Dinomaly

PyTorch Implementation of CVPR 2025 “Dinomaly: The Less Is More Philosophy in Multi-Class Unsupervised Anomaly Detection”. The first multi-class UAD model that can compete with single-class SOTAs !!!

• Give me a ⭐️ if you like it.

• This project is released under Apache-2.0 License. If your method is based on Dinomaly, please acknowledge it in your work (papers, products, patents, competitions, etc.), preferably as a “Preliminary” section.

• See here for the code of Dinomaly2.

News

• 05.2024: Arxiv preprint and github code released🚀

• 09.2024: Rejected by NeurIPS 2024 with 5 positive scores and no negative score, because "AC: lack of novelty"😭. Wish me good luck.

• 02.2025: Accepted by CVPR 2025🎉

• 07.2025: Spoil alert: We will come back with Dinomaly2😛

• 07.2025: Dinomaly has been integrated in Intel open-edge Anomalib in v2.1.0. Great thanks to the contributors for the nice reproduction and integration. Anomalib is a comprehensive library for benchmarking, developing and deploying deep learning anomaly detection algorithms.

• 08.2025: I have sucessfully implement DINOv3 on Dinomaly. The pixel-level performance is much better, with slightly lower image-level performance. DINOv3-Large on MVTecAD: I-Auroc:0.9970, P-AUROC:0.9878, P-AP:0.7422, P-F1:0.7184, P-AUPRO:0.9580. Due to DINOv3 requiring newer versions of Python (>3.10) and PyTorch (>2.7), it is not provided in this repository. You can refer to cnlab’s forked branch. Great thanks!

• 10.2025: We are thrilled to present the extended version of Dinomaly, now evolved into Dinomaly2!!! We introduce the first unified framework for full-spectrum UAD that seamlessly handles diverse data 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). Of course, Dinomaly2 achieves unprecedented UAD performance. Check it out😎

• 05.2026:Preview code of Dinomaly2 released at here. We also support super-large multi-class datasets: Real-IAD_Variety (160 classes) and ADNet (380 classes). How great is that!🚀

Abstract

Recent studies highlighted a practical setting of unsupervised anomaly detection (UAD) that builds a unified model for multi-class images. Despite various advancements addressing this challenging task, the detection performance under the multi-class setting still lags far behind state-of-the-art class-separated models. Our research aims to bridge this substantial performance gap. In this paper, we introduce a minimalistic reconstruction-based anomaly detection framework, namely Dinomaly, which leverages pure Transformer architectures without relying on complex designs, additional modules, or specialized tricks. Given this powerful framework consisted of only Attentions and MLPs, we found four simple components that are essential to multi-class anomaly detection: (1) Foundation Transformers that extracts universal and discriminative features, (2) Noisy Bottleneck where pre-existing Dropouts do all the noise injection tricks, (3) Linear Attention that naturally cannot focus, and (4) Loose Reconstruction that does not force layer-to-layer and point-by-point reconstruction. Extensive experiments are conducted across popular anomaly detection benchmarks including MVTec-AD, VisA, and Real-IAD. Our proposed Dinomaly achieves impressive image-level AUROC of 99.6%, 98.7%, and 89.3% on the three datasets respectively (99.8%, 98.9%, 90.1% with ViT-L), which is not only superior to state-of-the-art multi-class UAD methods, but also achieves the most advanced class-separated UAD records.

1. Environments

Create a new conda environment and install required packages.

conda create -n my_env python=3.8.12
conda activate my_env
pip install -r requirements.txt

Experiments are conducted on NVIDIA GeForce RTX 3090 (24GB). Same GPU and package version are recommended.

2. Prepare Datasets

Noted that ../ is the upper directory of Dinomaly code. It is where we keep all the datasets by default. You can also alter it according to your need, just remember to modify the data_path in the code.

MVTec AD

Download the MVTec-AD dataset from URL. Unzip the file to ../mvtec_anomaly_detection.

|-- mvtec_anomaly_detection
    |-- bottle
    |-- cable
    |-- capsule
    |-- ....

