Out-of-Distribution Detection on CelebA (ID) vs Multiple OOD Sets
100AUROC (CIFAR-100)GEPC-CelebA
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| GEPC-CelebAProtocol=Diffusion-based, single-checkpoint, Logical Cost=8F2026.05 | 100 | 100 | 100 | — | 90.8 | |
| CFS(1×2)-CelebAProtocol=Diffusion-based, single-checkpoint, Logical Cost=1F2026.05 | 99.9 | 99.9 | 100 | 100 | 93.5 | |
| CFS(1×2)-ImageNetProtocol=Diffusion-based, single-checkpoint, Logical Cost=1F2026.05 | 99.9 | 99.9 | 100 | 99.8 | 96.2 | |
| CFS(1×2)-ImageNet (kNN)Logical Cost=1F, Head=kNN2026.05 | 99.9 | 99.9 | 100 | 99.9 | 97.4 | |
| DiffPathTraining data regime=Full-data, Baselines trained on=SVHN2026.05 | 99.8 | 99.8 | 100 | 99.9 | 93.1 | |
| DiffPathTraining Data Regime=Full-data2026.05 | 99.8 | 99.8 | 100 | 99.9 | 93.1 | |
| DiffPathTraining Mode=Full-data, Training Set=CelebA2026.05 | 99.8 | 99.8 | 100 | 99.9 | 93.1 | |
| DiffPath-6D-CelebAProtocol=Diffusion-based, single-checkpoint, Logical Cost=10F2026.05 | 99.8 | 99.8 | 100 | 99.9 | 93.1 | |
| Unified Few-shotTraining data regime=Few-shot, samples per task=80-1502026.05 | 99.1 | 99.5 | 100 | 99.6 | 93.7 | |
| UFCODTraining Data Regime=Few-shot2026.05 | 99.1 | 99.5 | 100 | 99.6 | 93.7 | |
| Unified Few-shotTraining Mode=Few-shot, Training Set=CelebA, Samples per task=80-1502026.05 | 99.1 | 99.5 | 100 | 99.6 | 93.7 | |
| LMDProtocol=Diffusion-based, ID-specific, Logical Cost=10^4F2026.05 | 97.9 | 98.9 | 100 | 97.2 | 86.5 | |
| Label-Free MahalanobisBackbone=DINOv3 ViT-L, Computational cost (#F + #J)=1F + 0J2026.05 | 97 | 100 | 99.9 | — | — | |
| ReSCOPEDBackbone=DINOv3 ViT-L, Computational cost (#F + #J)=2F + 1J2026.05 | 96.9 | 100 | 99.7 | — | — | |
| SCOPED-CelebAProtocol=Diffusion-based, single-checkpoint, Logical Cost=2F+2J2026.05 | 96.2 | 92.5 | 99.4 | — | 89.2 | |
| EigenScoreProtocol=Diffusion-based, ID-specific, Logical Cost=300F2026.05 | 94.4 | 96.5 | 88.8 | — | 92.5 | |
| MSMAProtocol=Diffusion-based, ID-specific, Logical Cost=10F2026.05 | 92.7 | 91 | 99.6 | 99.9 | 94.5 | |
| Label-Free MahalanobisBackbone=DINOv3 ViT-B, Computational cost (#F + #J)=1F + 0J2026.05 | 92.3 | 100 | 100 | — | — | |
| ReSCOPEDBackbone=DINOv3 ViT-B, Computational cost (#F + #J)=2F + 1J2026.05 | 91.9 | 100 | 100 | — | — | |
| DoSTraining Mode=Full-data, Training Set=CelebA2026.05 | 88.9 | 89.9 | 22 | 76.9 | 62.5 | |
| NLLTraining Mode=Full-data, Training Set=CelebA2026.05 | 88.1 | 89.5 | 24.3 | 75 | 62.6 | |
| Label-Free MahalanobisBackbone=DINOv3 ViT-S, Computational cost (#F + #J)=1F + 0J2026.05 | 86.4 | 100 | 100 | — | — | |
| ReSCOPEDBackbone=DINOv3 ViT-S, Computational cost (#F + #J)=2F + 1J2026.05 | 86.4 | 99.9 | 100 | — | — | |
| DiffPath-6D-ImageNetProtocol=Diffusion-based, single-checkpoint, Logical Cost=10F2026.05 | 84.3 | 80.7 | 98.1 | 96.4 | 85 | |
| Improved CD2026.05 | 83 | — | — | — | — | |
| VAEBMTraining Mode=Full-data, Training Set=CelebA2026.05 | 82.4 | 85.2 | 85.6 | 83.9 | 54 | |
