Visual Task Adaptation on VTAB 1K
78.02Average AccuracyPVP
Evaluation results
| Method | Links | |||||||||||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| PVPMethod Category=PETuning Methods2023.04 | 78.02 | 76.3 | 94.4 | 73.1 | 99.7 | 92.3 | 87.3 | 58.6 | 87.5 | 95.6 | 87.4 | 76.9 | 76.4 | 64.6 | 54.6 | 82 | 88 | 58.5 | 36.2 | 52.8 | 83.09 | 86.84 | 64.13 | |
| HyperAdapterBackbone=ViT-B/16, # Param (M)=0.442026.06 | 77.6 | 74.1 | 93.3 | 72.8 | 99.3 | 91.7 | 88.3 | 56.7 | 87.5 | 96.3 | 86 | 76.5 | 84 | 64.4 | 55.2 | 83.9 | 88.5 | 54.6 | 36.3 | 43.7 | — | — | — | |
| LilyParams(M)=0.318, Backbone=ViT-B, Pre-training=ImageNet-21K2024.07 | 77.3 | — | — | — | 99.3 | 91.6 | 89 | 56.6 | 87.9 | 95.2 | 84.9 | 75.7 | 83.9 | 65.4 | 53.4 | 81.6 | 88.2 | 54.5 | 37 | 45.4 | 73.9 | 93 | 72.9 | |
| Bi-AdaptFormer# param (M)=0.642025.03 | 77 | 74.1 | 92.4 | 72.1 | 99.3 | 91.6 | 89 | 56.3 | 88.2 | 95.2 | 86 | 76.2 | 83.9 | 63.6 | 53 | 81.4 | 86.2 | 54.8 | 35.2 | 41.3 | — | — | — | |
| Bi-Adaptformer (Ours)# param (M)=0.642025.03 | 77 | 74.1 | 92.8 | 72.5 | 99.4 | 91.1 | 89.6 | 56.2 | 88.3 | 94.9 | 86.3 | 75.3 | 83.8 | 63 | 52.8 | 81.4 | 85.7 | 54.4 | 35.9 | 42.2 | — | — | — | |
| AdaptFormerParams(M)=0.588, Backbone=ViT-B, Pre-training=ImageNet-21K2024.07 | 76.8 | — | — | — | 99.3 | 91.7 | 88.9 | 56.4 | 87.2 | 95.1 | 85.7 | 75.9 | 84.2 | 62.2 | 53 | 81 | 87.1 | 53.6 | 35.3 | 42.3 | 74 | 92.2 | 71.7 | |
| Bi-LoRA# param (M)=1.172025.03 | 76.7 | 72.1 | 91.7 | 71.2 | 99.1 | 91.4 | 90.2 | 55.8 | 87 | 95.4 | 85.5 | 75.5 | 83.1 | 64.1 | 52.2 | 81.3 | 86.4 | 53.5 | 36.7 | 44.4 | — | — | — | |
| MISTBackbone=ViT-B/16, Support Edge Tuning=true, Transfer Overhead (MB/img)=0.577, Trainable Params (M)=0.38, MACs Cloud (x10^9)=17.57, MACs Edge (x10^9)=0.09, GPU Mem Cloud (GB)=0.57, GPU Mem Edge (GB)=0.452024.06 | 76.7 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 81.4 | 86.4 | 62.4 | |
| Bi-LORAParams(M)=1.180, Backbone=ViT-B, Pre-training=ImageNet-21K2024.07 | 76.7 | — | — | — | 99.1 | 91.4 | 90.2 | 55.8 | 87 | 95.4 | 85.5 | 75.5 | 83.1 | 64.1 | 52.2 | 81.3 | 86.4 | 53.5 | 36.7 | 44.4 | 72.1 | 91.7 | 71.2 | |
| LASTBackbone=ViT-B/16, Support Edge Tuning=true, Transfer Overhead (MB/img)=4.04, Trainable Params (M)=0.66, MACs Cloud (x10^9)=17.57, MACs Edge (x10^9)=0.15, GPU Mem Cloud (GB)=0.57, GPU Mem Edge (GB)=0.772024.06 | 76.5 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 80.9 | 86.2 | 62.3 | |
| Bi-LORA (Ours)# param (M)=1.182025.03 | 76.4 | 71.9 | 91.1 | 70.9 | 99.1 | 90.5 | 89.4 | 55.9 | 87.4 | 94.7 | 84.4 | 74.9 | 83.5 | 65.1 | 52.1 | 79.7 | 85.8 | 54.2 | 36.7 | 44.4 | — | — | — | |
