Domain Generalization on DomainBed (test)
84.3VLCS AccuracyCLIP + DPL
Evaluation results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| CLIP + DPLcategory=Prompt learning, Backbone=CLIP ViT-B162021.11 | 84.3 | — | — | 97.3 | 84.2 | 52.6 | — | 79.6 | |
| DANNcategory=Fine-tuning, Backbone=CLIP ViT-B162021.11 | 83.2 | — | — | 93.8 | 78.8 | 52.2 | — | 77 | |
| ERM (CLIP)category=Fine-tuning, Backbone=CLIP ViT-B162021.11 | 82.7 | — | — | 92.9 | 78.1 | 50.2 | — | 75.9 | |
| CLIP (template prompt)category=Zero-shot, Backbone=CLIP ViT-B162021.11 | 82.3 | — | — | 96.1 | 82.3 | 34.1 | — | 73.7 | |
| CORALcategory=Fine-tuning, Backbone=CLIP ViT-B162021.11 | 82 | — | — | 93.2 | 78.9 | 53.5 | — | 76.9 | |
| EoABackbone=RegNetY-16GF, Parameters=81M, Pre-training=3.6B Images2021.10 | 81.1 | — | — | 95.8 | 83.9 | 61.1 | 60.9 | 76.6 | |
| EoABackbone=RegNetY-16GF, Parameters=81M, Pre-training=Instagram 3.6B Images2021.10 | 81.1 | — | — | 95.8 | 83.9 | 61.1 | 60.9 | 76.6 | |
| SMABackbone=RegNetY-16GF, Parameters=81M, Pre-training=3.6B Images2021.10 | 80.7 | — | — | 95.5 | 82 | 59.7 | 60 | 75.6 | |
| SMABackbone=RegNetY-16GF, Parameters=81M, Pre-training=Instagram 3.6B Images2021.10 | 80.7 | — | — | 95.5 | 82 | 59.7 | 60 | 75.6 | |
| Ensemble (our runs)Backbone=RegNetY-16GF, Parameters=81M, Pre-training=3.6B Images2021.10 | 80.6 | — | — | 95.1 | 80.5 | 59.5 | 57.8 | 74.7 | |
| EnsembleBackbone=RegNetY-16GF, Parameters=81M, Pre-training=Instagram 3.6B Images, Source=Our runs2021.10 | 80.6 | — | — | 95.1 | 80.5 | 59.5 | 57.8 | 74.7 | |
| EoABackbone=ResNeXt-50 32x4d, Parameters=25M, Pre-training=1B Images2021.10 | 80.4 | — | — | 93.2 | 80.2 | 55.2 | 54.6 | 72.7 | |
| EoABackbone=ResNeXt-50 32x4d, Parameters=25M, Pre-training=Instagram 1B Images2021.10 | 80.4 | — | — | 93.2 | 80.2 | 55.2 | 54.6 | 72.7 | |
| Ensemble (our runs)Backbone=ResNeXt-50 32x4d, Parameters=25M, Pre-training=1B Images2021.10 | 80.3 | — | — | 91.2 | 77.8 | 53.5 | 52.8 | 71.1 | |
| EnsembleBackbone=ResNeXt-50 32x4d, Parameters=25M, Pre-training=Instagram 1B Images, Source=Our runs2021.10 | 80.3 | — | — | 91.2 | 77.8 | 53.5 | 52.8 | 71.1 | |
| EVIL-SAMFLOPs=0.89×2025.12 | 80.1 | 40.4 | 98.8 | 87.8 | 70.3 | 50.5 | 45 | 67.5 | |
| MIROBackbone=RegNetY-16GF, Parameters=81M, Pre-training=3.6B Images2021.10 | 79.9 | — | — | 97.4 | 80.4 | 58.9 | 53.8 | 74.1 | |
| MIROBackbone=RegNetY-16GF, Parameters=81M, Pre-training=Instagram 3.6B Images2021.10 | 79.9 | — | — | 97.4 | 80.4 | 58.9 | 53.8 | 74.1 | |
| FISHSDG=true2023.08 | 79.9 | 51.9 | 98 | 87.3 | 68.4 | 48.4 | 42.6 | 68.1 | |
| SMABackbone=ResNeXt-50 32x4d, Parameters=25M, Pre-training=1B Images2021.10 | 79.7 | — | — | 92.7 | 78.6 | 53.3 | 53.5 | 71.6 | |
