Domain Generalization on PACS (test)
99.7Average AccuracyText-Anchored Information Bottleneck
Evaluation results
| Method | Links | ||||||||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Text-Anchored Information BottleneckBackbone=RegNetY-16GF, Pre-training source=IG-3B2026.07 | 99.7 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Text-Anchored Information BottleneckBackbone=RegNetY-16GF, Pre-training source=IN-12K2026.07 | 99.4 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| CLIPoodBackbone=RegNetY-16GF, Pre-training source=IG-3B2026.07 | 97.8 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Text-Anchored Information BottleneckBackbone=Swin Transformer, Pre-training source=IN-21K2026.07 | 97.7 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Text-Anchored Information BottleneckBackbone=DINOv2, Pre-training source=LVD-142M2026.07 | 97.7 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| DAMPEvaluation Protocol=Prompt Learning, Backbone=ViT-B/162024.03 | 97.4 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| MIROBackbone=RegNetY-16GF, Pre-training source=IG-3B2026.07 | 97.4 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| DPLEvaluation Protocol=Prompt Learning, Backbone=ViT-B/162024.03 | 97.3 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| MetaPromptFew-shot=5-shot, CLIP=true2022.12 | 96.9 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| CLIPoodBackbone=DINOv2, Pre-training source=LVD-142M2026.07 | 96.8 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| CoCoOpFew-shot=5-shot, CLIP=true2022.12 | 96.7 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| MetaPromptFew-shot=1-shot, CLIP=true2022.12 | 96.7 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| VL2VBackbone=RegNetY-16GF, Pre-training source=IG-3B2026.07 | 96.7 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| SEDGEEnsemble=true, CLIP=true2022.12 | 96.1 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| CLIP (template)Zero-shot=true, CLIP=true2022.12 | 96.1 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| CLIP (Teacher)Backbone=ViT B/16, Ens/MA=No, Inference Mode=Zero-shot2023.09 | 96.1 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| CLIPEvaluation Protocol=Zero-shot, Backbone=ViT-B/162024.03 | 96.1 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| CLIP + DPLBackbone=CLIP ViT-B16, Test-time batch size=642021.11 | 95.9 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| EoAEnsemble=true, CLIP=true2022.12 | 95.8 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Text-Anchored Information BottleneckBackbone=DeiT, Pre-training source=IN-1K2026.07 | 95.8 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| LPBackbone=DINOv2, Pre-training source=LVD-142M2026.07 | 95.8 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| CLIPZero-shot=true, CLIP=true2022.12 | 95.7 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| RISEBackbone=RegNetY-16GF, Pre-training source=IG-3B2026.07 | 95.5 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| MIROBackbone=DINOv2, Pre-training source=LVD-142M2026.07 | 95.3 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| VL2VBackbone=DINOv2, Pre-training source=LVD-142M2026.07 | 95.1 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| RISEBackbone=Swin Transformer, Pre-training source=IN-21K2026.07 | 94.2 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| CLIPoodBackbone=RegNetY-16GF, Pre-training source=IN-12K2026.07 | 94.1 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| ERM + SHOTBackbone=CLIP ViT-B16, Test-time batch size=64, Test-time adaptation method=SHOT2021.11 | 93.8 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| DANNEvaluation Protocol=Fine-tuning (CLIP), Backbone=ViT-B/162024.03 | 93.8 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| MIROBackbone=Swin Transformer, Pre-training source=IN-21K2026.07 | 93.5 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| LPBackbone=Swin Transformer, Pre-training source=IN-21K2026.07 | 93.4 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| CORALEvaluation Protocol=Fine-tuning (CLIP), Backbone=ViT-B/162024.03 | 93.2 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| CORAL + Pseudo LabelBackbone=CLIP ViT-B16, Test-time batch size=64, Test-time adaptation method=Pseudo Label2021.11 | 93.1 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| CORAL + SHOT**Backbone=CLIP