CLIPFUSION

Official Implementation for “CLIP Meets Diffusion: A Synergistic Approach to Anomaly Detection” (TMLR 2025)

Environment

Ubuntu 22.04.4 LTS, CUDA 12.4

Installation

conda create -n clipfusion python==3.12.8
conda activate clipfusion
pip install -r requirements.txt

Running Tests

To run tests, run the following command

bash cls_mvtec.sh   # MVTec-AD Classification
bash seg_mvtec.sh   # MVTec-AD Segmentation
bash cls_visa.sh    # VisA Classification
bash seg_visa.sh    # VisA Classification

For all bash files, default setting is 0-shot inference. To use the few-shot setting, change the ZERO_SHOT variable to false for running of 1,2,4-shot CLIPFUSION.

Data

Download the MVTec-AD dataset from here

Download the VisA dataset from here

In each bash file, modify the DATA_PATH variable to corresponding data folder. For both datasets, data folder configuration is assumed to be as follows:

DATA_PATH/
    object_1/
        ground_truth/
            defect_class_1/
            ...
        test/
            good/
            defect_class_1/
            ...
        train/
            good/
    ...

To preprocess the VisA to the configuration above, follow the official splitting code from here