UniADet

Official PyTorch Implementation of One Language-Free Foundation Model Is Enough for Universal Vision Anomaly Detection, 2026.

• 首尔大学 SNU DSBA Lab 视频讲解 in Korean
• “我爱计算机视觉”解读 in Chinese
• "木迅"解读 in Chinese
• alphaxiv blog in English

📢 News

• [2026-06-02]: Thank you for your interest in UniADet. We are truly grateful for the widespread attention and valuable feedback from researchers, students, and CV practitioners worldwide, including those from the University of Cambridge, Texas A&M University, Institute of Automation of the Chinese Academy of Sciences, Xi'an Jiaotong-Liverpool University, Shenzhen University, University of Jinan, and other institutions. The paper is currently under peer review. The full source code and pre-trained models will be officially released upon the acceptance. We sincerely appreciate your patience and understanding.
• [2026-02-02]: ✅ UniADet has been independently reproduced by a third-party team, confirming our state-of-the-art results.
• [2026-01-13]: 🚀 Initialized the official UniADet code repository.
• [2026-01-09]: 📄 The UniADet paper is now available on arXiv.

Contents

• Introduction
• UniADet Framework
• Language-Free UniADet with Different Foundation Models
• Comparison with State-of-the-Arts
• Complexity and Efficiency Comparisons
• Ablation Studies
• Comparison with Language-Dependent AnomalyCLIP
• ToDo List
• Citation

📝 Introduction

UniADet is a language-free universal (Zero- and Few-shot) visual anomaly detection framework. It achieves superior performance outperforming state-of-the-art language-dependent Zero- and Few-shot AD models while also exhibiting remarkable simplicity and efficiency.

• We rethink vision-language ADs and find that language prompts and encoders are unnecessary. This insight leads to an embarrassingly simple(language-free+dual-decoupling), efficient(0.015M or 0.02M learnable params), effective (SOTA zero-/few-shot) and general (support VLMs and Pure VMs) framework for universal anomaly detection.
• We fully decouple global anomaly classification and local anomaly segmentation across multi-scale hierarchical features, i.e., learning layer-wise cls/seg weights, effectively mitigating the learning conflict between different feature manifolds and substantially improving AD performance.
• Comprehensive experiments conclusively validate that our approach achieves state-of-the-art zero-shot and few-shot performance. Notably, our few-shot UniADet is the first to outperform full-shot state-of-the-art.

💎 UniADet Framework

📊 Language-Free UniADet with Different Foundation Models

Backbone	Shot	MVTec-AD	VisA	Real-IAD
CLIP (ViT-L/14@336px)	0	92.4 / 42.8	88.0 / 28.0	78.6 / 33.6
CLIP (ViT-L/14@336px)	4	97.7 / 58.8	93.3 / 36.7	84.3 / 37.2
DINOv2 (Register ViT-L/14)	0	93.5 / 50.9	91.3 / 32.7	82.5 / 43.1
DINOv2 (Register ViT-L/14)	4	98.7 / 65.4	96.9 / 45.2	90.3 / 48.5
DINOv3 (ViT-L/16)	0	94.0 / 52.7	91.9 / 32.5	81.2 / 41.6
DINOv3 (ViT-L/16)	4	98.2 / 69.0	97.1 / 45.5	88.5 / 49.8

Note: The performance is mesured by Image-AUROC / Pixel-AUPR, and the same below.

🏆 Comparison with State-of-the-Arts

Methods	Venue	Language-Free	Shots	MVTec	VisA	Real-IAD
UniADet $^‡$	ours	$\color{red}{\checkmark}$	0	93.5 / 50.9	91.3 / 32.7	82.5 / 43.1
WinCLIP	CVPR 23	$\color{green}{✘}$	0	90.4 / 18.2	75.5 / 5.4	67.0 / 3.3
APRIL-GAN	CVPRW 23	$\color{green}{✘}$	0	86.1 / -	78.0 / -	-
AnomalyCLIP	ICLR 24	$\color{green}{✘}$	0	91.6 / 34.5	82.0 / 21.3	69.5 / 26.7
AdaCLIP	ECCV 24	$\color{green}{✘}$	0	90.7 / 39.1	81.7 / 31.0	73.3 / 30.5
VCPCLIP	ECCV 24	$\color{green}{✘}$	0	92.1 / 49.4	83.8 / 30.1	-
Bayes-PFL	CVPR 25	$\color{green}{✘}$	0	92.5 / 48.3	87.0 / 29.8	70.0 / 27.6
AA-CLIP	CVPR 25	$\color{green}{✘}$	0	90.5 / -	84.6 / -	-
FE-CLIP	ICCV 25	$\color{green}{✘}$	0	91.9 / -	84.6 / -	-
FAPrompt	ICCV 25	$\color{green}{✘}$	0	91.9 / -	84.6 / -	-
RareCLIP	ICCV 25	$\color{green}{✘}$	0	91.5 / 46.1	86.1 / 27.0	-
AdaptCLIP	AAAI 26	$\color{green}{✘}$	0	93.5 / 38.3	84.8 / 26.1	74.2 / 28.2

