Research map / Vision-language and multimodal / CLIP and zero-shot
AnomalyVFM -- Transforming Vision Foundation Models into Zero-Shot Anomaly Detectors
Matic Fučka, Vitjan Zavrtanik, Danijel Skočaj
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 synthetic dataset and feature adapters together to DINOv2 | DINOv2 · 9 industrial datasets avg | image-level AUROC: 83.0 -> 90.2 | Improved |
| add synthetic dataset and feature adapters to DINOv3 | DINOv3 · 9 industrial datasets avg | pixel-level AUROC: 88.0 -> 95.3 | Improved |
| add synthetic dataset and feature adapters to RADIO | RADIO · 9 industrial datasets avg | image-level AUROC: 89.1 -> 94.1 | Improved |
| use AnomalyVFM (RADIO) over best prior zero-shot method | AnomalyVFM · 9 industrial datasets avg | image-level AUROC: 90.8 -> 94.1 | Improved |
| use AnomalyVFM over AdaCLIP on pixel-level F1-max | AnomalyVFM · 9 industrial datasets avg | pixel-level F1-max: 51.0 -> 44.3 | Hurt |
| AnomalyVFM on medical image-level detection vs prior best | AnomalyVFM · medical datasets avg | image-level AUROC: 94.9 -> 94.0 | No clear effect |
| AnomalyVFM on medical pixel-level localisation vs AdaCLIP | AnomalyVFM · medical datasets avg | pixel-level AUROC: 88.8 -> 90.0 | Improved |
| add feature-based dataset filtering (ablation) | AnomalyVFM · 9 industrial datasets avg | pixel-level AUROC: -14.6 without filtering | Improved |
| add foreground-guided anomaly location selection (ablation) | AnomalyVFM · 9 industrial datasets avg | pixel-level AUROC: -5.8 without foreground selection | Improved |
| add confidence-weighted loss (ablation) | AnomalyVFM · 9 industrial datasets avg | pixel-level AUROC: -2.0 without it | Improved |
| replace FLUX with QWEN-Image or WAN generator (ablation) | AnomalyVFM · 9 industrial datasets avg | image-level AUROC: -0.1 (QWEN), -0.4 (WAN) | No clear effect |
| replace LoRA with VPT adapters (ablation) | AnomalyVFM · 9 industrial datasets avg | pixel-level F1-max: image AUROC -1.0, pixel F1 +4.0 | Mixed |
Closest papers by meaning
- Towards Zero-Shot Anomaly Detection and Reasoning with Multimodal Large Language Models (CVPR 2025 · similarity 0.89)
- UniVAD: A Training-free Unified Model for Few-shot Visual Anomaly Detection (CVPR 2025 · similarity 0.89)
- Unseen Visual Anomaly Generation (CVPR 2025 · similarity 0.89)
- AnomalyCLIP: Object-agnostic Prompt Learning for Zero-shot Anomaly Detection (ICLR 2024 · similarity 0.88)
- One-to-Normal: Anomaly Personalization for Few-shot Anomaly Detection (NeurIPS 2024 · similarity 0.88)
- Adapting Visual-Language Models for Generalizable Anomaly Detection in Medical Images (CVPR 2024 · similarity 0.88)
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.