VisA

Download the VisA dataset from URL. Unzip the file to ../VisA/. Preprocess the dataset to ../VisA_pytorch/ in 1-class mode by their official splitting code.

You can also run the following command for preprocess, which is the same to their official code.

python ./prepare_data/prepare_visa.py --split-type 1cls --data-folder ../VisA --save-folder ../VisA_pytorch --split-file ./prepare_data/split_csv/1cls.csv

../VisA_pytorch will be like:

|-- VisA_pytorch
    |-- 1cls
        |-- candle
            |-- ground_truth
            |-- test
                    |-- good
                    |-- bad
            |-- train
                    |-- good
        |-- capsules
        |-- ....

Real-IAD

Contact the authors of Real-IAD URL to get the net disk link.

Download and unzip realiad_1024 and realiad_jsons in ../Real-IAD. ../Real-IAD will be like:

|-- Real-IAD
    |-- realiad_1024
        |-- audiokack
        |-- bottle_cap
        |-- ....
    |-- realiad_jsons
        |-- realiad_jsons
        |-- realiad_jsons_sv
        |-- realiad_jsons_fuiad_0.0
        |-- ....

3. Run Experiments

Multi-Class Setting

python dinomaly_mvtec_uni.py --data_path ../mvtec_anomaly_detection

python dinomaly_visa_uni.py --data_path ../VisA_pytorch/1cls

python dinomaly_realiad_uni.py --data_path ../Real-IAD

Conventional Class-Separted Setting

python dinomaly_mvtec_sep.py --data_path ../mvtec_anomaly_detection

python dinomaly_visa_sep.py --data_path ../VisA_pytorch/1cls

python dinomaly_realiad_sep.py --data_path ../Real-IAD

Trained model weights

2x default training iterations for better performance.

Results

A. Compare with MUAD SOTAs:

Dinomaly can perfectly scale with model size, input image size, and the choice of foundation model.

B. Model Size:

C. Input Size:

D. Choice of Foundaiton Model:

Eval discrepancy of anomaly localization

In our code implementation, we binarize the GT mask using gt.bool() after down-sampling, specifically gt[gt>0]=1. As pointed out in an issue, the previous common practice is to use gt[gt>0.5]=1. The difference between these two binarization approaches is that gt[gt>0]=1 may result in anomaly regions being one pixel larger compared to gt[gt>0.5]=1. This difference does not affect image-level performance metrics, but it has a slight impact on pixel-level evaluation metrics.

We think gt[gt>0]=1 is a more reasonable choice. It can be seen as max pooling, so that in the down-sampled GT map, any position that corresponds to a region containing at least one anomaly pixel in the original map is marked as anomalous. If an anomaly region is extremely small in the original image (say 2 pixels), gt[gt>0.5]=1 will erase it while gt[gt>0]=1 can keep it.

Loss NaN

If you encounter Loss=NaN during training on other datasets (very rare in common datasets), simply add a small eps (1e-6 by default) in the LinearAttention2 module:

z = 1.0 / (torch.einsum('...sd,...d->...s', q, k.sum(dim=-2)) + self.eps)

Citation

@inproceedings{guo2025dinomaly,
  title={Dinomaly: The less is more philosophy in multi-class unsupervised anomaly detection},
  author={Guo, Jia and Lu, Shuai and Zhang, Weihang and Chen, Fang and Li, Huiqi and Liao, Hongen},
  booktitle={Proceedings of the Computer Vision and Pattern Recognition Conference},
  pages={20405--20415},
  year={2025}
}

@article{guo2025one,
  title={One Dinomaly2 Detect Them All: A Unified Framework for Full-Spectrum Unsupervised Anomaly Detection},
  author={Guo, Jia and Lu, Shuai and Fan, Lei and Li, Zelin and Di, Donglin and Song, Yang and Zhang, Weihang and Zhu, Wenbing and Yan, Hong and Chen, Fang and Li, Huiqi and Liao, Hongen},
  journal={arXiv preprint arXiv:2510.17611},
  year={2025}
}