Current anomaly detection methods are primarily designed for low-resolution images. However, in modern industrial production, anomalies often appear as subtle and hard-to-detect defects, making them difficult to identify effectively under low-resolution conditions. To address the industry challenge of large images with small defects, we conducted a systematic study focusing on high-resolution industrial image anomaly detection. We thoroughly analyzed the key challenges of this task, established a comprehensive evaluation benchmark, and proposed HiAD, a practical and efficient high-resolution anomaly detection framework. This framework can accurately detect subtle anomalies in images ranging from 1K to 4K resolution, while ensuring fast inference speeds on mainstream consumer-grade GPUs. If you are a researcher in this field, we invite you to read our paper for more technical details.

2048 × 2048	4096 × 4096

• [09.2025]: Updated DINOv3-based Dinomaly and INP-Former.

$ pip install hiad[cuda11] # for Linux with cuda11 
$ pip install hiad[cuda12] # for Linux with cuda12
$ pip install hiad[cuda]   # for Linux with other cuda versions
$ pip install hiad         # for Windows

_{Since faiss-gpu is not supported on Windows, some features of HiAD may be limited on Windows systems.}

Datasets	Hugging Face	☁️Google Drive
MVTec-2K	XimiaoZhang/MVTec-2K	MVTec-2K.zip
VisA-2K	XimiaoZhang/VisA-2K	VisA-2K.zip
MVTec-4K	XimiaoZhang/MVTec-4K	MVTec-4K.zip

If you would like to reproduce our experiments, please clone our repository and install:

$ git clone https://github.com/cnulab/HiAD.git
$ cd HiAD
$ pip install -e .[cuda11] # for Linux with cuda11 
$ pip install -e .[cuda12] # for Linux with cuda12
$ pip install -e .[cuda]   # for Linux with other cuda versions
$ pip install -e .         # for Windows

Refer to data/README for dataset preparation.

The experiment scripts are located in the runs directory. Run them using the following command:

# taking PatchCore as an example, for 2 GPUs
python runs/run_patchcore.py --data_root data/MVTec-2K --category bottle --gpus 0,1

If you encounter any issues during usage, feel free to open an issue and reach out to us.

If you find it useful, consider giving us a ⭐, we’d really appreciate it!

@inproceedings{zhang2025towards,
      title={Towards High-Resolution Industrial Image Anomaly Detection}, 
      author={Ximiao Zhang, Min Xu, and Xiuzhuang Zhou},
      year={2025},
      eprint={2508.12931},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}