Methods	Venue	Language-Free	Shots	MVTec	VisA	Real-IAD
UniADet $^‡$	ours	$\color{red}{\checkmark}$	1	97.6 / 63.1	95.2 / 42.1	88.7 / 48.4
UniADet $^‡$	ours	$\color{red}{\checkmark}$	2	98.0 / 64.1	96.1 / 44.2	89.0 / 46.7
UniADet $^‡$	ours	$\color{red}{\checkmark}$	4	98.7 / 65.4	96.9 / 45.2	90.3 / 48.5
MetaUAS	NeurIPS 24	$\color{red}{\checkmark}$	1	90.7 / 59.3	81.2 / 42.7	80.0 / 36.6
APRIL-GAN	CVPRW 23	$\color{green}{✘}$	4	92.8 / 54.5	92.6 / 32.2	-
PromptAD	CVPR 24	$\color{green}{✘}$	4	96.6 / 52.9	89.1 / 31.5	-
UniVAD	CVPR 25	$\color{green}{✘}$	1	97.8 / 55.6	93.5 / 42.8	85.1 / 37.6
AdaptCLIP	AAAI 26	$\color{green}{✘}$	1	94.5 / 53.7	90.5 / 38.9	81.8 / 36.6
AdaptCLIP	AAAI 26	$\color{green}{✘}$	2	95.7 / 55.1	92.2 / 40.7	82.9 / 37.8
AdaptCLIP	AAAI 26	$\color{green}{✘}$	4	96.6 / 57.2	93.1 / 41.8	83.9 / 39.1

Methods	Venue	Language-Free	Setting	MVTec	VisA	Real-IAD
Dinomaly	CVPR 25	$\color{red}{\checkmark}$	multi-class (full train set)	99.6 / 69.3	98.7 / 53.2	89.3 / 42.8
UniAD	NeurIPS 24	$\color{red}{\checkmark}$	multi-class (full train set)	96.5 / 44.7	90.8 / 33.6	83.0 / 21.1
MuSc	ICLR 24	$\color{red}{\checkmark}$	online (full test set)	97.8 / 62.7	92.8 / 45.1	-

Important

Note1: If you find that any existing zero-shot/few-shot AD methods are missing from the table above, please feel free to open an issue so we can add them.

Note2: Dinomaly and UniAD are multi-class unsupervised AD algorithms, and they require dataset-specific training with full normal images.

Note3: MuSc is an Online algorithm that requires access to statistics from the entire test dataset to evaluate the current image. Therefore, it is not a strictly zero-shot AD.

🚀 Complexity and Efficiency Comparisons

Shots	Methods	Models	Input Size	# Params (M)	Inf. Time (ms)
0	AdaCLIP	CLIP ViT-L/14@336px	518×518	428.8 + 1.1e+1	107.4
0	AnomalyCLIP	CLIP ViT-L/14@336px	518×518	427.9 + 5.6e+0	70.7
0	Bayes-PFL	CLIP ViT-L/14@336px	518×518	427.9 + 2.7e+1	154.9
0	AdaptCLIP-Zero	CLIP ViT-L/14@336px	518×518	427.9 + 6.0e-1	57.5
0	UniADet $^†$	CLIP ViT-L/14@336px	518×518	342.9 + 1.5e-2	15.7
0	UniADet $^‡$	DINOv2 ViT-L/14	518×518	303.2 + 2.0e-2	41.9
1	InCtrl	CLIP ViT-B-16+240	240×240	208.4 + 3.0e-1	59.0
1	AnomalyCLIP+	CLIP ViT-L/14@336px	518×518	427.9 + 5.6e+0	76.2
1	AdaptCLIP	CLIP ViT-L/14@336px	518×518	342.9 + 1.8e+0	58.7
1	UniADet $^†$	CLIP ViT-L/14@336px	518×518	342.9 + 1.5e-2	22.4
1	UniADet $^‡$	DINOv2 ViT-L/14	518×518	303.2 + 2.0e-2	48.4

