CCAD : Compressed Global Feature Conditioned Anomaly Detection

This repository is a reproduction of DiAD (AAAI 2024) and related CCAD experiments. Original Paper: DiAD: A Diffusion-based Framework for Multi-class Anomaly Detection

📂 Repository Structure

• train.py: Main entry point for training and demo runs.
• requirements.txt: Python dependencies.
• models/: Directory for model checkpoints (pretrained weights should be placed here).
• training/: Directory for datasets (MVTec-AD, VisA).

🚀 Quick Start

1. Installation

Create a fresh environment:

conda create -n ccad python=3.8 -y
conda activate ccad
pip install -r requirements.txt

2. Data Preparation

Download the MVTec-AD dataset and organize it as follows:

training/
└── MVTec-AD/
    └── mvtec_anomaly_detection/
        ├── bottle/
        ├── cable/
        └── ...

Ensure train.json and test.json are present in training/MVTec-AD/.

3. Pretrained Models

This code requires pretrained weights. Please download:

• Stable Diffusion v1.5 (v1-5-pruned.ckpt) -> Place in models/
• Autoencoders (if needed) -> Place in models/
• CLIP ViT-L/14 (if running offline) -> Place in models/clip-vit-large-patch14/

(Note: Large weights are ignored by git. You must download them manually.)

4. Running the Demo

To start training/testing:

python train.py