ViT-B16, Test-time batch size=64, Test-time adaptation method=SHOT, Update target=feature extractor2021.11 | 93.1 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| VL2VBackbone=Swin Transformer, Pre-training source=IN-21K2026.07 | 93.1 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| VL2VBackbone=RegNetY-16GF, Pre-training source=IN-12K2026.07 | 93.1 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| CORAL + SHOTBackbone=CLIP ViT-B16, Test-time batch size=64, Test-time adaptation method=SHOT2021.11 | 93 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| ERMEvaluation Protocol=Fine-tuning (CLIP), Backbone=ViT-B/162024.03 | 92.9 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| LPBackbone=RegNetY-16GF, Pre-training source=IG-3B2026.07 | 92.4 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Text-Anchored Information BottleneckBackbone=EfficientNet, Pre-training source=IN-1K2026.07 | 92.3 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| ERM + Pseudo LabelBackbone=CLIP ViT-B16, Test-time batch size=64, Test-time adaptation method=Pseudo Label2021.11 | 92.1 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| ERMBackbone=CLIP ViT-B16, Test-time batch size=642021.11 | 91.9 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| ERM + SHOT**Backbone=CLIP ViT-B16, Test-time batch size=64, Test-time adaptation method=SHOT, Update target=feature extractor2021.11 | 91.2 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| CORAL + T3ABackbone=CLIP ViT-B16, Test-time batch size=64, Test-time adaptation method=T3A2021.11 | 91.2 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| CORAL + Pseudo Label*Backbone=CLIP ViT-B16, Test-time batch size=64, Test-time adaptation method=Pseudo Label, Update target=linear classifier2021.11 | 91.2 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| CORAL + Tent*Backbone=CLIP ViT-B16, Test-time batch size=64, Test-time adaptation method=Tent, Update target=linear classifier2021.11 | 91.2 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| CLIPoodBackbone=EfficientNet, Pre-training source=IN-1K2026.07 | 91.2 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| CORALBackbone=CLIP ViT-B16, Test-time batch size=642021.11 | 91.1 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| VL2VBackbone=DeiT, Pre-training source=IN-1K2026.07 | 90.8 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| CLIPoodBackbone=DeiT, Pre-training source=IN-1K2026.07 | 90.6 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| RISEBackbone=DeiT, Pre-training source=IN-1K2026.07 | 90.6 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| RISE (ERM + Hint + AD + RD + MT)Backbone=ResNet50, Ens/MA=Yes2023.09 | 90.2 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| RISE (ERM + Hint + AD + RD)Backbone=ResNet50, Ens/MA=No2023.09 | 89.4 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| LPBackbone=DeiT, Pre-training source=IN-1K2026.07 | 89.4 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| GeoTexAugBackbone=ResNet-502023.04 | 89.17 | 89.98 | 83.84 | 98.1 | 84.75 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| RISEBackbone=RegNetY-16GF, Pre-training source=IN-12K2026.07 | 89.1 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| ERM + Hint + ADBackbone=ResNet50, Ens/MA=No2023.09 | 89 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| ERM + Hint + RDBackbone=ResNet50, Ens/MA=No2023.09 | 88.8 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| EoABackbone=ResNet50, Ens/MA=Yes2023.09 | 88.6 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| RICEBackbone=ResNet502022.03 | 88.4 | 87.8 | 84.3 | 96.8 | 84.7 | 84.3 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| ERM + HintBackbone=ResNet50, Ens/MA=No2023.09 | 88.4 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| RISEBackbone=DINOv2, Pre-training source=LVD-142M2026.07 | 88.3 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Pro-RandConvBackbone=ResNet-502023.04 | 88.27 | 89.28 | 84.13 | 97.83 | 81.85 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| ERM + T3ABackbone=CLIP ViT-B16, Test-time batch size=64, Test-time adaptation method=T3A2021.11 | 88.2 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| FADBackbone=ResNet-50, Pre-training=ImageNet, Training Iterations=5,000, Evaluation Protocol=DomainBed (Training-domain Validation Set)2026.06 | 88.2 