MRAD: Memory Retrieval Anomaly Detection

MRAD: Zero-Shot Anomaly Detection with Memory-Driven Retrieval

Introduction

MRAD is a novel approach for zero-shot anomaly detection that leverages memory banks and CLIP features. By building retrieval-augmented memory banks from auxiliary datasets, MRAD achieves strong cross-domain anomaly detection and localization without any target domain training data. This repository provides three model variants with different trade-offs:

Model	Description	Training	Weights
MRAD-CLIP	Full model with memory bank + CLIP image-text alignment	6 epochs	patch_proj, image_proj, prompt_learner, prompt_proj
MRAD-FT	Fine-tuned model with memory bank projection	1 epoch	patch_proj, image_proj
MRAD-TF	Train-free baseline using raw CLIP features	None	None

Install dependencies --- Python 3.9, PyTorch 2.0.0+cu117

pip install -r requirements.txt

Prepare your dataset

Download the dataset below:

• Industrial Domain: MVTec, VisA, MPDD, BTAD, DTD-Synthetic

• Medical Domain: HeadCT, BrainMRI, Br35H, CVC-ClinicDB, CVC-ColonDB, Kvasir, ISIC, Endo

Generate the dataset JSON (same as AnomalyCLIP)

Take MVTec AD for example:

Structure of MVTec Folder:

mvtec/
│
├── meta.json
│
├── bottle/
│   ├── ground_truth/
│   │   ├── broken_large/
│   │   │   └── 000_mask.png
│   │   │   └── ...
│   │   └── ...
│   └── test/
│       ├── broken_large/
│       │   └── 000.png
│       │   └── ...
│       └── ...
│
└── ...

cd generate_dataset_json
python mvtec.py

Select the corresponding script and run it. The generated JSON stores all the information that MRAD needs.

Run MRAD

• Quick start with pre-trained weights (MRAD-CLIP, trained on VisA, tested on MVTec):

python test.py \
    --model_type mrad-clip \
    --checkpoint_path ./checkpoints/test_on_mvtec.pth \
    --dataset mvtec \
    --data_path ./data/MVTEC \
    --cache_dir ./cache \
    --save_path ./results/test_on_mvtec \
    --device cuda:0

• Train your own weights:

python train.py \
    --model_type mrad-clip \
    --dataset visa \
    --data_path ./data/VISA \
    --save_path ./checkpoints/mrad_clip \
    --cache_dir ./cache \
    --device cuda:0

Before running the above commands, please download the official memory banks from this link and the corresponding MRAD-CLIP model weights from this link, and place them under ./cache and ./checkpoints respectively.

The ./cache directory stores the memory banks built from MVTec and VisA. If you want to rebuild the memory banks from scratch, please set the load_cache argument of both build_cache_model and build_patch_cache_model to False in train.py.

• End-to-end pipeline (train on VisA + test on MVTec):

If you want to retrain from scratch, you can use the following command as a quick start, but please set the GPU device according to your actual environment.

bash run.sh mrad-clip

You can switch model variants by changing --model_type to mrad-ft or mrad-clip.

Experiment setting

We train on VisA and test on all other datasets. For VisA, MRAD is trained on MVTec.

Acknowledgments

We thank the authors of the following code repositories: AnomalyCLIP, open_clip, CLIP_Surgery.

BibTex Citation

If you find this paper and repository useful, please cite our paper.

@inproceedings{
xu2026mrad,
title={{MRAD}: Zero-Shot Anomaly Detection with Memory-Driven Retrieval},
author={Chaoran Xu and Chengkan Lv and Qiyu Chen and Feng Zhang and Zhengtao Zhang},
booktitle={The Fourteenth International Conference on Learning Representations},
year={2026}
}