| SWADBackbone=RegNetY-16GF, Parameters=81M, Pre-training=3.6B Images2021.10 | 79.7 | — | — | 94.7 | 80 | 57.9 | 53.6 | 73.2 | |
| SMABackbone=ResNeXt-50 32x4d, Parameters=25M, Pre-training=Instagram 1B Images2021.10 | 79.7 | — | — | 92.7 | 78.6 | 53.3 | 53.5 | 71.6 | |
| SWADBackbone=RegNetY-16GF, Parameters=81M, Pre-training=Instagram 3.6B Images2021.10 | 79.7 | — | — | 94.7 | 80 | 57.9 | 53.6 | 73.2 | |
| CORALSDG=true2023.08 | 79.7 | 52.5 | 98.2 | 86.9 | 68.7 | 49.9 | 41.4 | 68.2 | |
| EVIL-MIROFLOPs=0.45×2025.12 | 79.4 | 40.2 | 98.6 | 85.8 | 71.2 | 50.9 | 45 | 67.3 | |
| SAMFLOPs=2×2025.12 | 79.4 | 38.5 | 98.1 | 85.8 | 69.6 | 43.3 | 44.3 | 65.6 | |
| EVIL-SWADFLOPs=0.43×2025.12 | 79.3 | 38.7 | 98.3 | 88.3 | 71.7 | 51.2 | 46.9 | 67.7 | |
| EVIL-CORALFLOPs=0.43×2025.12 | 79.2 | 34.5 | 98.6 | 86.9 | 69 | 49.2 | 42.6 | 65.7 | |
| SWADBackbone=ResNet-50, Parameters=25M, Pre-training=ImageNet2021.10 | 79.1 | — | — | 88.1 | 70.6 | 50 | 46.5 | 66.9 | |
| EoABackbone=ResNet-50, Parameters=25M, Pre-training=ImageNet2021.10 | 79.1 | — | — | 88.6 | 72.5 | 52.3 | 47.4 | 68 | |
| SWADBackbone=ResNet-50, Parameters=25M, Pre-training=ImageNet2021.10 | 79.1 | — | — | 88.1 | 70.6 | 50 | 46.5 | 66.9 | |
| EoABackbone=ResNet-50, Parameters=25M, Pre-training=ImageNet2021.10 | 79.1 | — | — | 88.6 | 72.5 | 52.3 | 47.4 | 68 | |
| EVIL-RExFLOPs=0.42×2025.12 | 79.1 | 41.2 | 98.7 | 86 | 68 | 48.4 | 34.5 | 65.1 | |
| SWADFLOPs=1×2025.12 | 79.1 | 38.3 | 98.1 | 88.1 | 70.6 | 50 | 46.5 | 67.2 | |
| LPABackbone=ResNet-50, Number of random seeds=32026.05 | 79.1 | — | — | 87.8 | 69.5 | 49 | — | — | |
| MIROBackbone=ResNet-50, Parameters=25M, Pre-training=ImageNet2021.10 | 79 | — | — | 85.4 | 70.5 | 50.4 | 44.3 | 65.9 | |
| ERM (our runs)Backbone=ResNeXt-50 32x4d, Parameters=25M, Pre-training=1B Images2021.10 | 79 | — | — | 88.9 | 70.9 | 51.4 | 48.1 | 67.7 | |
| MIROBackbone=ResNet-50, Parameters=25M, Pre-training=ImageNet2021.10 | 79 | — | — | 85.4 | 70.5 | 50.4 | 44.3 | 65.9 | |
| ERMBackbone=ResNeXt-50 32x4d, Parameters=25M, Pre-training=Instagram 1B Images, Source=Our runs2021.10 | 79 | — | — | 88.9 | 70.9 | 51.4 | 48.1 | 67.7 | |
| MIROFLOPs=1×2025.12 | 79 | 39.4 | 97.5 | 85.4 | 70.5 | 50.4 | 44.3 | 66.6 | |
| VRexSDG=true2023.08 | 78.9 | 52.7 | 98.1 | 86.4 | 68.2 | 48.1 | 41.8 | 67.7 | |
| CORALlearning domain-invariant features=false, Adversarial learning strategy=false, inter-domain mix-up=false, contrastive learning=false, unsupervised domain adaptation=true, model selection method=domain validation set2021.04 | 78.8 | 51.5 | 98 | 86.2 | 68.7 | 47.6 | 41.5 | 67.5 | |
| CORALSDG=false2023.08 | 78.8 | 51.5 | 98 | 86.2 | 68.7 | 47.6 | 41.5 | 67.5 | |
