• Official PyTorch implementation for paper: "Is Task-Specific Training Necessary for Anomaly Detction?"
  

• [2026/01/28] The code of RAD is released! 🎉
• [2026/05/01] RAD is accepted by ICML2026! 🎉

• Setup
• Prepare Datasets
• Download pretrained Encoder
• Multi-class Anomaly Detection

First, clone this repository to your local machine, and install the dependencies.

conda create -n rad python=3.11 -y
conda activate rad
cd RAD
pip install -r requirements.txt

❗ You can modify the PyTorch version to suit your machine. Experiments are conducted on an NVIDIA GeForce RTX 5090 (32GB). Same GPU and package version are recommended.

Noted that ../ is the upper directory of RAD 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.

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

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

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
        |-- ....

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
        |-- ....

The 3D-ADAM dataset can be found publicly available in the HuggingFace repository.

Ignore the /xyz/ since we only use the RGB modality and download the dataset to ../3D-ADAM_anomalib

python ./prepare_data/prepare_3dadam.py --data-folder ../3D-ADAM_anomalib --save-folder ../3D-ADAM

../3D-ADAM will be like:

|-- 3D-ADAM
    |-- 1m1
        |-- ground_truth
        |-- test
            |-- good
            |-- <defect_1>
            |-- <defect_2>
            |-- ...
        |-- train
            |-- good
    |-- 1m2
    |-- 1m3
    |-- 2m1
    |-- 2m2h
    |-- 2m2m
    |-- 3m1
    |-- ...

• DINOv3 ViT-B/16: #parameters: 86M

First build multi-layer anomaly-free memory bank.

python build_bank_multilayer.py \
    --data_path ../mvtec_anomaly_detection \
    --item_list carpet grid leather tile wood \
               bottle cable capsule hazelnut metal_nut \
               pill screw toothbrush transistor zipper \
    --encoder_name dinov3_vitb16 \
    --encoder_weight ../weights/dinov3_vitb16_pretrain_lvd1689m-73cec8be.pth \
    --image_size 512 \
    --crop_size 448 \
    --batch_size 32 \
    --num_workers 4 \
    --layer_idx_list 3 6 9 11 \
    --bank_path ./bank/mvtec_dinov3_vitb16_multilayer36911_448_bank.pth

• data_path is the folder dir that stores the datasets.
• item_list specifics the item classes of the datasets.
• bank_path is the folder dir that stores the bank.
• layer_idx_list specifics the layer index (zero-based) of the encoder we extract features.

Using RAD as a detecter.

python rad_mvtec_visa_3dadam.py \
    --data_path ../mvtec_anomaly_detection \
    --bank_path ./bank/mvtec_dinov3_vitb16_multilayer36911_448_bank.pth \
    --encoder_name dinov3_vitb16 \
    --encoder_weight ../weights/dinov3_vitb16_pretrain_lvd1689m-73cec8be.pth \
    --save_dir ./saved_results_mvtec \
    --save_name mvtec_patch_knn_multilayer_36911_448 \
    --k_image 150 \
    --resize_mask 256 \
    --max_ratio 0.01 \
    --vis_max 8 \
    --use_positional_bank \
    --pos_radius 1
    --item_list carpet grid leather tile wood \
               bottle cable capsule hazelnut metal_nut \
               pill screw toothbrush transistor zipper \

• data_path is the folder dir that stores the datasets.
• save_dir specifies the directory where logs, metrics, and visualization results are saved.
• k_image sets how many nearest-neighbor training images are used to build the local patch memory bank for each test image.
• resize_mask sets the spatial resolution to which anomaly maps and ground-truth masks are resized for evaluation and visualization.
• max_ratio determines what fraction of the highest anomaly pixels are pooled to compute the image-level anomaly score.
• vis_max limits the maximum number of test samples that will be visualized per class.
• use_positional_bank enables the position-aware patch memory bank so that each patch only compares with patches from nearby spatial locations rather than all positions.
• pos_radius controls the spatial neighborhood radius (in patch units) used for the position-aware patch memory bank around each query patch.
• item_list specifics the item classes of the datasets.

First build multi-layer anomaly-free memory bank.

python build_bank_multilayer.py \
    --data_path ../VisA_pytorch/1cls \
    --item_list candle capsules cashew chewinggum fryum \
               macaroni1 macaroni2 pcb1 pcb2 pcb3 pcb4 pipe_fryum \
    --encoder_name dinov3_vitb16 \
    --encoder_weight ../weights/dinov3_vitb16_pretrain_lvd1689m-73cec8be.pth \
    --image_size 512 \
    --crop_size 448 \
    --batch_size 32 \
    --num_workers 4 \
    --layer_idx_list 3 6 9 11 \
    --bank_path ./bank/visa_dinov3_vitb16_multilayer36911_448_bank.pth

• data_path is the folder dir that stores the datasets.
• item_list specifics the item classes of the datasets.
• bank_path is the folder dir that stores the bank.
• layer_idx_list specifics the layer index (zero-based) of the encoder we extract features.

