UniADC

Official Implementation of the paper UniADC: A Unified Framework for Anomaly Detection and Classification.

Introduction

We propose a novel unified framework for anomaly detection and classification, called UniADC, which can accurately detect and classify anomalous regions in images using only a few or even no anomalous images. Moreover, UniADC exhibits powerful anomaly synthesis capabilities, enabling the generation of class-specific anomalous samples guided by anomaly priors or few-shot anomaly examples.

What can UniADC do?

• Anomaly Detection
• Anomaly Localization
• Anomaly Classification (Image-level and Pixel-level)
• Zero-shot Anomaly Synthesis (Training-free, Text-to-image generation)
• Few-shot Anomaly Synthesis (Training-free, Image-to-image generation)

Experiments

Installation

To run experiments, first clone the repository and install requirements.txt.

$ git clone https://github.com/cnulab/UniADC.git
$ cd UniADC
$ pip install -r requirements.txt

Download Pre-trained Models

1. Request DINOv3 weights, and download DINOv3-B/16 (dinov3_vitb16_pretrain_lvd1689m-73cec8be.pth) and DINOv3-based dino.txt (dinov3_vitl16_dinotxt_vision_head_and_text_encoder-a442d8f5.pth).
2. Run sh download.sh to download other dependency weights.

All weight files are placed in the ckpt directory. The complete directory structure is as follows:

  |--ckpt                         
      |--birefnet  #BiRefNet
          |--BiRefNet_config.py
          |--birefnet.py
          |--model.safetensors
          |--...
      |--realisticVisionV60B1_v51VAE      #Stable Diffusion v1.5
          |--feature_extractor
          |--...
      |--random_mask_brushnet_ckpt        #BrushNet
          |--config.json
          |--diffusion_pytorch_model.safetensors
      |--segmentation_mask_brushnet_ckpt  #BrushNet
          |--config.json
          |--diffusion_pytorch_model.safetensors
      |--clip_l14_336_grit_20m_4xe.pth    #AlphaCLIP
      |--dinov3_vitb16_pretrain_lvd1689m-73cec8be.pth  #DINOv3
      |--dinov3_vitl16_dinotxt_vision_head_and_text_encoder-a442d8f5.pth #dino.txt

If you cannot use download.sh for automatic downloading, you can download manually via the following links:

• BiRefNet
• BrushNet and Stable Diffusion
• AlphaCLIP
• DINO

Download Datasets

Note: Please ensure you download the dataset via our provided links. The original MVTec dataset is not supported as it lacks fine-grained anomaly class labels.

Datasets	Hugging Face	☁️Google Drive
MVTec-FS	XimiaoZhang/MVTec-FS	MVTec-FS.zip
WFDD	XimiaoZhang/WFDD	WFDD.zip

Place the downloaded dataset in the data folder.

Zero-shot Anomaly Synthesis (Anomaly Prior Guidance)

Run the following command to experience UniADC's zero-shot anomaly synthesis feature via Gradio:

$ python zero_shot_gen_app.py

Few-shot Anomaly Synthesis (Anomaly Sample Guidance)

Run the following command to experience UniADC's few-shot anomaly synthesis feature via Gradio:

$ python few_shot_gen_app.py

Anomaly synthesis and discriminator training

For zero-shot anomaly detection and classification:

$ sh run_zero.sh

For few-shot anomaly detection and classification:

$ sh run_few.sh

For model testing:

$ python test.py --checkpoint_path <your-checkpoint-path>

Citation

@inproceedings{zhang2025uniadc,
      title={UniADC: A Unified Framework for Anomaly Detection and Classification}, 
      author={Ximiao Zhang, Min Xu, Zheng Zhang, Junlin Hu, and Xiuzhuang Zhou},
      year={2025},
      eprint={2511.06644},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}