Clean reproduction of UniAD (You et al., NeurIPS 2022 Spotlight) on MVTec-AD - multi-class anomaly detection and localization with one unified transformer-based reconstruction model.
Paper: A Unified Model for Multi-class Anomaly Detection - You et al., NeurIPS 2022 Spotlight
Official repo: https://github.com/zhiyuanyou/UniAD
End-to-end reimplementation of UniAD: a frozen EfficientNet-b4 backbone, multi-stage feature concatenation at 14×14, and a transformer encoder-decoder with neighbor-masked attention, layer-wise learnable queries, and feature jittering. One model is trained jointly across all 15 MVTec-AD categories - the unified one-for-all setting that defines the paper.
This is the third reproduction in a series of visual anomaly detection methods I'm building toward UniVAD (CVPR 2025), the current state-of-the-art training-free unified VAD method. UniAD is the canonical unified-model baseline that every later "one model for all categories" paper compares against, including UniVAD.
In active development. Code lands over the coming week, results follow once GPU access returns.
UniAD is the paper that put "one model for all categories" on the map. Earlier methods (PatchCore, EfficientAD) train a separate model per category, which is fine for benchmarks but unrealistic for a factory with hundreds of products. UniAD trains a single transformer across every category at once and shows you can match or beat per-category baselines while serving the whole catalog from one checkpoint.
To get there, the paper diagnoses an "identical shortcut" failure mode in plain reconstruction networks - they trivially copy anomalies through to the output - and fixes it with three design choices:
- Neighbor-masked attention stops each token from attending to itself or its immediate neighbors, forcing reconstruction from non-local context.
- Layer-wise learnable queries in the decoder use a fresh query set per layer instead of propagating decoded tokens, blocking input leakage.
- Feature jittering perturbs encoder input tokens during training so the model recovers the underlying normal pattern instead of memorizing exact features.
| Detection AUROC | Localization AUROC | |
|---|---|---|
| UniAD (paper, unified setting) | 96.7 | 96.8 |
(Numbers from the UniAD paper, mean across 15 MVTec-AD categories under the unified one-for-all setting.)
Match the paper's reported numbers within ±0.5 points on MVTec-AD (mean across 15 categories) under the unified setting:
| Setting | Dataset | Detection AUROC | Localization AUROC |
|---|---|---|---|
| Unified | MVTec-AD | 96.7 → TBD | 96.8 → TBD |
Per-category table will land here as runs complete.
Clone and create the environment:
git clone https://github.com/hammadhaideer/uniad-reproduced.git
cd uniad-reproduced
conda env create -f environment.yml
conda activate uniad
Or with pip in an existing environment:
pip install -r requirements.txt
Download MVTec-AD from the official page (registration required, free for non-commercial use). Extract it anywhere on your machine, then point the code to it via an environment variable:
export UNIAD_DATA_ROOT=/path/to/mvtec_ad
Expected layout:
$UNIAD_DATA_ROOT/
├── bottle/
│ ├── train/good/*.png
│ ├── test/<defect>/*.png
│ └── ground_truth/<defect>/*_mask.png
├── cable/
└── ... (15 categories)
| Dataset | Source | Categories |
|---|---|---|
| MVTec-AD | https://www.mvtec.com/company/research/datasets/mvtec-ad | 15 |
Train a unified model across all 15 categories (default: 1000 epochs, batch size 8, single GPU):
python scripts/train.py --config configs/default.yaml
Evaluate a checkpoint:
python scripts/eval.py --config configs/default.yaml \
--checkpoint checkpoints/uniad_epoch1000.pth \
--out results/results.json
Build the comparison table after the run completes:
python scripts/aggregate_results.py --results results/results.json
Per-category results land in results/results.json. The aggregated table prints to stdout, comparing your numbers against the paper's.
For an interactive end-to-end pipeline check on a single category - feature shapes, neighbor mask visualization, and qualitative heatmaps before/after training:
jupyter notebook notebooks/01_walkthrough.ipynb
Run cells top-to-bottom. The notebook reads the same config and dataset path as the scripts.
- Repo scaffold, configs, dataset loader
- EfficientNet-b4 backbone + multi-stage feature concatenation
- Transformer with neighbor-masked attention + layer-wise queries + feature jittering
- Training loop with AMP, cosine schedule, warmup
- Image-AUROC, pixel-AUROC, AUPRO metrics
- Eval runner and results aggregation script
- Walkthrough notebook with neighbor-mask and heatmap visualizations
- Empirical reproduction across all 15 categories (pending GPU)
- Qualitative heatmap figures
- Medium walkthrough post
Part of a broader series reproducing UniVAD's full comparison set:
-
patchcore-reproduced- PatchCore (CVPR 2022) -
winclip-reproduced- WinCLIP (CVPR 2023) -
uniad-reproduced- UniAD (NeurIPS 2022) ← this repo -
anomalygpt-reproduced- AnomalyGPT (AAAI 2024) -
comad-reproduced- ComAD (PR 2024) -
medclip-reproduced- MedCLIP (EMNLP 2022) -
univad-reproduced- UniVAD (CVPR 2025) ← target
- You et al., A Unified Model for Multi-class Anomaly Detection, NeurIPS 2022 (Spotlight) - arXiv:2206.03687
- Tan & Le, EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks, ICML 2019
- Bergmann et al., MVTec-AD, CVPR 2019
- Official UniAD repo: https://github.com/zhiyuanyou/UniAD
MIT