rashidrao/AD_MVTec

VAE-GAN Checkpoints for MVTec Anomaly Detection

This repository contains pre-trained VAE-GAN model checkpoints for visual anomaly detection on the MVTec AD dataset.

Overview

The models were trained in a one-class anomaly detection setting using only normal training images. During inference, each input image is reconstructed by the VAE-GAN model, and anomaly scores are computed from the reconstruction error between the input and reconstructed image.

These checkpoints are useful for:

• —Reconstruction-based anomaly detection
• —Threshold selection experiments
• —Multi-point threshold evaluation
• —Anomaly localization
• —Explainable anomaly detection
• —Baseline comparison with AE, VAE, PatchCore, PaDiM, and other methods

Dataset

The checkpoints are trained on MVTec AD object and texture categories.

Reference:

@article{bergmann2021mvtec,
  title={The MVTec Anomaly Detection Dataset: A Comprehensive Real-World Dataset for Unsupervised Anomaly Detection},
  author={Bergmann, Paul and Batzner, Kilian and Fauser, Michael and Sattlegger, David and Steger, Carsten},
  journal={International Journal of Computer Vision},
  year={2021}
}

Available Checkpoints

Model Details

Checkpoint Structure

Each checkpoint contains:

{
    "encoder_state_dict": ...,
    "decoder_state_dict": ...,
    "discriminator_state_dict": ...
}

Loading a Checkpoint

import torch

checkpoint = torch.load(
    "model_bottle_64.pt",
    map_location="cpu",
    weights_only=False
)

encoder.load_state_dict(checkpoint["encoder_state_dict"])
decoder.load_state_dict(checkpoint["decoder_state_dict"])
discriminator.load_state_dict(checkpoint["discriminator_state_dict"])

encoder.eval()
decoder.eval()
discriminator.eval()

Example Anomaly Score

with torch.no_grad():
    mu, logvar = encoder(image)
    z = reparameterize(mu, logvar)
    reconstruction = decoder(z)

anomaly_map = torch.abs(image - reconstruction).mean(dim=1)
anomaly_score = anomaly_map.mean().item()

Device Support

The checkpoints can be loaded on:

• —CPU
• —NVIDIA CUDA GPUs
• —Apple Silicon (MPS)

import torch

if torch.cuda.is_available():
    device = "cuda"
elif torch.backends.mps.is_available():
    device = "mps"
else:
    device = "cpu"

Intended Use

This repository is intended for research on:

• —Visual Anomaly Detection
• —Reconstruction-Based Anomaly Scoring
• —Threshold Calibration
• —Multi-Point Thresholding
• —Explainable Anomaly Detection
• —Industrial Inspection Systems

Limitations

• —Models are trained on resized 128×128 images.
• —Performance depends on preprocessing, anomaly score design, and threshold selection.
• —These checkpoints are intended for research purposes and should be validated before deployment in industrial or safety-critical environments.

Citation

@misc{rao2026vaeganmvtec,
  title={VAE-GAN Checkpoints for MVTec Anomaly Detection},
  author={Rao, Rashid},
  year={2026},
  publisher={Hugging Face},
  url={https://huggingface.co/rashidrao/AnomalyDetection}
}

Author

Rashid Rao Industrial PhD Researcher University of Turin, Italy

Research Areas:

• —Explainable AI (XAI)
• —Visual Anomaly Detection
• —Trustworthy AI
• —Industrial AI Systems