| SPT-LoRABackbone=ViT-B/16, Support Edge Tuning=false, Trainable Params (M)=0.54, MACs Cloud (x10^9)=0, MACs Edge (x10^9)=20.44, GPU Mem Cloud (GB)=0, GPU Mem Edge (GB)=5.132024.06 | 76.4 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 81.9 | 85.9 | 61.3 | |
| LORAParams(M)=1.180, Backbone=ViT-B, Pre-training=ImageNet-21K2024.07 | 76.4 | — | — | — | 99.1 | 91.4 | 89.6 | 56 | 87.6 | 95.3 | 84 | 75 | 83.6 | 64.3 | 51.6 | 80.9 | 86 | 51.8 | 36.8 | 42.3 | 72.5 | 91.5 | 71.9 | |
| Res-TuningBackbone=ViT-B/16, # Param (M)=0.552026.06 | 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 | — | — | — | |
| BiT-LPre-training strategy=Generalist, Hyperparameter setting=BiT-HyperRule, Evaluation protocol=Fine-tuning2019.12 | 76.29 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| AdaptFormer# param (M)=0.162025.03 | 76.2 | 72 | 92.7 | 70.2 | 99.3 | 91 | 87.5 | 54.8 | 87.4 | 95.2 | 85.2 | 75.2 | 83.6 | 62.6 | 52.1 | 81 | 86.2 | 53.1 | 34.5 | 40.3 | — | — | — | |
| AdaptFormer (Ours)# param (M)=0.172025.03 | 76.2 | 71.3 | 93.5 | 69.9 | 99.3 | 90.7 | 87 | 54.7 | 87.5 | 95.1 | 84.5 | 76.2 | 83.6 | 63.1 | 52.2 | 81.3 | 87.1 | 54.1 | 33.5 | 40.2 | — | — | — | |
| DA-VPT+Backbone=ViT-B, Pre-training=Supervised on ImageNet-21k, Params (M)=0.242025.05 | 76.14 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 81.98 | 86.47 | 59.96 | |
| RepAdapter# param (M)=0.222025.03 | 76.1 | 72.4 | 91.6 | 71 | 99.2 | 91.4 | 90.7 | 55.1 | 85.3 | 95.9 | 84.6 | 75.9 | 82.3 | 68 | 50.4 | 79.9 | 80.4 | 49.2 | 38.6 | 41 | — | — | — | |
| RepAdapter (Ours)# param (M)=0.232025.03 | 76.1 | 72.1 | 91.5 | 71.8 | 99.3 | 90.7 | 90.3 | 55 | 85.2 | 96.2 | 84.5 | 75.6 | 82.2 | 67.7 | 49.7 | 79.9 | 82.2 | 48.7 | 36.9 | 41.7 | — | — | — | |
| RepAdapterBackbone=ViT-B/16, # Param (M)=0.222026.06 | 76.1 | 72.4 | 91.6 | 71 | 99.2 | 91.4 | 90.7 | 55.1 | 85.3 | 95.9 | 84.6 | 75.9 | 82.3 | 68 | 50.4 | 79.9 | 80.4 | 49.2 | 38.6 | 41 | — | — | — | |
| ARCBackbone=ViT-B/16, # Param (M)=0.132026.06 | 75.8 | 72.2 | 90.1 | 72.7 | 99 | 91 | 91.9 | 54.4 | 84.9 | 95.7 | 86.7 | 75.8 | 80.7 | 67.1 | 48.7 | 81.6 | 79.2 | 51 | 31.4 | 39.9 | — | — | — | |
| SSF# param (M)=0.242025.03 | 75.7 | 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 | — | — | — | |
| LORA# param (M)=0.292025.03 | 75.7 | 70.9 | 93 | 70 | 99.1 | 91.2 | 86 | 55.8 | 87.1 | 94.6 | 82.4 | 74.7 | 83.5 | 63.1 | 51.7 | 79.3 | 85.2 | 51.8 | 33.5 | 43.6 | — | — | — | |
| SSFBackbone=ViT-B/16, # Param (M)=0.242026.06 | 75.7 | 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 | — | — | — | |
| SSFMethod Category=PETuning Methods2023.04 | 75.69 | 69 | 92.6 | 71.5 | 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 | 81.57 | 86.55 | 58.96 | |
| LORA (Ours)# param (M)=0.302025.03 | 75.6 | 70.1 | 92.7 | 69.4 | 99.1 | 90.8 | 85.4 | 55.6 | 87.2 | 94.6 | 82.5 | 74.1 | 83 | 63.4 | 50.6 | 81.6 | 84.7 | 51.5 | 34.3 | 43.3 | — | — | — | |
