Research map / Label-efficient and robust learning / OOD and anomaly detection
CLIPN for Zero-Shot OOD Detection: Teaching CLIP to Say No
Hualiang Wang, Yi Li, Huifeng Yao, Xiaomeng Li
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add CLIPN-A (learnable no text encoder, agreeing-to-differ) versus MCM | CLIP ViT-B-16 · ImageNet-1K zero-shot OOD | AUROC / FPR95: +2.34 AUROC, -11.64 FPR95 | Improved |
| add text semantic-opposite loss to image-text binary-opposite loss (ablation) | CLIPN-A · ImageNet-1K OOD sets | AUROC: iNaturalist 93.08 -> 95.27 | Improved |
| use learnable no prompts instead of handcrafted for competing-to-win (ablation) | CLIPN-C · ImageNet-1K OOD sets | AUROC: +6.83, +6.85, +0.66, +6.86 | Improved |
| use learnable no prompts instead of handcrafted for agreeing-to-differ (ablation) | CLIPN-A · ImageNet-1K OOD sets | AUROC: +2.52, +2.38, +2.63; Texture -1.22 | Mixed |
| use agreeing-to-differ instead of competing-to-win at inference (ablation) | CLIPN · ImageNet-1K avg, ViT-B-16 | AUROC: 86.35 -> 93.10 | Improved |
| CLIPN-A versus best baselines on CIFAR-100 ID small-scale | CLIP ViT-B-32 · CIFAR-100 ID | AUROC / FPR95: +5.23 AUROC avg; ODIN better on ImageNet_R and LSUN FPR95 | Mixed |
| use CLIPN-A instead of MSP thresholds on CIFAR-10 vs CIFAR-100 | CLIPN-A · CIFAR-10 ID, CIFAR-100 OOD | misclassified OOD samples: 666/410/171 -> 45 | Improved |
| train only the no text encoder with frozen image and text encoders instead of full CLIP training | CLIPN · training ViT-L-14 batch 16 | time per iteration / GPU memory: 0.58 -> 0.24 s; 18.91 -> 6.62 G | Improved |
| add a second text encoder (no encoder) at inference | CLIP ViT-B-32 · ImageNet-1K inference | avg AUROC: 85.59 -> 90.53 (+4.94) | Improved |
| use CLIPN-A on in-domain CIFAR splits | CLIP · CIFAR-10 / CIFAR-100 in-domain | AUROC: +4.5 CIFAR10, +7.75 CIFAR100 over MSP | Improved |
Closest papers by meaning
- ID-like Prompt Learning for Few-Shot Out-of-Distribution Detection (CVPR 2024 · similarity 0.90)
- Adaptive Multi-prompt Contrastive Network for Few-shot Out-of-distribution Detection (ICML 2025 · similarity 0.89)
- CLIPure: Purification in Latent Space via CLIP for Adversarially Robust Zero-Shot Classification (ICLR 2025 · similarity 0.89)
- Learning Transferable Negative Prompts for Out-of-Distribution Detection (CVPR 2024 · similarity 0.89)
- Envisioning Outlier Exposure by Large Language Models for Out-of-Distribution Detection (ICML 2024 · similarity 0.88)
- Does CLIP’s generalization performance mainly stem from high train-test similarity? (ICLR 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
