SubspaceAD: Training-Free Few-Shot Anomaly Detection via Subspace Modeling (CVPR2026)

This repository contains the official implementation of the paper:

SubspaceAD: Training-Free Few-Shot Anomaly Detection via Subspace Modeling (CVPR 2026)

---

Figure 1. SubspaceAD consists of two training-free stages: (1) extracting DINOv2 patch features from a few normal exemplars, and (2) estimating a low-dimensional PCA subspace to model normal appearance. Anomalies are detected via reconstruction residuals.

Introduction

Detecting visual anomalies in industrial inspection often requires operating with only a few normal exemplars per category. While many recent approaches rely on large memory banks, auxiliary datasets, or multi-modal tuning, SubspaceAD introduces a minimalist, training-free alternative.

SubspaceAD consists of two stages:

1. Feature Extraction: Patch-level features are extracted from a small set of normal images using a frozen DINOv2 backbone.
2. Subspace Modeling: A PCA model is fit to these features to estimate the low-dimensional manifold of normal appearance.

At inference time, anomalies are detected using the reconstruction residual with respect to this learned subspace. Despite its simplicity, SubspaceAD achieves state-of-the-art performance in one-shot and few-shot settings.

Main Results (1-Shot)

• MVTec-AD: 97.1% Image AUROC; 97.5% Pixel AUROC
• VisA: 93.4% Image AUROC; 98.2% Pixel AUROC

---

MVTec-AD

VisA

Figure 2. Qualitative comparison on VisA and MVTec-AD (1-shot). SubspaceAD produces sharper and more precise anomaly maps than PromptAD and AnomalyDINO, with fewer false activations and better alignment with ground-truth defects across both datasets.

Environment Setup

# 1. Create environment
conda create -n subspacead python=3.10
conda activate subspacead

# 2. Install dependencies and the package
pip install -r requirements.txt
pip install -e .

Data Preparation

MVTec-AD

Download the dataset from the MVTec website and extract it to:

datasets/mvtec-ad/

VisA

# 1. Download and extract
mkdir -p datasets/VisA_20220922
wget https://amazon-visual-anomaly.s3.us-west-2.amazonaws.com/VisA_20220922.tar
tar -xvf VisA_20220922.tar -C datasets/VisA_20220922

# 2. Preprocess (reorganize folder structure)
python tools/prepare_visa.py \
    --data-folder datasets/VisA_20220922 \
    --save-folder datasets/VisA_pytorch

Folder Structure

pca-dino/
├── datasets/                   # Dataset root
│   ├── mvtec-ad/
│   └── VisA_pytorch/
├── logs/                       # Experiment logs
├── scripts/                    # SLURM/Bash benchmark scripts
│   ├── benchmark_few_shot.sh
│   ├── benchmark_full_shot.sh
│   └── ...
├── src/
│   └── subspacead/
│       ├── core/               # PCA + feature extraction code
│       ├── data/               # Dataset loaders & transforms
│       ├── post_process/       # Scoring, pixel maps, filters
│       └── utils/              # Visualization, logging
├── tools/
│   └── prepare_visa.py
├── main.py
└── README.md

Usage

Benchmark Scripts

Scripts are provided in scripts/ to reproduce all results. Edit MVTEC_PATH and VISA_PATH inside the scripts as needed.

Few-Shot (1, 2, 4 Shots) Selects k normal images per category.

bash scripts/benchmark_few_shot.sh

Batched Zero-Shot Fits PCA on the full unlabeled test set.

bash scripts/benchmark_batched0shot.sh

Full-Shot Uses all training images.

bash scripts/benchmark_full_shot.sh

Manual Execution

Example: 1-shot MVTec-AD on bottle:

python main.py \
    --dataset_name mvtec_ad \
    --dataset_path datasets/mvtec-ad \
    --categories bottle \
    --model_ckpt facebook/dinov2-with-registers-giant \
    --image_res 672 \
    --k_shot 1 \
    --aug_count 30 \
    --pca_ev 0.99 \
    --outdir results/debug_run

Benchmark Results

Few-Shot (Image AUROC)

Setting	Method	MVTec-AD	VisA
1-Shot	PromptAD	94.2	86.9
	AnomalyDINO	96.6	87.4
	SubspaceAD	97.1	93.2
2-Shot	PromptAD	95.7	88.3
	AnomalyDINO	96.9	89.7
	SubspaceAD	97.5	93.8
4-Shot	PromptAD	96.6	89.1
	AnomalyDINO	97.7	92.6
	SubspaceAD	98.0	94.7

Batched Zero-Shot

Method	MVTec-AD	VisA
MuSc	97.8	94.1
AnomalyDINO	94.2	90.7
SubspaceAD	96.6	94.1

Citation

If you find this repository useful, please consider citing:

@inproceedings{lendering2026subspacead,
  title={SubspaceAD: Training-Free Few-Shot Anomaly Detection via Subspace Modeling},
  author={Lendering, Camile and Akdag, Erkut and Bondarev, Egor},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  pages={28557--28566},
  year={2026}
}