| FacTBackbone=ViT-B/16, Support Edge Tuning=false, Trainable Params (M)=0.07, MACs Cloud (x10^9)=0, MACs Edge (x10^9)=18.67, GPU Mem Cloud (GB)=0, GPU Mem Edge (GB)=4.812024.06 | 75.6 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 80.6 | 85.3 | 60.7 | |
| FacTBackbone=ViT-B/16, # Param (M)=0.072026.06 | 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 | — | — | — | |
| FacTMethod Category=PETuning Methods2023.04 | 75.55 | 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 | 80.64 | 85.3 | 60.71 | |
| NOAH# param (M)=0.362025.03 | 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 | — | — | — | |
| NOAHBackbone=ViT-B/16, # Param (M)=0.362026.06 | 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 | — | — | — | |
| MORAParams(M)=1.058, Backbone=ViT-B, Pre-training=ImageNet-21K2024.07 | 75.4 | — | — | — | 99.2 | 91.1 | 90.1 | 56 | 87.1 | 94.8 | 85.1 | 75.4 | 76.7 | 62.3 | 49.7 | 78.3 | 83.1 | 53 | 34.5 | 34.5 | 72.1 | 90 | 71.7 | |
| FacT-TT# param (M)=0.042025.03 | 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 | — | — | — | |
| Pruned RepAdapter# param (M)=0.182025.03 | 74.8 | 71.4 | 87.3 | 68.1 | 96 | 89.9 | 89.3 | 53.4 | 85 | 95.3 | 81.9 | 75.2 | 80.9 | 69.8 | 50.5 | 80.7 | 80.5 | 47.1 | 35.7 | 41 | — | — | — | |
| NOAHMethod Category=PETuning Methods2023.04 | 74.8 | 70.7 | 91.6 | 68.2 | 98.9 | 90.2 | 88.4 | 54 | 85.9 | 95.3 | 84.2 | 73.6 | 81.7 | 63.1 | 49 | 78.5 | 82.3 | 45 | 31.8 | 43.5 | 80.29 | 84.75 | 59.36 | |
| AdaptFormerBackbone=ViT-B/16, Support Edge Tuning=false, Trainable Params (M)=0.16, MACs Cloud (x10^9)=0, MACs Edge (x10^9)=17.6, GPU Mem Cloud (GB)=0, GPU Mem Edge (GB)=4.112024.06 | 74.7 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 80.6 | 84.9 | 59 | |
| AdaptFormerBackbone=ViT-B/16, # Param (M)=0.162026.06 | 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 | — | — | — | |
| DA-VPTBackbone=ViT-B, Pre-training=Supervised on ImageNet-21k, Params (M)=0.212025.05 | 74.69 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 80.25 | 85.12 | 58.71 | |
| LoRABackbone=ViT-B/16, Support Edge Tuning=false, Trainable Params (M)=0.29, MACs Cloud (x10^9)=0, MACs Edge (x10^9)=21.87, GPU Mem Cloud (GB)=0, GPU Mem Edge (GB)=3.42024.06 | 74.5 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 79.5 | 84.6 | 59.8 | |
| ConvpassBackbone=ViT-B/16, # Param (M)=0.332026.06 | 74.5 | 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 | — | — | — | |
| LoRABackbone=ViT-B/16, # Param (M)=0.292026.06 | 74.5 | 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 | — | — | — | |
