | 49.3 | 80.3 | 76.3 | 45.7 | 31.7 | 41.1 | — | — | |
| LoR-VPBackbone=ViT-Base/16, Extra Params=true2026.06 | 74.69 | — | — | — | — | — | — | — | 80.25 | — | — | — | — | 85.12 | — | — | — | — | — | — | — | — | 58.71 | — | |
| EViT + AdaptFormerFLOPs (G)=11.77, Param. (M)=0.16, Throughput (img/s)=1152.382024.03 | 74.63 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| LoRAFLOPs (G)=17.58, V100 Throughput (img/s)=806.24, T4 Throughput (img/s)=435.41, Xeon(R) 8163 Throughput (img/s)=2.122024.03 | 74.6 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| LoRATuning Protocol=Parameter-efficient tuning, Params. (M)=0.292024.03 | 74.6 | 67.1 | 91.4 | 69.4 | 98.8 | 90.4 | 85.3 | 54 | — | 84.9 | 95.3 | 84.4 | 73.6 | — | 82.9 | 69.2 | 49.8 | 78.5 | 75.7 | 47.1 | 31 | 44 | — | — | |
| LoRAParam (10^6)=0.29, Backbone=ViT-B2024.02 | 74.5 | 67.1 | 91.4 | 69.4 | 98.8 | 90.4 | 85.3 | 54 | 79.5 | 84.9 | 95.3 | 84.4 | 73.6 | 84.6 | 82.9 | 69.2 | 49.8 | 78.5 | 75.7 | 47.1 | 31 | 44 | 59.8 | — | |
| RLRRBackbone=ViT-B/16, Pre-training=ImageNet-21k, Params.(M)=0.332024.03 | 74.5 | 75.6 | 92.4 | 72.9 | 99.3 | 91.5 | 89.8 | 57 | 82.7 | 86.8 | 95.2 | 85.3 | 75.9 | 85.8 | 79.7 | 64.2 | 53.9 | 82.1 | 83.9 | 53.7 | 33.4 | 43.6 | 61.8 | — |