Out-of-distribution Detection on CIFAR-10 (ID) vs Celeb-A (OOD)
100AUROCForte+SVM
Evaluation results
| Method | Links | ||
|---|---|---|---|
| Forte+SVMSupervision=Unsupervised2024.10 | 100 | 0 | |
| Forte+GMMSupervision=Unsupervised2024.10 | 100 | 0 | |
| SPEM-noiseDensity Estimator=Glow2026.02 | 99.93 | — | |
| Forte+KDESupervision=Unsupervised2024.10 | 99.75 | 0.06 | |
| WAICapproach_type=Generative-based2019.09 | 99.7 | — | |
| Label-Free MahalanobisBackbone=DINOv3 ViT-L, Computational cost (#F + #J)=1F + 0J2026.05 | 99.1 | — | |
| lambda (Similarity-only)2026.02 | 98.94 | — | |
| ReSCOPEDBackbone=DINOv3 ViT-L, Computational cost (#F + #J)=2F + 1J2026.05 | 98.8 | — | |
| SPEM-noiseFlow Model=ResFlow, Feature Extractor=ResNet-152, Hyperparameter alpha=0.12026.02 | 98.73 | — | |
| Label-Free MahalanobisBackbone=DINOv3 ViT-B, Computational cost (#F + #J)=1F + 0J2026.05 | 98.5 | — | |
| SPEMFlow Model=ResFlow, Feature Extractor=ResNet-152, Hyperparameter alpha=0.42026.02 | 98.3 | — | |
| SPEM-noiseDensity Estimator=Glow2026.02 | 98.3 | — | |
| ReSCOPEDBackbone=DINOv3 ViT-B, Computational cost (#F + #J)=2F + 1J2026.05 | 98 | — | |
| SPEMDensity Estimator=Glow2026.02 | 97.82 | — | |
| Label-Free MahalanobisBackbone=DINOv3 ViT-S, Computational cost (#F + #J)=1F + 0J2026.05 | 97.7 | — | |
| DOSESupervision=Unsupervised2024.10 | 97.6 | 12.82 | |
| ReSCOPEDBackbone=DINOv3 ViT-S, Computational cost (#F + #J)=2F + 1J2026.05 | 97.5 | — | |
| SPEMDensity Estimator=Glow2026.02 | 97.01 | — | |
| LIDDensity Estimator=Glow2026.02 | 93.9 | — | |
| GMMDensity Estimator=Glow2026.02 | 93.36 | — | |
| WAICSupervision=Unsupervised2024.10 | 92.8 | 46.04 | |
| Typicality (sqrt(d))Density Estimator=Glow2026.02 | 92.53 | — | |
| q(X|theta_eta)Supervision=Unsupervised2024.10 | 91.4 | 51.1 | |
| DoSE2026.05 | 90.4 | — | |
| MSMAComputational cost (#F + #J)=10F + 0J2026.05 | 90.1 | — | |
| DiffPathComputational cost (#F + #J)=10F + 0J2026.05 | 89.9 | — | |
| MOODFLOPs=0.79 x 10^82021.04 | 88 | — | |
| S scoregenerative_model=Glow, compression_algorithm=FLIF2019.09 | 86.3 | — | |
| Serrà et altype=likelihood ratio-based, variant=best results, FLOPs=2.78 x 10^102021.04 | 86 | — | |
| LMDComputational cost (#F + #J)=10^4F + 0J2026.05 | 85.4 | — | |
| SCOPEDComputational cost (#F + #J)=2F + 2J2026.05 | 85.2 | — | |
| TTSupervision=Unsupervised2024.10 | 84.8 | 74 | |
| AENIBScore=max_y p(y|x)2023.03 | 81 | — | |
| LLRSupervision=Unsupervised2024.10 | 80.08 | 64.79 | |
| AENIBScore=log Dir_0.05(y)2023.03 | 80 | — | |
| JEMtype=generative-based, variant=best results2021.04 | 79 | — | |
| JEMScore=max_y p(y|x)2023.03 | 79 | — | |
| AENIBScore=log Dir_0.05(y) + log N(z_n; 0, I)2023.03 | 79 | — | |
| VIBScore=log Dir_0.05(y)2023.03 | 78 | — | |
| S scoregenerative_model=PixelCNN++, compression_algorithm=FLIF2019.09 | 77.6 | — | |
| VIBScore=max_y p(y|x)2023.03 | 76 | — | |
| TT2026.05 | 75.9 | — | |
| ComplexityDensity Estimator=Glow2026.02 | 75.64 | — | |
| JEMScore=log p(x)2023.03 | 75 | — | |
| LikelihoodDensity Estimator=Glow2026.02 | 74.08 | — | |
| VIBapproach_type=Classifier-based2019.09 | 73.5 | — | |
| Likelihood RatioDensity Estimator=Glow2026.02 | 73.34 | — | |
| Typicality (Entropy)Density Estimator=Glow2026.02 | 72.37 | — | |
| Likelihood RatioFlow Model=ResFlow2026.02 | 70.18 | — | |
| SITN2026.05 | 67.1 | — | |
| SITN2026.05 | 67.1 | — | |
| LIDFlow Model=ResFlow2026.02 | 65.5 | — | |
| LIDDensity Estimator=Glow2026.02 | 65.5 | — | |
| Cross-entropyScore=max_y p(y|x)2023.03 | 64 | — | |
| LikelihoodFlow Model=ResFlow2026.02 | 63.46 | — | |
| DDPM-OODComputational cost (#F + #J)=350F + 0J2026.05 | 63.2 | — | |
| Typicality (sqrt(d))Density Estimator=Glow2026.02 | 62.97 | — | |
| ComplexityFlow Model=ResFlow2026.02 | 62.52 | — | |
| GMMDensity Estimator=Glow2026.02 | 61.74 | — | |
| Complexity2026.05 | 61.5 | — | |
| Cross-entropyScore=log Dir_0.05(y)2023.03 | 61 | — | |
| PS-CDTraining Set=CIFAR-102021.11 | 58 | — | |
| ComplexityDensity Estimator=Glow2026.02 | 55.66 | — | |
| Typicality (sqrt(d))Flow Model=ResFlow2026.02 | 54.29 | — | |
| Glowtype=generative-based, FLOPs=4.09 x 10^92021.04 | 54 | — | |
| Likelihood RatioDensity Estimator=Glow2026.02 | 51.46 | — | |
| LikelihoodDensity Estimator=Glow2026.02 | 51.44 | — | |
| CDTraining Set=CIFAR-102021.11 | 51 | — | |
| Log-likelihood2026.05 | 49.6 | — | |
| Log-likelihood2026.05 | 49.6 | — | |
| NLLComputational cost (#F + #J)=1000F + 0J2026.05 | 48.2 | — | |
| WAIC2026.05 | 46.4 | — | |
| WAICensemble size=52026.05 | 45 | — | |
| DoSE2026.05 | 43.2 | — | |
| Typicality (Entropy)Density Estimator=Glow2026.02 | 42.83 | — | |
| Typicality2026.05 | 42.1 | — | |
| GMMFlow Model=ResFlow2026.02 | 40.5 | — | |
| LR2026.05 | 36.8 | — | |
| Typicality (Entropy)Flow Model=ResFlow2026.02 | 36.33 | — |