| LSTBackbone=ViT-B/16, Support Edge Tuning=true, Transfer Overhead (MB/img)=7.503, Trainable Params (M)=2.38, MACs Cloud (x10^9)=17.57, MACs Edge (x10^9)=0.51, GPU Mem Cloud (GB)=0.57, GPU Mem Edge (GB)=2.082024.06 | 74.3 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 77.6 | 85.6 | 59.7 | |
| E2VPTBackbone=ViT-B, Pre-training=Supervised on ImageNet-21k, Params (M)=0.332025.05 | 73.94 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 80.01 | 84.43 | 57.39 | |
| E2VPTBackbone=ViT-B/16, # Param (M)=0.252026.06 | 73.9 | 78.6 | 89.4 | 67.8 | 98.2 | 88.5 | 85.3 | 52.3 | 82.5 | 96.8 | 84.8 | 73.6 | 71.7 | 61.2 | 47.9 | 75.8 | 80.8 | 48.1 | 31.7 | 41.9 | — | — | — | |
| AdapterBackbone=ViT-B/16, # Param (M)=0.162026.06 | 73.9 | 69.2 | 90.1 | 68 | 98.8 | 89.9 | 82.8 | 54.3 | 84 | 94.9 | 81.9 | 75.5 | 80.9 | 65.3 | 48.6 | 78.3 | 74.8 | 48.5 | 29.9 | 41.6 | — | — | — | |
| DA-VPT+Backbone=ViT-B, Pre-training=MoCo-v3 on ImageNet-1K, Params (M)=0.242025.05 | 73.53 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 76.86 | 84.71 | 58.98 | |
| SynQT# param (M)=2.722025.03 | 72.9 | 70.9 | 89.7 | 68.8 | 98.5 | 89.6 | 77.8 | 50.6 | 82.3 | 96.7 | 83.5 | 75.2 | 71.8 | 62.7 | 48.5 | 75.4 | 74.1 | 49 | 31.7 | 36.1 | — | — | — | |
| E2VPTBackbone=ViT-B, Pre-training=MoCo-v3 on ImageNet-1K, Params (M)=0.112025.05 | 72.88 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 76.47 | 87.28 | 54.91 | |
| FourierFTParams(M)=0.936, Backbone=ViT-B, Pre-training=ImageNet-21K2024.07 | 72.7 | — | — | — | 99 | 91 | 79 | 55.6 | 84.9 | 93 | 83.2 | 74.9 | 70.7 | 61.1 | 45.2 | 74.8 | 78 | 53 | 24.8 | 30.8 | 69.1 | 88.8 | 71.9 | |
| ARCBackbone=Swin-Base, Pre-trained=ImageNet-21k, Params.=0.272023.10 | 72.6 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 79 | 86.6 | 59.9 | |
| Full fine-tuningBackbone=Swin-Base, Pre-trained=ImageNet-21k, Params.=86.82023.10 | 72.4 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 79.1 | 86.2 | 59.7 | |
| VPT-DEEP# param (M)=0.532025.03 | 72 | 78.8 | 90.8 | 65.8 | 98 | 88.3 | 78.1 | 49.6 | 81.8 | 96.1 | 83.4 | 68.4 | 68.5 | 60 | 46.5 | 72.8 | 73.6 | 47.9 | 32.9 | 37.8 | — | — | — | |
| VPTBackbone=ViT-B/16, Support Edge Tuning=false, Trainable Params (M)=0.56, MACs Cloud (x10^9)=0, MACs Edge (x10^9)=23.18, GPU Mem Cloud (GB)=0, GPU Mem Edge (GB)=5.632024.06 | 72 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 78.5 | 82.4 | 54.9 | |
| VPT-DeepBackbone=ViT-B/16, # Param (M)=0.532026.06 | 72 | 78.8 | 90.8 | 65.8 | 98 | 88.3 | 78.1 | 49.6 | 81.8 | 96.1 | 83.4 | 68.4 | 68.5 | 60 | 46.5 | 72.8 | 73.6 | 47.9 | 32.9 | 37.8 | — | — | — | |
| VPT-DeepMethod Category=PETuning Methods, Variant=Deep2023.04 | 71.96 | 78.8 | 90.8 | 65.8 | 98 | 88.3 | 78.1 | 49.6 | 81.8 | 96.1 | 83.4 | 68.4 | 68.5 | 60 | 46.5 | 72.8 | 73.6 | 47.9 | 32.9 | 37.8 | 78.48 | 82.43 | 54.98 | |
