OOD Detection on ImageNet-1K OOD (Average of OpenImage-O, Texture, iNaturalist, ImageNet-O) (test)
97.95AUROCCoEvoNegLabel
Evaluation results
| Method | Links | ||
|---|---|---|---|
| CoEvoNegLabelEncoder=CLIP ViT-B/16, Method Category=Training-free2026.01 | 97.95 | 10.22 | |
| CoEvoCSPEncoder=CLIP ViT-B/16, Method Category=Training-free2026.01 | 97.85 | 10.94 | |
| AdaNegEncoder=CLIP ViT-B/16, Method Category=Training-free2026.01 | 96.66 | 18.92 | |
| CSPEncoder=CLIP ViT-B/16, Method Category=Training-free2026.01 | 95.76 | 17.51 | |
| NegPromptEncoder=CLIP ViT-B/16, Method Category=Training-based2026.01 | 94.81 | 23.01 | |
| LAPTEncoder=CLIP ViT-B/16, Method Category=Training-based2026.01 | 94.68 | 23.4 | |
| NegLabelEncoder=CLIP ViT-B/16, Method Category=Training-free2026.01 | 94.21 | 25.4 | |
| ViMBackbone=Swin [26], Source=feat+logit, Pre-trained=timm [35]2022.03 | 94.11 | 31.04 | |
| LoCoOpEncoder=CLIP ViT-B/16, Method Category=Training-based2026.01 | 93.52 | 28.66 | |
| EnergyBackbone=CLIP-L, Training Protocol=Fine-tuned2022.11 | 93.5 | 29.42 | |
| CLIPNEncoder=CLIP ViT-B/16, Method Category=Training-based2026.01 | 93.1 | 31.1 | |
| ResidualBackbone=Swin [26], Source=feat, Pre-trained=timm [35]2022.03 | 92.88 | 37.38 | |
| EnergyBackbone=CLIP-B, Training Protocol=Fine-tuned2022.11 | 92.26 | 35.92 | |
| MahalanobisBackbone=Swin [26], Source=feat+label, Pre-trained=timm [35]2022.03 | 92.16 | 40.39 | |
| EnergyEncoder=CLIP ViT-B/16, Method Category=Training-based2026.01 | 91.54 | 39.89 | |
| MCMBackbone=CLIP-L, Training Protocol=Zero-shot2022.11 | 91.49 | 38.17 | |
| NPOSEncoder=CLIP ViT-B/16, Method Category=Training-based2026.01 | 91.22 | 37.93 | |
| MCMEncoder=CLIP ViT-B/16, Method Category=Training-free2026.01 | 90.82 | 43.93 | |
| MCMBackbone=CLIP-B, Training Protocol=Zero-shot2022.11 | 90.77 | 42.74 | |
| ReActBackbone=Swin [26], Source=feat+logit, Pre-trained=timm [35]2022.03 | 90.17 | 31.36 | |
| MOSBackbone=BiT, Training Protocol=Fine-tuned2022.11 | 90.11 | 39.97 | |
| ViMBackbone=Res50d [11], Source=feat+logit, Pre-trained=timm [35]2022.03 | 89.22 | 52.61 | |
| Fort et al.Backbone=ViT-L, Training Protocol=Fine-tuned2022.11 | 88.96 | 43.65 | |
| KL MatchingBackbone=Swin [26], Source=prob, Pre-trained=timm [35]2022.03 | 88.87 | 46.99 | |
| MaxLogitBackbone=Swin [26], Source=logit, Pre-trained=timm [35]2022.03 | 88.4 | 35.28 | |
| MahalanobisBackbone=Res50d [11], Source=feat+label, Pre-trained=timm [35]2022.03 | 88.33 | 55.7 | |
| Fort et al.Backbone=ViT-B, Training Protocol=Fine-tuned2022.11 | 88.25 | 45.12 | |
| ODINBackbone=Swin [26], Source=prob+grad, Pre-trained=timm [35]2022.03 | 88 | 36.58 | |
| ViMBackbone=RepVGG [7], Source=feat+logit, Pre-trained=timm [35]2022.03 | 87.81 | 50.5 | |
| EnergyBackbone=Swin [26], Source=logit, Pre-trained=timm [35]2022.03 | 87.77 | 35.08 | |