Note: The number of learnable parameters (1.5e-3 and 2.0e-3) is not correct for our UniADet $^†$ and UniADet $^‡$, respectively. The correct is 1.5e-2 and 2.0e-2 for UniADet $^†$ and UniADet $^‡$, respectively.

🔍 Ablation Studies

Ablation studies about different components.

No	DCS	DHF	CAA	Shot	MVTec	VisA
0	$\color{green}{✘}$	$\color{green}{✘}$	$\color{green}{✘}$	0	85.4 / 36.4	77.9 / 26.1
1	$\color{red}{\checkmark}$	$\color{green}{✘}$	$\color{green}{✘}$	0	91.8 / 38.3	85.9 / 27.2
2	$\color{red}{\checkmark}$	$\color{red}{\checkmark}$	$\color{green}{✘}$	0	92.2 / 40.7	86.0 / 27.6
3	$\color{red}{\checkmark}$	$\color{red}{\checkmark}$	$\color{red}{\checkmark}$	0	92.4 / 42.8	88.0 / 28.0
4	$\color{red}{\checkmark}$	$\color{red}{\checkmark}$	random	0	91.3 / 41.5	87.5 / 26.6
5	$\color{red}{\checkmark}$	$\color{red}{\checkmark}$	$\color{red}{\checkmark}$	1	95.9 / 54.6	91.3 / 32.5

Note: The ablation studies are conducted by UniADet $^†$ (i.e., using CLIP ViT-L/14@336px).

⚖️ Comparison with Language-Dependent AnomalyCLIP

• Essential Differences

Feature	AnomalyCLIP	UniADet (Ours)
🧠 Paradigm	🔴 Language-Dependent	✅ Language-Free
⚡ Task Decoupling	Shared Cls/Seg Weight $W$	✅ Decoupled Cls/Seg Weights $W_{cls}$, $W_{seg}$
🏗️ Hierarchical Decoupling	Single / Last Layer	✅ Layer-Wise Cls/Seg Weights $W_{cls}^l$, $W_{seg}^l$
🤖 Backbones	CLIP Only	✅ CLIP, DINOv2-R, DINOv3
📉 Params	⚠️ 130M Text-Encoder + 5.6M (Heavy)	🚀 0.015M or 0.020M(Efficient)

Important

Note1: The official AnomalyCLIP team has clarified that their initial multi-layer claim was a code bug; the effective implementation relies only on the last layer.

Note2: A naive extension of AnomalyCLIP to multiple blocks leads to significant performance degradation.

• Commonality

Feature	AnomalyCLIP & UniADet
Core Objective	Learning Normal / Anomaly Weights
Loss Function	CE for Cls and Focal + Dice for Seg
Training Data	Auxiliary Data (e.g., VisA or MVTec)

• Performace on VisA

Shots	Backbones	AnomalyCLIP	UniADet (Ours)
0-Shot	CLIP	82.0 / 21.3	88.0 / 28.0
0-Shot	DINOv2-R	Not supported	91.3 / 32.7
0-Shot	DINOv3	Not supported	91.9 / 32.5
4-Shot	CLIP	Not supported	93.3 / 36.7
4-Shot	DINOv2-R	Not supported	96.9 / 45.2
4-Shot	DINOv3	Not supported	97.1 / 45.5

📌 ToDo List

• release pre-trained UniADet models
• deploy online UniADet Demo on huggingface
• open training and testing code

📖 Citation

If you find this work useful in your research, please consider citing:

@inproceedings{uniadet,
  title={One Language-Free Foundation Model Is Enough for Universal Vision Anomaly Detection},
  author={Gao, Bin-Bin and Wang, Chengjie},
  booktitle={arXiv:2601.05552},
  year={2026}
}

Star History