| VPT-DeepBackbone=ViT-B, Pre-training=Supervised on ImageNet-21k, Params (M)=0.642025.05 | 71.96 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 78.48 | 82.43 | 54.98 | |
| DA-VPTBackbone=ViT-B, Pre-training=MoCo-v3 on ImageNet-1K, Params (M)=0.212025.05 | 70.9 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 74.24 | 83.21 | 55.23 | |
| Generalist SOTAPre-training strategy=Generalist2019.12 | 70.5 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| DA-VPT+Backbone=ViT-B, Pre-training=MAE on ImageNet-1K, Params (M)=0.222025.05 | 69.61 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 66.59 | 82.96 | 59.28 | |
| FullBackbone=ViT-B, Pre-training=MoCo-v3 on ImageNet-1K2025.05 | 69.55 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 71.95 | 84.72 | 51.98 | |
| GateVPTBackbone=ViT-B, Pre-training=MoCo-v3 on ImageNet-1K, Params (M)=0.172025.05 | 69.11 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 74.84 | 83.38 | 49.1 | |
| Full TuneMethod Category=Traditional Methods2023.04 | 68.96 | 68.9 | 87.7 | 64.3 | 97.2 | 86.9 | 87.4 | 38.8 | 79.7 | 95.7 | 84.2 | 73.9 | 56.3 | 58.6 | 41.7 | 65.5 | 57.5 | 46.7 | 25.7 | 29.1 | 75.88 | 83.36 | 47.64 | |
| Full fine-tuningBackbone=ViT-B, Pre-training=Supervised on ImageNet-21k2025.05 | 68.96 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 75.88 | 83.36 | 47.64 | |
| Full# param (M)=85.82025.03 | 68.9 | 68.9 | 87.7 | 64.3 | 97.2 | 86.9 | 87.4 | 38.8 | 79.7 | 95.7 | 84.2 | 73.9 | 56.3 | 58.6 | 41.7 | 65.5 | 57.5 | 46.7 | 25.7 | 29.1 | — | — | — | |
| Full fine-tuningBackbone=ViT-B/16, Support Edge Tuning=false, Trainable Params (M)=85.8, MACs Cloud (x10^9)=0, MACs Edge (x10^9)=17.57, GPU Mem Cloud (GB)=0, GPU Mem Edge (GB)=6.092024.06 | 68.9 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 75.9 | 83.4 | 47.6 | |
| FFTParams(M)=86, Backbone=ViT-B, Pre-training=ImageNet-21K2024.07 | 68.9 | — | — | — | 97.2 | 86.9 | 87.4 | 38.8 | 79.7 | 95.7 | 84.2 | 73.9 | 56.3 | 58.6 | 41.7 | 65.5 | 57.5 | 46.7 | 25.7 | 29.1 | 68.9 | 87.7 | 64.3 | |
| Full fine-tuningBackbone=ViT-B/16, # Param (M)=85.82026.06 | 68.9 | 68.9 | 87.7 | 64.3 | 97.2 | 86.9 | 87.4 | 38.8 | 79.7 | 95.7 | 84.2 | 73.9 | 56.3 | 58.6 | 41.7 | 65.5 | 57.5 | 46.7 | 25.7 | 29.1 | — | — | — | |
| VPT-ShallowMethod Category=PETuning Methods, Variant=Shallow2023.04 | 67.82 | 77.7 | 86.9 | 62.6 | 97.5 | 87.3 | 74.5 | 51.2 | 78.2 | 92 | 75.6 | 72.9 | 55.5 | 58.6 | 40.5 | 67.1 | 68.7 | 36.1 | 20.2 | 34.1 | 76.81 | 79.66 | 46.98 | |
| VPT-ShallowBackbone=ViT-B, Pre-training=Supervised on ImageNet-21k, Params (M)=0.112025.05 | 67.82 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 76.81 | 79.68 | 46.98 | |