| MSPBackbone=Swin [26], Source=prob, Pre-trained=timm [35]2022.03 | 87.57 | 43.44 | |
| ResidualBackbone=Res50d [11], Source=feat, Pre-trained=timm [35]2022.03 | 87.01 | 58.55 | |
| MahalanobisBackbone=RepVGG [7], Source=feat+label, Pre-trained=timm [35]2022.03 | 86.07 | 59.39 | |
| MSPBackbone=CLIP-L, Training Protocol=Fine-tuned2022.11 | 85.68 | 53.71 | |
| ViMBackbone=DeiT [33], Source=feat+logit, Pre-trained=timm [35]2022.03 | 85.25 | 69.95 | |
| MahalanobisBackbone=DeiT [33], Source=feat+label, Pre-trained=timm [35]2022.03 | 85.03 | 73.18 | |
| ResidualBackbone=RepVGG [7], Source=feat, Pre-trained=timm [35]2022.03 | 84.19 | 59 | |
| ResidualBackbone=DeiT [33], Source=feat, Pre-trained=timm [35]2022.03 | 84.15 | 74.13 | |
| KL MatchingBackbone=DeiT [33], Source=prob, Pre-trained=timm [35]2022.03 | 83.49 | 64.8 | |
| ReActBackbone=Res50d [11], Source=feat+logit, Pre-trained=timm [35]2022.03 | 82.93 | 58.63 | |
| KL MatchingBackbone=Res50d [11], Source=prob, Pre-trained=timm [35]2022.03 | 82.72 | 64.41 | |
| MSPBackbone=CLIP-B, Training Protocol=Fine-tuned2022.11 | 82.04 | 59.89 | |
| ZOCEncoder=CLIP ViT-B/16, Method Category=Training-based2026.01 | 81.79 | 85.19 | |
| MSPEncoder=CLIP ViT-B/16, Method Category=Training-based2026.01 | 81.63 | 69.61 | |
| KL MatchingBackbone=RepVGG [7], Source=prob, Pre-trained=timm [35]2022.03 | 81.35 | 61.65 | |
| EnergyEncoder=CLIP ViT-B/16, Method Category=Training-free2026.01 | 79.57 | 82.21 | |
| MSPBackbone=DeiT [33], Source=prob, Pre-trained=timm [35]2022.03 | 79.48 | 66.43 | |
| MSPBackbone=RepVGG [7], Source=prob, Pre-trained=timm [35]2022.03 | 78.1 | 70.55 | |
| MSPBackbone=Res50d [11], Source=prob, Pre-trained=timm [35]2022.03 | 77.99 | 67.96 | |
| ODINBackbone=RepVGG [7], Source=prob+grad, Pre-trained=timm [35]2022.03 | 77.72 | 72.68 | |
| MaxLogitBackbone=RepVGG [7], Source=logit, Pre-trained=timm [35]2022.03 | 77.56 | 73.5 | |
| ReActBackbone=DeiT [33], Source=feat+logit, Pre-trained=timm [35]2022.03 | 77.37 | 67 | |
| ODINBackbone=DeiT [33], Source=prob+grad, Pre-trained=timm [35]2022.03 | 77.13 | 63.92 | |
| MaxLogitBackbone=DeiT [33], Source=logit, Pre-trained=timm [35]2022.03 | 76.79 | 64.49 | |
| EnergyBackbone=RepVGG [7], Source=logit, Pre-trained=timm [35]2022.03 | 76.38 | 78.99 | |
| MaxLogitBackbone=Res50d [11], Source=logit, Pre-trained=timm [35]2022.03 | 75.39 | 69.34 | |
| ODINBackbone=Res50d [11], Source=prob+grad, Pre-trained=timm [35]2022.03 | 75.27 | 68.56 | |
| EnergyBackbone=DeiT [33], Source=logit, Pre-trained=timm [35]2022.03 | 72.8 | 70.14 | |
| GradNormEncoder=CLIP ViT-B/16, Method Category=Training-based2026.01 | 72.35 | 80.82 | |
| EnergyBackbone=Res50d [11], Source=logit, Pre-trained=timm [35]2022.03 | 71.08 | 78.39 | |
| ReActBackbone=RepVGG [7], Source=feat+logit, Pre-trained=timm [35]2022.03 | 49.14 | 98.96 |