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| FACTBackbone=ResNet-502023.04 | 88.15 | 89.63 | 81.77 | 96.75 | 84.46 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| SWADEnsemble=true2022.12 | 88.1 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| SWADBackbone=ResNet50, Ens/MA=Yes2023.09 | 88.1 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| RISEBackbone=EfficientNet, Pre-training source=IN-1K2026.07 | 88 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| ERM + Tent*Backbone=CLIP ViT-B16, Test-time batch size=64, Test-time adaptation method=Tent, Update target=linear classifier2021.11 | 87.8 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| GENIEBackbone=ResNet-50, Pre-training=ImageNet, Training Iterations=5,000, Evaluation Protocol=DomainBed (Training-domain Validation Set)2026.06 | 87.8 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| ERM + Pseudo Label*Backbone=CLIP ViT-B16, Test-time batch size=64, Test-time adaptation method=Pseudo Label, Update target=linear classifier2021.11 | 87.6 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| EnsembleBackbone=ResNet50, Ens/MA=Yes2023.09 | 87.6 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Single sample generalizationBackbone=ResNet-502022.02 | 87.51 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| VL2VBackbone=EfficientNet, Pre-training source=IN-1K2026.07 | 86.8 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Seo et al. (2020)Backbone=ResNet-502022.02 | 86.64 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| RandConvBackbone=ResNet-50, Reproduced=true2023.04 | 86.63 | 86.13 | 82 | 96.65 | 81.72 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Best SOTA competitorBackbone=ResNet50, Ens/MA=No2023.09 | 86.6 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Wang et al. (2021)Backbone=ResNet-502022.02 | 86.23 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| CORAL2022.12 | 86.2 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| CLIPoodBackbone=Swin Transformer, Pre-training source=IN-21K2026.07 | 86.2 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| LPBackbone=RegNetY-16GF, Pre-training source=IN-12K2026.07 | 86.2 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| GAMBackbone=ResNet-50, Pre-training=ImageNet, Training Iterations=5,000, Evaluation Protocol=DomainBed (Training-domain Validation Set)2026.06 | 86.1 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| EISNetBackbone=ResNet-502023.04 | 85.84 | 86.64 | 81.53 | 97.11 | 78.07 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| ERMBackbone=ResNet50, Ens/MA=No2023.09 | 85.7 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Baseline (ERM)Backbone=ResNet-502023.04 | 85.61 | 87.15 | 83.82 | 97.77 | 73.71 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Gulrajani & Lopez-Paz (2020)Backbone=ResNet-502022.02 | 85.5 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| ERMBackbone=ResNet502022.03 | 85.5 | 84.7 | 80.8 | 97.2 | 79.3 | 79.3 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Fish2022.12 | 85.5 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| MIROBackbone=RegNetY-16GF, Pre-training source=IN-12K2026.07 | 85.4 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Zhao et al. (2020)Backbone=ResNet-502022.02 | 85.34 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| RISE (ERM + Hint + AD + RD)Backbone=ResNet18, Ens/MA=No2023.09 | 85.3 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| SAMBackbone=ResNet-50, Pre-training=ImageNet, Training Iterations=5,000, Evaluation Protocol=DomainBed (Training-domain Validation Set)2026.06 | 85.3 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Seo et al. (2020)Backbone=ResNet-182022.02 | 85.11 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| ERM + Hint + ADBackbone=ResNet18, Ens/MA=No2023.09 | 85.1 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Zhou et al. (2020a)Backbone=ResNet-502022.02 | 84.9 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| ERM + Hint + RDBackbone=ResNet18, Ens/MA=No2023.09 | 84.9 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| MLDG2022.12 | 84.8 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| MixupBackbone=ResNet502022.03 | 84.6 | 86.1 | 78.9 | 97.6 | 75.8 | 75.8 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| MTLBackbone=ResNet502022.03 | 84.6 | 87.5 | 77.1 | 96.4 | 77.3 | 77.1 | — | — | — | — | — | — | — | — | — | — | — | — | — | — |