Using RAD as a detecter.

python rad_mvtec_visa_3dadam.py \
    --data_path ../VisA_pytorch/1cls \
    --bank_path ./bank/visa_dinov3_vitb16_multilayer36911_448_bank.pth \
    --encoder_name dinov3_vitb16 \
    --encoder_weight ../weights/dinov3_vitb16_pretrain_lvd1689m-73cec8be.pth \
    --save_dir ./saved_results_patch_knn \
    --save_name visa_patch_knn_multilayer_36911_448 \
    --k_image 900 \
    --resize_mask 256 \
    --max_ratio 0.01 \
    --vis_max 8 \
    --use_positional_bank \
    --pos_radius 2

• data_path is the folder dir that stores the datasets.
• save_dir specifies the directory where logs, metrics, and visualization results are saved.
• k_image sets how many nearest-neighbor training images are used to build the local patch memory bank for each test image.
• resize_mask sets the spatial resolution to which anomaly maps and ground-truth masks are resized for evaluation and visualization.
• max_ratio determines what fraction of the highest anomaly pixels are pooled to compute the image-level anomaly score.
• vis_max limits the maximum number of test samples that will be visualized per class.
• use_positional_bank enables the position-aware patch memory bank so that each patch only compares with patches from nearby spatial locations rather than all positions.
• pos_radius controls the spatial neighborhood radius (in patch units) used for the position-aware patch memory bank around each query patch.
• item_list specifics the item classes of the datasets.

First build multi-layer anomaly-free memory bank.

python build_realiad_bank_multilayer.py \
  --data_path ../Real-IAD \
  --encoder_name dinov3_vitb16 \
  --encoder_weight ../weights/dinov3_vitb16_pretrain_lvd1689m-73cec8be.pth \
  --image_size 512 \
  --crop_size 448 \
  --layer_idx_list 3 6 9 11 \
  --bank_path ./bank/realiad_dinov3_vitb16_multilayer36911_448_bank.pth

• data_path is the folder dir that stores the datasets.
• bank_path is the folder dir that stores the bank.
• layer_idx_list specifics the layer index (zero-based) of the encoder we extract features.

Using RAD as a detecter.

python rad_real_iad.py \
  --data_path ../Real-IAD \
  --bank_path ./bank/realiad_dinov3_vitb16_multilayer36911_448_bank.pth \
  --encoder_name dinov3_vitb16 \
  --encoder_weight ../weights/dinov3_vitb16_pretrain_lvd1689m-73cec8be.pth \
  --save_dir ./saved_results_realiad \
  --save_name realiad_multilayer_36911_448 \
  --k_image 900 \
  --resize_mask 256 \
  --max_ratio 0.01 \
  --vis_max 8 \
  --use_positional_bank \
  --pos_radius 0

• data_path is the folder dir that stores the datasets.
• save_dir specifies the directory where logs, metrics, and visualization results are saved.
• k_image sets how many nearest-neighbor training images are used to build the local patch memory bank for each test image.
• resize_mask sets the spatial resolution to which anomaly maps and ground-truth masks are resized for evaluation and visualization.
• max_ratio determines what fraction of the highest anomaly pixels are pooled to compute the image-level anomaly score.
• vis_max limits the maximum number of test samples that will be visualized per class.
• use_positional_bank enables the position-aware patch memory bank so that each patch only compares with patches from nearby spatial locations rather than all positions.
• pos_radius controls the spatial neighborhood radius (in patch units) used for the position-aware patch memory bank around each query patch.

First build multi-layer anomaly-free memory bank.

python build_bank_multilayer.py \
    --data_path ../3D-ADAM \
--item_list 1m1 1m2 1m3 2m1 2m2h 2m2m \
           3m1 3m2 3m2c 4m1 4m2 4m2c \
           gripper_closed gripper_open helicalgear1 helicalgear2 \
           rackgear spiralgear spurgear tapa2m1 tapa3m1 tapa4m1 tapatbb \
    --encoder_name dinov3_vitb16 \
    --encoder_weight ../weights/dinov3_vitb16_pretrain_lvd1689m-73cec8be.pth \
    --image_size 512 \
    --crop_size 448 \
    --batch_size 32 \
    --num_workers 4 \
    --layer_idx_list 3 6 9 11 \
    --bank_path ./bank/3dadam_dinov3_vitb16_multilayer36911_448_bank.pth

• data_path is the folder dir that stores the datasets.
• item_list specifics the item classes of the datasets.
• bank_path is the folder dir that stores the bank.
• layer_idx_list specifics the layer index (zero-based) of the encoder we extract features.

Using RAD as a detecter.

python rad_mvtec_visa_3dadam.py \
    --data_path ../3D-ADAM \
    --bank_path ./bank/3dadam_dinov3_vitb16_multilayer36911_448_bank.pth \
    --encoder_name dinov3_vitb16 \
    --encoder_weight ../weights/dinov3_vitb16_pretrain_lvd1689m-73cec8be.pth \
    --save_dir ./saved_results_patch_knn \
    --save_name 3dadam_patch_knn_multilayer_36911_448 \
    --k_image 48 \
    --resize_mask 256 \
    --max_ratio 0.01 \
    --vis_max 8 \
    --use_positional_bank \
    --pos_radius 1

• data_path is the folder dir that stores the datasets.
• save_dir specifies the directory where logs, metrics, and visualization results are saved.
• k_image sets how many nearest-neighbor training images are used to build the local patch memory bank for each test image.
• resize_mask sets the spatial resolution to which anomaly maps and ground-truth masks are resized for evaluation and visualization.
• max_ratio determines what fraction of the highest anomaly pixels are pooled to compute the image-level anomaly score.
• vis_max limits the maximum number of test samples that will be visualized per class.
• use_positional_bank enables the position-aware patch memory bank so that each patch only compares with patches from nearby spatial locations rather than all positions.
• pos_radius controls the spatial neighborhood radius (in patch units) used for the position-aware patch memory bank around each query patch.
• item_list specifics the item classes of the datasets.

This repository is built using the Dinomaly, the DINOv3 and many other inspiring works in the community.

This project is licensed under the license found in the LICENSE file in the root directory of this source tree. Microsoft Open Source Code of Conduct