Research map / Label-efficient and robust learning / OOD and anomaly detection

Dinomaly: The Less Is More Philosophy in Multi-Class Unsupervised Anomaly Detection

CVPR 2025 · Label-efficient and robust learning › OOD and anomaly detection · approach: Unified multi-class anomaly detection · 103 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Dinomaly builds a minimalist, pure-Transformer reconstruction framework for multi-class unsupervised anomaly detection, using noisy MLP bottleneck, linear attention, and loose group-wise reconstruction loss on a DINOv2-pretrained ViT.

Problem

Multi-class unsupervised anomaly detection (MUAD) with a single unified model lags far behind class-separated UAD models, largely due to the identity-mapping (over-generalization) problem.

Approach

Reframes identity mapping as over-generalization and fixes it without extra modules: Dropout in the MLP bottleneck as pseudo feature noise, Linear Attention whose unfocusing limits copying, and loose reconstruction that groups layers and discards well-reconstructed points, all on a scalable DINOv2 ViT encoder.

Key results

• Dinomaly reaches image-level AUROC of 99.6% on MVTec-AD, 98.7% on VisA, and 89.3% on Real-IAD in the multi-class setting, beating prior multi-class SoTA by large margins.
• Scaling to ViT-Large raises results to 99.8/98.9/90.1 image-level AUROC (with longer training for further gains), while ViT-Small already reaches SoTA.
• Multi-class Dinomaly is nearly on par with class-separated SoTA on MVTec-AD and VisA, with moderate drop on Real-IAD.

Measured effects reported in the paper

Surprising findings

• Unlike prior anomaly detection networks, Dinomaly benefits from scaling model size (ViT-Small to ViT-Large), reversing the previously reported lack of scaling law.
• Increasing input size degrades competing methods such as RD4AD and ReContrast, while Dinomaly benefits from larger inputs for localization.

Limitations

• Performance drop when moving from class-separated to multi-class setting on Real-IAD remains moderate.
• Image-level performance on MVTec-AD is nearly saturated, limiting discriminative comparison there.

Experimental setup

Datasets: MVTec-AD VisA Real-IAD

Benchmarks: MVTec-AD VisA Real-IAD MPDD BTAD UniMedical

Metrics: image-level AUROC image-level AP image-level F1-max pixel-level AUROC pixel-level AP pixel-level F1-max AUPRO

Baselines: UniAD RD4AD ReContrast DiAD ViTAD MambaAD DeSTSeg SimpleNet PatchCore

Builds on: DINOv2 RD4AD UniAD ReContrast Linear Attention

Closest papers by meaning

• AnomalyNCD: Towards Novel Anomaly Class Discovery in Industrial Scenarios (CVPR 2025 · similarity 0.88)
• Generating and Reweighting Dense Contrastive Patterns for Unsupervised Anomaly Detection (AAAI 2024 · similarity 0.88)
• ResAD: A Simple Framework for Class Generalizable Anomaly Detection (NeurIPS 2024 · similarity 0.88)
• Correcting Deviations from Normality: A Reformulated Diffusion Model for Multi-Class Unsupervised Anomaly Detection (CVPR 2025 · similarity 0.87)
• UniVAD: A Training-free Unified Model for Few-shot Visual Anomaly Detection (CVPR 2025 · similarity 0.87)
• Long-Tailed Anomaly Detection with Learnable Class Names (CVPR 2024 · similarity 0.87)