SubspaceAD

This repository contains the official implementation of the paper:

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

---

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: 98.0% Image AUROC; 97.6% Pixel AUROC
• VisA: 93.3% Image AUROC; 98.3% 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.

🎁 Project Highlights:

1. Training method upgraded 🎉 to DINOv3
2. Added save_model.py for PCA parameter model persistence and inference.py for convenient model inference
3. Added comprehensive project documentation and usage guides

🚀 Quick Start

Environment Setup

# 1. Create environment
uv venv
source venv/bin/activate

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

Download DINO Model

pip install modelscope

mkdir -p DINO

modelscope download --model facebook/dinov3-vitl16-pretrain-lvd1689m --local_dir ./DINO

Training on Custom Data:

Dataset Preparation:

Training Directory Requirements:

• The directory should contain normal (defect-free) sample images
• Supported image formats: .png, .jpg
• Images will be automatically shuffled, then trained using the number specified by the --k_shot parameter

Validation Set Support: If you need to use a validation set to calculate detection thresholds in a custom directory, you can create a val/ directory at the same level as the training directory:

your_datasets/
├── train/    # Training images (normal samples)
│   ├── 001.png
│   ├── 002.png
│   └── ...
└── val/                    # Validation set (optional)
    ├── good/               # Normal samples (for threshold calculation)
    │   ├── 001.png
    │   └── ...
    └── defect/             # Defective samples (optional)
        └── ...

🎯 Training:

python save_model.py \
    --dataset_name phone_screen \
    --train_dir your_datasets \
    --model_ckpt DINO/dinov3-vitl16-pretrain-lvd1689m \
    --image_res 1024 \
    --k_shot 10 \
    --layers="-7,-8,-9,-10" \
    --aug_count 10 \
    --aug_list rotate hflip vflip affine \
    --pca_ev 0.99 \
    --score_method reconstruction \
    --img_score_agg mtop1p

🍭 Model Inference

python inference.py \
    --model_path models/phone_screen_train_k10 \
    --image_path datasets/phone_screen/val/defect_croped \
    --output_dir results/inference \
    --visualize \
    --save_summary

Benchmark Results

Few-Shot (Image AUROC)

Setting	Method	MVTec-AD	VisA
1-Shot	PromptAD	94.2	86.9
	AnomalyDINO	96.6	87.4
	SubspaceAD	98.0	93.3
2-Shot	PromptAD	95.7	88.3
	AnomalyDINO	96.9	89.7
	SubspaceAD	98.1	94.1
4-Shot	PromptAD	96.6	89.1
	AnomalyDINO	97.7	92.6
	SubspaceAD	98.4	94.5

Batched Zero-Shot

Method	MVTec-AD	VisA
MuSc	97.8	94.1
AnomalyDINO	94.2	90.7
SubspaceAD	96.6	97.7

Citation

If you find this repository useful, please consider citing:

@misc{lendering2026subspaceadtrainingfreefewshotanomaly,
      title={SubspaceAD: Training-Free Few-Shot Anomaly Detection via Subspace Modeling}, 
      author={Camile Lendering and Erkut Akdag and Egor Bondarev},
      year={2026},
      eprint={2602.23013},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2602.23013}, 
}