| EVILFLOPs=0.42×2025.12 | 78.8 | 39.4 | 98.4 | 86 | 68.2 | 49.1 | 43.8 | 66.2 | |
| EVIL-IRMFLOPs=0.42×2025.12 | 78.8 | 39.1 | 98.3 | 85.1 | 66.4 | 48.3 | 36 | 64.6 | |
| CORALFLOPs=1×2025.12 | 78.8 | 29.9 | 98.1 | 86.2 | 68.7 | 47.7 | 41.5 | 64.4 | |
| DANN2021.04 | 78.7 | 51.5 | 97.9 | 84.6 | 65.4 | 48.7 | 38.4 | 66.4 | |
| DANNBackbone=ResNet-50, Parameters=25M, Pre-training=ImageNet2021.10 | 78.7 | — | — | 84.6 | 65.4 | 48.4 | 38.4 | 63.1 | |
| DANNlearning domain-invariant features=true, Adversarial learning strategy=true, inter-domain mix-up=false, contrastive learning=false, unsupervised domain adaptation=true, model selection method=domain validation set2021.04 | 78.6 | 51.5 | 97.8 | 83.6 | 65.9 | 46.7 | 38.3 | 66.1 | |
| ERM (our runs)Backbone=RegNetY-16GF, Parameters=81M, Pre-training=3.6B Images2021.10 | 78.6 | — | — | 92 | 73.8 | 55.6 | 53.2 | 70.6 | |
| ERMBackbone=RegNetY-16GF, Parameters=81M, Pre-training=3.6B Images2021.10 | 78.6 | — | — | 89.6 | 71.9 | 51.4 | 48.5 | 68 | |
| ERMBackbone=RegNetY-16GF, Parameters=81M, Pre-training=Instagram 3.6B Images, Source=Our runs2021.10 | 78.6 | — | — | 92 | 73.8 | 55.6 | 53.1 | 70.6 | |
| ERMBackbone=RegNetY-16GF, Parameters=81M, Pre-training=Instagram 3.6B Images2021.10 | 78.6 | — | — | 89.6 | 71.9 | 51.4 | 48.5 | 68 | |
| DANN2025.12 | 78.6 | 29 | 89.1 | 83.7 | 65.9 | 46.7 | 38.3 | 61.6 | |
| IRMFLOPs=1×2025.12 | 78.6 | 36.3 | 97.7 | 83.5 | 64.3 | 47.6 | 33.9 | 63.1 | |
| IRMlearning domain-invariant features=true, Adversarial learning strategy=false, inter-domain mix-up=false, contrastive learning=false, unsupervised domain adaptation=false, model selection method=domain validation set2021.04 | 78.5 | 52 | 97.7 | 83.5 | 64.3 | 47.6 | 33.9 | 65.4 | |
| Ensemble (our runs)Backbone=ResNet-50, Parameters=25M, Pre-training=ImageNet2021.10 | 78.5 | — | — | 87.6 | 70.8 | 49.2 | 47.7 | 66.8 | |
| EnsembleBackbone=ResNet-50, Parameters=25M, Pre-training=ImageNet, Source=Our runs2021.10 | 78.5 | — | — | 87.6 | 70.8 | 49.2 | 47.7 | 66.8 | |
| SWADBackbone=ResNet-50, Number of random seeds=32026.05 | 78.5 | — | — | 87.2 | 68.8 | 48.2 | — | — | |
| VREXlearning domain-invariant features=true, Adversarial learning strategy=false, inter-domain mix-up=false, contrastive learning=false, unsupervised domain adaptation=false, model selection method=domain validation set2021.04 | 78.3 | 51.8 | 97.9 | 84.9 | 66.4 | 46.4 | 33.6 | 65.6 | |
| VRexSDG=false2023.08 | 78.3 | 51.8 | 97.9 | 84.9 | 66.4 | 46.4 | 33.6 | 65.6 | |
| RExFLOPs=1×2025.12 | 78.3 | 39.2 | 97.3 | 84.9 | 66.4 | 46.4 | 33.6 | 63.7 | |
| CDANN2021.04 | 78.2 | 51.9 | 98 | 82.8 | 65.6 | 47.6 | 38.9 | 66.1 | |
