[AAAI 2026] AnoStyler: Text-Driven Localized Anomaly Generation via Lightweight Style Transfer

Yulim So¹, Seokho Kang¹
¹Sungkyunkwan University

---

Abstract

Anomaly generation has been widely explored to address the scarcity of anomaly images in real-world data. However, existing methods typically suffer from at least one of the following limitations, hindering their practical deployment: (1) lack of visual realism in generated anomalies; (2) dependence on large amounts of real images; and (3) use of memory-intensive, heavyweight model architectures. To overcome these limitations, we propose AnoStyler, a lightweight yet effective method that frames zero-shot anomaly generation as text-guided style transfer. Given a single normal image along with its category label and expected defect type, an anomaly mask indicating the localized anomaly regions and two-class text prompts representing the normal and anomaly states are generated using generalizable category-agnostic procedures. A lightweight U-Net model trained with CLIP-based loss functions is used to stylize the normal image into a visually realistic anomaly image, where anomalies are localized by the anomaly mask and semantically aligned with the text prompts. Extensive experiments on the MVTec-AD and VisA datasets show that AnoStyler outperforms existing anomaly generation methods in generating high-quality and diverse anomaly images. Furthermore, using these generated anomalies helps enhance anomaly detection performance.

---

Installation

Create a new conda environment:

conda create -n anostyler python=3.10
conda activate anostyler
pip install -r requirements.txt

Or install manually:

pip install torch==2.5.1 torchvision==0.20.1
pip install numpy==2.3.2 opencv-contrib-python==4.11.0.86 Pillow==11.3.0 scipy==1.16.1 tqdm==4.67.1
pip install git+https://github.com/openai/CLIP.git
pip install git+https://github.com/facebookresearch/segment-anything.git

---

Dataset & Checkpoint Download

We provide a helper script at script/download.sh (edit the links to your mirrors):

# /path/to/AnoStyler/script/download.sh

# Download MVTec-AD dataset
mkdir -p datasets/MVTec-AD && cd datasets/MVTec-AD
wget <download_link_to_MVTec_AD_dataset>
cd -

# Download VisA dataset
mkdir -p datasets/VisA && cd datasets/VisA
wget <download_link_to_VisA_dataset>
cd -

# Download SAM model checkpoint
mkdir -p checkpoints && cd checkpoints
wget <download_link_to_sam_checkpoint>
cd -

Usage:

bash script/download.sh

Datasets:

• MVTec-AD (Bergmann et al., CVPR 2019)
• VisA (Zou et al., ECCV 2022)

After download, expected tree:

AnoStyler/
├─ datasets/
│  ├─ MVTec-AD/
│  └─ VisA/
└─ checkpoints/
   └─ sam_vit_b_01ec64.pth

---

Quick Start (CLI)

The main entry is main.py, which reads a YAML config and generates anomalies.

Config template (save as config.yaml):

Run:

# default reads ./config.yaml
python main.py

# or specify a path
python main.py --config configs.yaml

This will:

• optionally compute a SAM foreground mask,
• procedurally generate meta‑shape masks until non‑empty,
• run style transfer and save results to <save_path>/<category>/<defect>/image/ and .../mask/.

---

Demo (Notebook)

A minimal demo is provided at demo/run_demo.ipynb.

Folder layout:

demo/
├─ images/
│  ├─ normal_image_ex1.png
│  └─ normal_image_ex2.png
└─ results/
   ├─ gen_ano.jpg
   └─ gen_mask.jpg

How to use:

1. Put a normal image into demo/images/ (e.g., normal_image_ex1.png).
2. Open and run demo/run_demo.ipynb.
3. Generated anomaly and mask will be saved to demo/results/ as gen_ano.jpg, gen_mask.jpg.

---

Project Structure

AnoStyler/
├─ datasets/
├─ demo/
│  ├─ images/
│  ├─ results/
│  └─ run_demo.ipynb
├─ figures/
│  ├─ framework.pdf
│  └─ result.pdf
├─ script/
│  └─ download.sh
├─ src/
│  ├─ StyleNet.py
│  ├─ def_train.py
│  ├─ meta_shape_priors.py
│  ├─ sam.py
│  ├─ two_class_prompt_template.py
│  └─ utils.py
├─ checkpoints/
├─ configs.yaml
├─ main.py
└─ README.md
├─ requirements.txt

---

Results

AnoStyler outperforms existing zero-shot methods on MVTec-AD and VisA datasets.

---

Citation

If you find this repository useful, please cite:

@inproceedings{so2026anostyler,
  title={AnoStyler: Text-Driven Localized Anomaly Generation via Lightweight Style Transfer},
  author={So, Yulim and Kang, Seokho},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  year={2026}
}

---

Acknowledgement

Thanks for the excellent inspiration from CLIPstyler.