Xijun Lu · Hongying Liu · Fanhua Shang · Yanming Hui · Liang Wan
Tianjin University · Medical School & College of Intelligence and Computing

CVPR 2026

Code | Paper | Project Page

PDD is a medical image anomaly detection framework based on manifold-prior diverse distillation. It uses two frozen teachers, VMamba-Tiny for global context and Wide-ResNet50-2 for local structure, and trains a lightweight PDD decoder to reconstruct and align multi-level features.

This repository is a cleaned training subset refactored from VAD-TS. It contains the single-class DDP training script, the multi-class DDP training script, dataset utilities, evaluation utilities, visualization code, and model definitions. Model weights are not stored in this project.

News 🎉

• [2026/02/21] PDD is accepted to CVPR 2026.
• [2026/08/24] Training code is now available.

Project Structure

PDD/
  models/
    mamba_decoder.py          # PDD decoder, entry: pdd_decoder
    resnet_encoder.py         # Wide-ResNet50-2 encoder
    vmamba.py                 # VMamba timm registration
  scripts/
    train_two_twins_ddp.py    # single-class DDP training
    train_two_twins_ddp_multi.py
                               # multi-class DDP training
    vis_eval.py               # test-set inference and visualization
  utils/
    dataset.py                # single-class dataset
    dataset_full.py           # multi-class meta.json dataset
    evaluation.py             # evaluation helpers
    losses.py                 # training losses
    utils.py                  # common helpers
  static/                     # README figures
  checkpoints/                # local checkpoint output, ignored by git
  logs/                       # TensorBoard output, ignored by git

Environment

Create the environment from the exported conda file:

cd /PDD
conda env create -f environment.yml
conda activate PDD

Or install from pip requirements if you already have a compatible CUDA/PyTorch environment:

# Our env. version is Python 3.10.18
conda create -n PDD python==3.10
pip install -r requirements.txt

Data

Single-Class Data

The single-class script expects an MVTec-like layout:

data_path/
  train/
    good/
      xxx.png
  test/
    good/
      xxx.png
    anomaly_type/
      xxx.png

Image suffixes are matched case-insensitively for common formats such as png, jpg, jpeg, and jepg.

Multi-Class Data

The multi-class script uses utils/dataset_full.py and expects a meta.json file under data_path:

data_path/
  meta.json
  class_a/...
  class_b/...

meta.json should contain train and test sections. Each class contains a list of records with at least:

{
  "img_path": "relative/path/to/image.png",
  "anomaly": 0
}

For training, only records with "anomaly": 0 are used.

Training

Single-Class DDP

CUDA_VISIBLE_DEVICES=4,5 torchrun --nproc_per_node=2 scripts/train_two_twins_ddp.py \
  --data_path /your/path/data/head_ct \
  --save_path /your/path/PDD/checkpoints/head_ct

Multi-Class DDP

CUDA_VISIBLE_DEVICES=4,5 torchrun --nproc_per_node=2 scripts/train_two_twins_ddp_multi.py \
  --data_path /your/path/data/medical \
  --class_list brain,liver,retinal \
  --save_path /your/path/PDD/checkpoints/multi_tao_0375

Default training settings follow the tao_0375 branch:

• res=10
• layerloss=4
• decoder entry: models.mamba_decoder.pdd_decoder

Visualization

Run test-set inference and save original image, heatmap, overlay, and comparison images:

CUDA_VISIBLE_DEVICES=5 python scripts/vis_eval.py \
  --device cuda:0 \
  --output_dir /your/path/PDD/vis/head_ct

The checkpoint path and default head_ct data path are currently defined in scripts/vis_eval.py.

Outputs

Training writes:

• checkpoints to checkpoints/
• TensorBoard logs to logs/

Visualization writes:

• original images to vis/*/org/
• heatmaps to vis/*/heatmap/
• overlays to vis/*/overlay/
• side-by-side comparisons to vis/*/compare/

Notes

• No trained weights are committed in this repository.
• Teacher pretrained weights are loaded through the model definitions and the local PyTorch/timm cache.
• Use torchrun for DDP training. CUDA_VISIBLE_DEVICES controls which physical GPUs are visible; inside the script each process uses its local rank.

Citation

If you find our code or paper useful, please cite:

@inproceedings{lu2026pdd,
  title={PDD: Manifold-Prior Diverse Distillation for Medical Anomaly Detection},
  author={Lu, Xijun and Liu, Hongying and Shang, Fanhua and Hui, Yanming and Wan, Liang},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={28534--28544},
  year={2026}
}

Acknowledgement

This project builds on VMamba, RD4AD. We thank the authors for their excellent work:

• VMamba
• RD4AD