RetriVAD: Training-Free Unified Visual Anomaly Detection via Single-Encoder Retrieval

---

Introduction

RetriVAD is a training-free unified visual anomaly detection method that detects anomalies across industrial, logical, and medical imaging domains using a single frozen encoder, without any domain-specific training, fine-tuning, or GPU.

Given a category, RetriVAD builds a memory bank offline by encoding normal reference images into 768-dimensional feature vectors stored in a FAISS index. At test time, the anomaly score for a query image is its nearest-neighbour distance to the memory bank. No thresholds are tuned and no labels are used. The same index generalises across industrial products, brain MRI, chest X-rays, and retinal images.

Pixel-level localisation requires 256 sequential DINOv2 forward passes, one per 14x14-pixel crop from the 16x16 spatial grid, producing a 16x16 distance map which is bilinearly upsampled to the original image resolution.

---

Results

Image-Level AUROC (1-shot setting)

Evaluated on 6 datasets spanning industrial, logical, and medical domains. All baseline numbers are taken from UniVAD (CVPR 2025) under the identical 1-normal-shot protocol.

Dataset	PatchCore	AnomalyGPT	WinCLIP	ComAD	UniAD	MedCLIP	UniVAD	RetriVAD
MVTec-AD	84.0	94.1	93.1	57.3	70.3	75.2	97.8	89.6
VisA	74.8	87.4	83.8	53.9	61.3	69.0	93.5	79.1
MVTec LOCO	62.0	60.4	58.0	62.2	50.9	54.9	71.0	72.1
BrainMRI	73.2	73.1	55.4	33.3	50.0	69.7	80.2	95.3
RESC	56.3	82.4	72.9	73.5	53.5	66.9	85.5	76.5
ChestXray	66.4	68.5	70.2	50.1	60.6	71.4	72.2	86.8
Mean	69.5	77.6	72.2	55.1	57.8	67.9	83.4	83.2

RetriVAD wins on 3 out of 6 datasets against UniVAD (CVPR 2025), matching its mean performance with a single frozen encoder and no GPU. VisA results use the official CSV split protocol.

Pixel-Level AUROC

Dataset	Pixel-AUROC
VisA (12 categories)	91.92
MVTec-AD (15 categories)	71.27
MVTec LOCO (5 categories)	56.21

Pixel-level evaluation is reported on datasets with ground-truth pixel masks only.

---

Environment Setup

Clone the repository:

git clone https://github.com/hammadhaideer/RetriVAD.git
cd RetriVAD

Create and activate the conda environment:

conda create -n retrivad python=3.10
conda activate retrivad

Install PyTorch (CPU):

pip install torch==2.10.0+cpu --index-url https://download.pytorch.org/whl/cpu

Install remaining dependencies:

pip install -r requirements.txt

Core dependencies:

torch==2.10.0+cpu
faiss-cpu==1.13.2
scikit-learn==1.7.2
numpy==2.2.6
scipy
Pillow
tqdm

Verify installation:

python -c "import torch, faiss, numpy; print('torch:', torch.__version__); print('faiss: OK'); print('numpy:', numpy.__version__)"

---

Data Preparation

Datasets are not included in this repository. Download each dataset and place it at the corresponding path.

MVTec-AD

Download from https://www.mvtec.com/company/research/datasets/mvtec-ad and extract to:

data/mvtec/
├── bottle/
├── cable/
└── ...

VisA

Download from https://amazon-visual-anomaly.s3.us-west-2.amazonaws.com/VisA_20220922.tar and extract to:

data/VisA_20220922/
├── candle/
├── capsules/
└── ...

MVTec LOCO

Download from https://www.mvtec.com/company/research/datasets/mvtec-loco and extract to:

data/mvtec_loco/
├── breakfast_box/
├── pushpins/
└── ...

Medical Datasets

Dataset	Source
BrainMRI	Kaggle
RESC	GitHub
ChestXray	Kaggle

---

Running Evaluation

Image-Level Evaluation

# All industrial and logical datasets
python benchmark.py --dataset mvtec --data_root /path/to/mvtec
python benchmark.py --dataset visa --data_root /path/to/VisA_20220922
python benchmark.py --dataset loco --data_root /path/to/mvtec_loco

# Medical datasets
python eval_brainmri.py --data_root /path/to/brain_mri
python eval_chestxray.py --data_root /path/to/chest_xray
python eval_resc.py --data_root /path/to/resc/RESC

Pixel-Level Evaluation

# MVTec-AD and MVTec LOCO
python scripts/eval_pixel_auroc.py --dataset mvtec --data_root /path/to/mvtec

# VisA
python eval_pixel_auroc_visa.py --data_root /path/to/VisA_20220922

Pixel-level evaluation is computationally intensive on CPU-only hardware. All results are pre-computed and saved in results/. Re-running is only necessary if the model code changes.

Latency Benchmark

python scripts/latency_benchmark.py --data_root /path/to/mvtec --category bottle --n_test 20

Qualitative Figure

python generate_heatmaps.py \
    --carpet_normal  /path/to/mvtec/carpet/train/good/000.png \
    --carpet_test    /path/to/mvtec/carpet/test/color/000.png \
    --carpet_gt      /path/to/mvtec/carpet/ground_truth/color/000_mask.png \
    --bottle_normal  /path/to/mvtec/bottle/train/good/000.png \
    --bottle_test    /path/to/mvtec/bottle/test/broken_large/000.png \
    --bottle_gt      /path/to/mvtec/bottle/ground_truth/broken_large/000_mask.png \
    --brain_normal   /path/to/brain_mri/Training/notumor/Tr-no_999.jpg \
    --brain_test     /path/to/brain_mri/Testing/glioma/Te-gl_1.jpg \
    --visa_normal    /path/to/VisA_20220922/pcb1/Data/Images/Normal/0000.JPG \
    --visa_test      /path/to/VisA_20220922/pcb1/Data/Images/Anomaly/009.JPG \
    --visa_gt        /path/to/VisA_20220922/pcb1/Data/Masks/Anomaly/009.png \
    --output         figures/fig_qualitative.png

---

Pre-computed Results

File	Contents
results/image_auroc_results.json	Image-level AUROC across all 6 datasets
results/mvtec_ad_retrivad.json	Per-category results for MVTec-AD
results/visa_retrivad.json	Per-category results for VisA
results/mvtecloco_retrivad.json	Per-category results for MVTec LOCO
results/pixel_auroc_final.json	Pixel-level AUROC for MVTec-AD and MVTec LOCO
results/visa_pixel_auroc.json	Pixel-level AUROC for VisA (12 categories)
results/latency_results.json	Per-image latency measured on CPU

---

Repository Structure

RetriVAD/
├── README.md
├── requirements.txt
├── benchmark.py
├── generate_heatmaps.py
├── ablation_features.py
├── eval_brainmri.py
├── eval_chestxray.py
├── eval_resc.py
├── eval_pixel_auroc_visa.py
├── eval_pixel_auroc_resc.py
├── models/
│   └── retrivad.py
├── utils/
│   └── metrics.py
├── scripts/
│   ├── eval_pixel_auroc.py
│   └── latency_benchmark.py
├── results/
└── figures/

---

Citation

@article{haider2026retrivad,
  title={RetriVAD: Training-Free Unified Visual Anomaly Detection via Single-Encoder Retrieval},
  author={Haider, Hammad Ali and Zheng, Panpan},
  year={2026}
}