One Dinomaly2 Detect Them All: A Unified Framework for Full-Spectrum Unsupervised Anomaly Detection

TL;DR

Dinomaly2 introduces a minimalist, unified framework for unsupervised anomaly detection that spans 2D, multi-view, RGB-3D, and RGB-IR data, while handling single-class, multi-class, inference-unified MUAD, and few-shot settings. Built on a Vision Transformer backbone, it combines a Noisy Bottleneck, Unfocused Linear Attention, Context-Aware Recentering, and Loose Reconstruction to achieve robust reconstruction-based anomaly detection without task-specific tailoring. Across 12 benchmarks and four modalities, it delivers state-of-the-art results, demonstrates strong scaling behavior with model size and input resolution, and shows a high correlation between foundation-model pretraining quality and anomaly detection performance. The work argues that architectural simplicity paired with principled components enables a truly universal UAD framework suitable for diverse real-world deployments.

Abstract

Unsupervised anomaly detection (UAD) has evolved from building specialized single-class models to unified multi-class models, yet existing multi-class models significantly underperform the most advanced one-for-one counterparts. Moreover, the field has fragmented into specialized methods tailored to specific scenarios (multi-class, 3D, few-shot, etc.), creating deployment barriers and highlighting the need for a unified solution. In this paper, we present Dinomaly2, the first unified framework for full-spectrum image UAD, which bridges the performance gap in multi-class models while seamlessly extending across diverse data modalities and task settings. Guided by the "less is more" philosophy, we demonstrate that the orchestration of five simple element achieves superior performance in a standard reconstruction-based framework. This methodological minimalism enables natural extension across diverse tasks without modification, establishing that simplicity is the foundation of true universality. Extensive experiments on 12 UAD benchmarks demonstrate Dinomaly2's full-spectrum superiority across multiple modalities (2D, multi-view, RGB-3D, RGB-IR), task settings (single-class, multi-class, inference-unified multi-class, few-shot) and application domains (industrial, biological, outdoor). For example, our multi-class model achieves unprecedented 99.9% and 99.3% image-level (I-) AUROC on MVTec-AD and VisA respectively. For multi-view and multi-modal inspection, Dinomaly2 demonstrates state-of-the-art performance with minimum adaptations. Moreover, using only 8 normal examples per class, our method surpasses previous full-shot models, achieving 98.7% and 97.4% I-AUROC on MVTec-AD and VisA. The combination of minimalistic design, computational scalability, and universal applicability positions Dinomaly2 as a unified solution for the full spectrum of real-world anomaly detection applications.

Figures (9)

• Figure 1: Overview: task settings, benchmarks, and scalability of Dinomaly2. (a) Task settings include single-class UAD, multi-class UAD (MUAD), and inference-unified MUAD. (b) Multi-modal capabilities span 2D images, multi-view 2D, RGB+3D point cloud, and RGB+Infrared. (c) Cross-domain validation across industrial inspection, medical imaging, and aerial surveillance. (d) Dinomaly2 achieves performance on MVTec-AD (15 categories), VisA (12 categories), Real-IAD (30 categories$\times$5 views), and MVTec3D (10 categories) that surpasses the corresponding state-of-the-art methods. (e) Scalability analysis shows consistent improvements with model size, input resolution, and foundation model quality (on MVTec-AD).
• Figure 2: The framework of Dinomaly2, built by simple and pure Transformer building blocks. A pre-trained encoder extracts multi-layer features, which are aggregated through a bottleneck and reconstructed by a decoder. During training, the model is optimized exclusively on normal images, while during inference, anomalies are detected based on reconstruction errors.
• Figure 3: Softmax Attention vs. Linear Attention. (a) Visualization of attention maps, and (b) Attention distribution, demonstrate both demonstrates Softmax Attention's tendency to focus on local regions versus Linear Attention's distributed attention pattern.
• Figure 4: Context-Aware Recentering and a toy example. (a) Separate detectors for sidewalk and highway show how identical objects (pedestrians/vehicles) may receive opposite anomaly labels depending on context. (b) A unified detector faces contextual ambiguity without proper conditioning. (c) Our Context-Aware Recentering uses class tokens as reference anchors to map patch features into class-specific coordinate systems, resolving multi-class confusion through simple subtraction.
• Figure 5: Schemes of reconstruction constraint. (a) Layer-to-layer (sparse). (b) Layer-to-cat-layer. (c) Layer-to-layer (dense). (d) Loose group-to-group, 1-group (Ours). (e) Loose group-to-group, 2-group (Ours).