| NLLProtocol=Diffusion-based, ID-specific, Logical Cost=1000F2026.05 | 78.6 | 81.4 | 10.5 | 80.9 | 68.9 | |
| DDPM-OODProtocol=Diffusion-based, ID-specific, Logical Cost=350F2026.05 | 77.8 | 79.5 | 63.6 | 77.3 | 74.6 | |
| DDPM-OODTraining Mode=Full-data, Training Set=CelebA2026.05 | 76 | 81.5 | 22.9 | 55.5 | 56.7 | |
| VAEBMTraining Data Regime=Full-data2026.05 | 71 | 72.2 | 85.2 | 85.7 | 54.3 | |
| LMDTraining Mode=Full-data, Training Set=CelebA2026.05 | 70.5 | 72.9 | 6.2 | 68.8 | 60.8 | |
| DoSTraining data regime=Full-data, Baselines trained on=SVHN2026.05 | 64.3 | 63.9 | 0 | 62.4 | 61 | |
| NLLTraining data regime=Full-data, Baselines trained on=SVHN2026.05 | 64 | 65.6 | 0.1 | 63 | 61.1 | |
| MSMAComputational cost (#F + #J)=10F + 0J2026.05 | 61.5 | 91 | 99.6 | — | — | |
| LMDComputational cost (#F + #J)=10^4F + 0J2026.05 | 60.4 | 98.9 | 100 | — | — | |
| DiffPathComputational cost (#F + #J)=10F + 0J2026.05 | 59 | 99.8 | 100 | — | — | |
| IGEBMTraining Mode=Full-data, Training Set=CelebA2026.05 | 58.9 | 61 | 73.5 | 58.9 | 49.5 | |
| DoSE2026.05 | 57.1 | 94.9 | 99.7 | — | — | |
| TT2026.05 | 54.8 | 63.4 | 98.2 | — | — | |
| VAEBMTraining data regime=Full-data, Baselines trained on=SVHN2026.05 | 54.7 | 56.6 | 42.1 | 42.9 | 48.9 | |
| DDPM-OODComputational cost (#F + #J)=350F + 0J2026.05 | 53.6 | 79.5 | 63.6 | — | — | |
| WAIC2026.05 | 53.2 | 50.7 | 13.9 | — | — | |
| NLLComputational cost (#F + #J)=1000F + 0J2026.05 | 52.1 | 81.4 | 10.5 | — | — | |
| LR2026.05 | 52 | 32.3 | 2.8 | — | — | |
| LMDTraining Data Regime=Full-data2026.05 | 51.5 | 50 | 1.2 | 59.3 | 59.4 | |
| ICProtocol=Diffusion-based, ID-specific, Logical Cost=1000F2026.05 | 51 | 48.5 | 97.2 | 55.9 | 45.1 | |
| DDPM-OODTraining data regime=Full-data, Baselines trained on=SVHN2026.05 | 50.8 | 48.6 | 0.7 | 34.7 | 53.8 | |
| NLLTraining Data Regime=Full-data2026.05 | 49.2 | 40.7 | 1.6 | 50.4 | 58 | |
| DDPM-OODTraining Data Regime=Full-data2026.05 | 49.1 | 49.1 | 5.6 | 37.1 | 53.2 | |
| DoSTraining Data Regime=Full-data2026.05 | 48.5 | 42 | 1.6 | 51.9 | 58.5 | |
| LMDTraining data regime=Full-data, Baselines trained on=SVHN2026.05 | 48.3 | 45.2 | 0 | 55 | 59.3 | |
| IGEBMTraining Data Regime=Full-data2026.05 | 47.9 | 47.1 | 55.3 | 45.4 | 47.9 | |
| SCOPEDComputational cost (#F + #J)=2F + 2J2026.05 | 47.7 | 92.5 | 99.4 | — | — | |
| MSMATraining Data Regime=Full-data2026.05 | 47.6 | 44.1 | 57.3 | 47.1 | 48.3 | |
| MSMATraining data regime=Full-data, Baselines trained on=SVHN2026.05 | 40.7 | 43.2 | 83.6 | 60.8 | 46.7 | |
| MSMATraining Mode=Full-data, Training Set=CelebA2026.05 | 39.3 | 44.4 | 61.9 | 44.6 | 46.2 | |
| IGEBMTraining data regime=Full-data, Baselines trained on=SVHN2026.05 | 37.3 | 34.2 | 38.2 | 40.4 | 48.9 | |
| DLSR (+LR)Note=Outside MBE protocol2026.05 | — | — | — | 98.5 | 96 | |
| DLSR (+MFsim)Note=Outside MBE protocol2026.05 | — | — | — | 100 | 96.1 | |
| DLSR (+MSE)Note=Outside MBE protocol2026.05 | — | — | — | 99.9 | 96.2 |