| VPT-Shallow# param (M)=0.062025.03 | 67.8 | 77.7 | 86.9 | 62.6 | 97.5 | 87.3 | 74.5 | 51.2 | 78.2 | 92 | 75.6 | 72.9 | 50.5 | 58.6 | 40.5 | 67.1 | 68.7 | 36.1 | 20.2 | 34.1 | — | — | — | |
| VPT-ShallowBackbone=ViT-B/16, # Param (M)=0.062026.06 | 67.8 | 77.7 | 86.9 | 62.6 | 97.5 | 87.3 | 74.5 | 51.2 | 78.2 | 92 | 75.6 | 72.9 | 50.5 | 58.6 | 40.5 | 67.1 | 68.7 | 36.1 | 20.2 | 34.1 | — | — | — | |
| VPT-DeepBackbone=Swin-Base, Pre-trained=ImageNet-21k, Params.=0.222023.10 | 67.7 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 76.8 | 84.5 | 53.4 | |
| VPT-DeepBackbone=ViT-B, Pre-training=MoCo-v3 on ImageNet-1K, Params (M)=0.22025.05 | 65.9 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 70.27 | 83.04 | 42.38 | |
| BitFit# param (M)=0.102025.03 | 65.2 | 72.8 | 87 | 59.2 | 97.5 | 85.3 | 59.9 | 51.4 | 78.7 | 91.6 | 72.9 | 69.8 | 61.5 | 55.6 | 32.4 | 55.9 | 66.6 | 40 | 15.7 | 25.1 | — | — | — | |
| BitFitBackbone=ViT-B/16, Support Edge Tuning=false, Trainable Params (M)=0.1, MACs Cloud (x10^9)=0, MACs Edge (x10^9)=17.57, GPU Mem Cloud (GB)=0, GPU Mem Edge (GB)=3.82024.06 | 65.2 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 73.3 | 78.2 | 44.1 | |
| BitFitBackbone=ViT-B/16, # Param (M)=0.102026.06 | 65.2 | 72.8 | 87 | 59.2 | 97.5 | 85.3 | 59.9 | 51.4 | 78.7 | 91.6 | 72.9 | 69.8 | 61.5 | 55.6 | 32.4 | 55.9 | 66.6 | 40 | 15.7 | 25.1 | — | — | — | |
| DA-VPTBackbone=ViT-B, Pre-training=MAE on ImageNet-1K, Params (M)=0.22025.05 | 65.19 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 62.14 | 79.14 | 54.31 | |
| FullBackbone=ViT-B, Pre-training=MAE on ImageNet-1K2025.05 | 64.27 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 59.31 | 79.68 | 53.82 | |
| VPT-ShallowBackbone=Swin-Base, Pre-trained=ImageNet-21k, Params.=0.052023.10 | 62.9 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 79.9 | 82.5 | 37.8 | |
| VPT-ShallowBackbone=ViT-B, Pre-training=MoCo-v3 on ImageNet-1K, Params (M)=0.112025.05 | 62.38 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 67.34 | 82.26 | 37.55 | |
| BiasBackbone=Swin-Base, Pre-trained=ImageNet-21k, Params.=0.252023.10 | 62.1 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 74.2 | 80.1 | 42.4 | |
| E2VPTBackbone=ViT-B, Pre-training=MAE on ImageNet-1K, Params (M)=0.062025.05 | 61.04 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 77.8 | 44.65 | 60.66 | |
| Partial-1Backbone=ViT-B/16, Support Edge Tuning=true, Transfer Overhead (MB/img)=0.577, Trainable Params (M)=6.75, MACs Cloud (x10^9)=17.57, MACs Edge (x10^9)=1.42, GPU Mem Cloud (GB)=0.57, GPU Mem Edge (GB)=0.522024.06 | 60.7 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 69.4 | 78.5 | 34.2 | |
| PartialBackbone=Swin-Base, Pre-trained=ImageNet-21k, Params.=12.652023.10 | 58.9 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 73.1 | 81.7 | 35 | |