| C-DANNBackbone=ResNet-50, Parameters=25M, Pre-training=ImageNet2021.10 | 78.2 | — | — | 82.8 | 65.6 | 47.6 | 38.9 | 62.6 | |
| SMABackbone=ResNet-50, Parameters=25M, Pre-training=ImageNet2021.10 | 78.2 | — | — | 87.5 | 70.6 | 50.3 | 46 | 66.5 | |
| SMABackbone=ResNet-50, Parameters=25M, Pre-training=ImageNet2021.10 | 78.2 | — | — | 87.5 | 70.6 | 50.3 | 46.4 | 66.5 | |
| IRM2021.04 | 78.1 | 51.8 | 97.9 | 84.4 | 66.6 | 47.9 | 35.7 | 66 | |
| IRMBackbone=ResNet-50, Parameters=25M, Pre-training=ImageNet2021.10 | 78.1 | — | — | 84.4 | 66.6 | 47.9 | 35.7 | 62.5 | |
| Mixstyle2025.12 | 77.9 | 38.5 | 97.2 | 85.2 | 60.4 | 44 | 34 | 62.4 | |
| Manifold MixupBackbone=ResNet-50, Number of random seeds=32026.05 | 77.9 | — | — | 86.5 | 67.5 | 47.1 | — | — | |
| SelfReglearning domain-invariant features=false, Adversarial learning strategy=false, inter-domain mix-up=true, contrastive learning=true, unsupervised domain adaptation=false, use IDCL=false, use SWA=false, model selection method=domain validation set2021.04 | 77.8 | 52.1 | 98 | 85.6 | 67.9 | 47 | 42.8 | 67.3 | |
| SagNetlearning domain-invariant features=true, Adversarial learning strategy=true, inter-domain mix-up=false, contrastive learning=false, unsupervised domain adaptation=false, model selection method=domain validation set2021.04 | 77.8 | 51.7 | 98 | 86.3 | 68.1 | 48.6 | 40.3 | 67.2 | |
| Fish2021.04 | 77.8 | 51.6 | 98 | 85.5 | 68.6 | 45.1 | 42.7 | 67.1 | |
| FishBackbone=ResNet-50, Parameters=25M, Pre-training=ImageNet2021.10 | 77.8 | — | — | 85.5 | 68.6 | 45.1 | 42.7 | 63.9 | |
| FishrBackbone=ResNet-50, Parameters=25M, Pre-training=ImageNet2021.10 | 77.8 | — | — | 85.5 | 67.8 | 47.4 | 41.7 | 65.7 | |
| FISHSDG=false2023.08 | 77.8 | 51.6 | 98 | 85.5 | 68.6 | 45.1 | 42.7 | 67.1 | |
| SagNet2025.12 | 77.8 | 34.2 | 96.4 | 86.3 | 68.1 | 48.6 | 40.3 | 64.5 | |
| CORALBackbone=ResNet-50, Number of random seeds=32026.05 | 77.8 | — | — | 86 | 66.8 | 46.5 | — | — | |
| Mixup2021.04 | 77.7 | 51.9 | 98.1 | 84.3 | 69 | 48.9 | 39.6 | 67.1 | |
| Coral2021.04 | 77.7 | 51.7 | 98.1 | 86 | 68.6 | 46.4 | 41.8 | 67.2 | |
| MixupBackbone=ResNet-50, Parameters=25M, Pre-training=ImageNet2021.10 | 77.7 | — | — | 84.3 | 69 | 48.9 | 39.6 | 63.9 | |
| CORALBackbone=ResNet-50, Parameters=25M, Pre-training=ImageNet2021.10 | 77.7 | — | — | 86 | 68.6 | 46.4 | 41.8 | 64.1 | |
| EVIL-DROFLOPs=0.42×2025.12 | 77.7 | 34.2 | 98.2 | 85.6 | 66.4 | 49.1 | 35.5 | 63.8 | |
| ARMlearning domain-invariant features=false, Adversarial learning strategy=false, inter-domain mix-up=false, contrastive learning=false, unsupervised domain adaptation=false, model selection method=domain validation set2021.04 | 77.6 | 56.2 | 98.2 | 85.1 | 64.8 | 45.5 | 35.5 | 66.1 | |
