One-to-More: High-Fidelity Training-Free Anomaly Generation with Attention Control

Installation

1. Create and activate a new Conda environment:

conda env create -f environment.yml
conda activate O2MAG

Mask Generation

We use AnomalyDiffusion masks for MVTec-AD and SeaS for VisA/Real-IAD.

We have released the generated image-mask pairs for MVTec-AD.

Generated 500 image-mask pairs: Google Drive Link

Normal Data Augmentation

See Appendix C.1. for detailed information. We encourage you to explore other augmentation strategies to achieve better performance.

python ./img_augment.py

Anomaly Generation

We provide three ways to run and evaluate our anomaly generation code:

1. Interactive Web UI (Requires approx. 14GB VRAM)

   Quickly edit and visualize anomalies in your browser.

python ./app_edit_anomaly_mask.py

---

2. Jupyter Notebook (Requires approx. 16GB VRAM)

edit_anomaly_mask.ipynb

---

3. Generate 1000 anomaly images per anomaly type. About 24G.

python edit_anomaly_moregpu_fewshot.py --root ./datasets/mvtec \
    --normal_path ./data_agument
    --sourece_image_mask ./anomalydiffusion/generated_mask \
    --embedding_file ./embed_bank/mvtec \
    --outputs_path ./generated_data_fewshot/mvtec \
    --pairs-file ./anomaly_name/name-mvtec.txt --devices cuda:2,cuda:3,cuda:4,cuda:5

• --root: Path to the MVTec-AD dataset.
• --normal_path: Path to the augmented normal data.
• --sourece_image_mask: Directory containing the generated anomaly masks.
• --pairs-file : Specifies the object category and anomaly type to generate (e.g., 'cable+combined').
• --devices : Specifies the GPUs to be used. For single-GPU execution, use --devices cuda:0, (note the trailing comma).

Attention Map Visualization

To visualize the attention maps, please follow these steps:

1. Enable Attention Storage: In triag/mca_p2p.py, change AttentionBase to AttentionStore. (Note: This process requires approx. 32GB of VRAM).
2. Run the Visualization Script: Execute the visualization_attention_map.py script. It is ready to run out-of-the-box, but please ensure you update the following variables inside the script to match your local environment: model_path and paths for the corresponding reference anomaly and normal images.

Evaluation

Compute KID

python eval/compute-kid.py --generated_path $path_to_the_generated_data  --real_path=$path_to_mvtec

Anomaly detection

Train U-Net

python train-localization.py \
  --mvtec_path=$path_to_mvtec \
  --generated_data_path=$path_to_the_generated_data \
  --save_path=$path_to_save_checkpoint

Test

python test-localization.py \
  --mvtec_path=$path_to_mvtec \
  --checkpoint_path=$path_to_save_checkpoint

Anomaly classification

Train ResNet-34

python train-classification.py --mvtec_path $path_to_mvtec \
 --generated_data_path $path_to_the_generated_data \
 --checkpoint_path $path_to_save_checkpoint

Test

python test-classification.py --mvtec_path $path_to_mvtec \
 --generated_data_path $path_to_the_generated_data \
 --checkpoint_path $path_to_save_checkpoint

Results of anomaly image generation

The generation results of anomaly images and normal images are shown as follows:

Citation

If you find our work useful in your research, please consider citing our paper.

@misc{rao2026onetomorehighfidelitytrainingfreeanomaly,
      title={One-to-More: High-Fidelity Training-Free Anomaly Generation with Attention Control}, 
      author={Haoxiang Rao and Zhao Wang and Chenyang Si and Yan Lyu and Yuanyi Duan and Fang Zhao and Caifeng Shan},
      year={2026},
      eprint={2603.18093},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2603.18093}, 
}