| Linear probingBackbone=Swin-Base, Pre-trained=ImageNet-21k, Params.=0.052023.10 | 58.2 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 73.5 | 80.8 | 33.5 | |
| MLP-4Backbone=Swin-Base, Pre-trained=ImageNet-21k, Params.=4.042023.10 | 57.7 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 70.6 | 80.7 | 31.2 | |
| Linear ProbeMethod Category=Traditional Methods2023.04 | 57.64 | 63.4 | 85 | 63.2 | 97 | 86.3 | 36.6 | 51 | 78.5 | 87.5 | 68.6 | 74 | 34.3 | 30.6 | 33.2 | 55.4 | 12.5 | 20 | 9.6 | 19.2 | 68.93 | 77.16 | 26.84 | |
| Linear# param (M)=02025.03 | 57.6 | 64.4 | 85 | 63.2 | 97 | 86.3 | 36.6 | 51 | 78.5 | 87.5 | 68.5 | 74 | 34.3 | 30.6 | 33.2 | 55.4 | 12.5 | 20 | 9.6 | 19.2 | — | — | — | |
| Linear probingBackbone=ViT-B/16, Support Edge Tuning=true, Transfer Overhead (MB/img)=0.003, Trainable Params (M)=0, MACs Cloud (x10^9)=17.57, MACs Edge (x10^9)=0, GPU Mem Cloud (GB)=0.57, GPU Mem Edge (GB)=0.072024.06 | 57.6 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 69.1 | 77.1 | 26.8 | |
| LPParams(M)=0, Backbone=ViT-B, Pre-training=ImageNet-21K2024.07 | 57.6 | — | — | — | 97 | 86.3 | 36.6 | 51 | 78.5 | 87.5 | 68.5 | 74 | 34.3 | 30.6 | 33.2 | 55.4 | 12.5 | 20 | 9.6 | 19.2 | 64.4 | 85 | 63.2 | |
| Linear probingBackbone=ViT-B/16, # Param (M)=02026.06 | 57.6 | 64.4 | 85 | 63.2 | 97 | 86.3 | 36.6 | 51 | 78.5 | 87.5 | 68.5 | 74 | 34.3 | 30.6 | 33.2 | 55.4 | 12.5 | 20 | 9.6 | 19.2 | — | — | — | |
| GateVPTBackbone=ViT-B, Pre-training=MAE on ImageNet-1K, Params (M)=0.172025.05 | 53.09 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 47.61 | 76.86 | 36.8 | |
| VPT-ShallowBackbone=ViT-B, Pre-training=MAE on ImageNet-1K, Params (M)=0.12025.05 | 45.7 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 39.96 | 69.65 | 27.5 | |
| VPT-DeepBackbone=ViT-B, Pre-training=MAE on ImageNet-1K, Params (M)=0.22025.05 | 41.73 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 36.02 | 60.61 | 26.57 | |
| Auto-PromptingBackbone=Swin-Base, Tuned=0.262026.06 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 85.52 | 86.49 | 62.36 | |
| BiasBackbone=Swin-Base, Tuned=0.302026.06 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 76.78 | 83.33 | 51.85 | |
| E2VPTBackbone=Swin-Base, Tuned=0.212026.06 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 83.31 | 84.95 | 57.35 | |
| FullBackbone=ViT-B, Pre-training Method=MAE, Pre-training Dataset=ImageNet-1K, Tuning Depth=Full2024.02 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 59.31 | 79.68 | 53.82 | |
| FullBackbone=ViT-B, Pre-training Method=MoCo-v3, Pre-training Dataset=ImageNet-1K, Tuning Depth=Full2024.02 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 71.95 | 84.72 | 51.98 |