| ARM2025.12 | 77.6 | 32.6 | 98.1 | 85.1 | 64.8 | 45.5 | 35.5 | 62.5 | |
| ERMlearning domain-invariant features=false, Adversarial learning strategy=false, inter-domain mix-up=false, contrastive learning=false, unsupervised domain adaptation=false, model selection method=domain validation set2021.04 | 77.5 | 51.5 | 98 | 85.5 | 66.5 | 46.1 | 40.9 | 66.6 | |
| CDANNlearning domain-invariant features=true, Adversarial learning strategy=true, inter-domain mix-up=false, contrastive learning=false, unsupervised domain adaptation=false, model selection method=domain validation set2021.04 | 77.5 | 51.7 | 97.9 | 82.6 | 65.8 | 45.8 | 38.3 | 65.6 | |
| MMDlearning domain-invariant features=true, Adversarial learning strategy=false, inter-domain mix-up=false, contrastive learning=false, unsupervised domain adaptation=false, model selection method=domain validation set2021.04 | 77.5 | 51.5 | 97.9 | 84.6 | 66.3 | 42.2 | 23.4 | 63.3 | |
| ERMSDG=false2023.08 | 77.5 | 51.5 | 98 | 85.5 | 66.5 | 46.1 | 40.9 | 66.6 | |
| MMD2025.12 | 77.5 | 42.6 | 98.1 | 84.7 | 66.4 | 42.2 | 23.4 | 62.1 | |
| CDANN2025.12 | 77.5 | 31.1 | 96.3 | 82.6 | 65.7 | 45.8 | 38.3 | 62.5 | |
| MixupBackbone=ResNet-50, Number of random seeds=32026.05 | 77.5 | — | — | 86.1 | 67.2 | 46.8 | — | — | |
| Mixuplearning domain-invariant features=false, Adversarial learning strategy=false, inter-domain mix-up=true, contrastive learning=false, unsupervised domain adaptation=false, model selection method=domain validation set2021.04 | 77.4 | 52.1 | 98 | 84.6 | 68.1 | 47.9 | 39.2 | 66.7 | |
| ERM2021.04 | 77.4 | 52 | 98 | 85.7 | 67.5 | 47.2 | 41.2 | 67 | |
| ERMBackbone=ResNet-50, Parameters=25M, Pre-training=ImageNet2021.10 | 77.4 | — | — | 85.7 | 67.5 | 47.2 | 41.2 | 63.8 | |
| ERMBackbone=ResNet-50, Parameters=25M, Pre-training=ImageNet2021.10 | 77.4 | — | — | 85.7 | 67.5 | 47.2 | 41.2 | 63.8 | |
| I-Mixup2025.12 | 77.4 | 31.3 | 97.8 | 84.6 | 68.1 | 47.9 | 39.2 | 63.7 | |
| ERMBackbone=ResNet-50, Number of random seeds=32026.05 | 77.3 | — | — | 85.5 | 66.5 | 46.2 | — | — | |
| MLDGlearning domain-invariant features=false, Adversarial learning strategy=false, inter-domain mix-up=false, contrastive learning=false, unsupervised domain adaptation=false, model selection method=domain validation set2021.04 | 77.2 | 51.5 | 97.9 | 84.9 | 66.8 | 47.7 | 41.2 | 66.7 | |
| MTLlearning domain-invariant features=false, Adversarial learning strategy=false, inter-domain mix-up=false, contrastive learning=false, unsupervised domain adaptation=false, model selection method=domain validation set2021.04 | 77.2 | 51.4 | 97.9 | 84.6 | 66.4 | 45.6 | 40.6 | 66.2 |