winclip-reproduced

Clean reproduction of WinCLIP (Jeong et al., CVPR 2023) on MVTec-AD and VisA — zero-shot and few-normal-shot anomaly classification and segmentation with CLIP.

Paper: WinCLIP: Zero-/Few-Shot Anomaly Classification and Segmentation - Jeong et al., CVPR 2023

What this is

End-to-end reproduction of WinCLIP, the first strong CLIP-based zero-shot anomaly detector on MVTec-AD and VisA. Implements the compositional prompt ensemble, multi-scale window aggregation, and the few-normal-shot extension WinCLIP+.

This is the second reproduction in a series of visual anomaly detection methods I'm building toward UniVAD (CVPR 2025), my main thesis baseline. WinCLIP is one of UniVAD's six comparison methods in Table 1.

Status

In active development. Code lands over the coming week, results follow once GPU access returns.

Why WinCLIP

WinCLIP is one of two foundation-model-based methods on UniVAD's comparison list (alongside AnomalyGPT). It shifted visual anomaly detection from "train a custom model per object" to "use a pretrained vision-language model with the right prompts" making it the conceptual bridge between PatchCore-era frozen-ImageNet-backbone methods and the current foundation-model era of AD.

	Zero-shot Image-AUROC (MVTec-AD)	1-shot Image-AUROC (MVTec-AD)
WinCLIP	91.8	—
WinCLIP+	—	93.1

(Numbers from the WinCLIP paper.)

Goal

Match the paper's reported numbers within ±0.5 points:

Setting	Dataset	Image-AUROC	Pixel-AUROC
Zero-shot	MVTec-AD	91.8 → TBD	85.1 → TBD
Zero-shot	VisA	78.1 → TBD	79.6 → TBD
1-normal-shot	MVTec-AD	93.1 → TBD	95.2 → TBD
1-normal-shot	VisA	83.8 → TBD	96.4 → TBD

Installation

Clone and create the environment:

git clone https://github.com/hammadhaideer/winclip-reproduced.git
cd winclip-reproduced
conda env create -f environment.yml
conda activate winclip

Or with pip in an existing environment:

pip install -r requirements.txt

Datasets

This repo evaluates on MVTec-AD and VisA. Download both, then point the code to a parent directory containing them:

export WINCLIP_DATA_ROOT=/path/to/anomaly_datasets

Expected layout:

$WINCLIP_DATA_ROOT/
├── mvtec_ad/
│   ├── bottle/
│   │   ├── train/good/*.png
│   │   ├── test/<defect>/*.png
│   │   └── ground_truth/<defect>/*_mask.png
│   ├── cable/
│   └── ... (15 categories)
└── visa/
    ├── candle/
    ├── ... (12 categories)
    └── split_csv/1cls.csv

Dataset	Source	Categories
MVTec-AD	https://www.mvtec.com/company/research/datasets/mvtec-ad	15
VisA	https://github.com/amazon-science/spot-diff	12

VisA is distributed with a split CSV file (split_csv/1cls.csv) that defines the train/test partition. The loader reads from this CSV — make sure it's present at the path above.

Run

Single category, zero-shot:

python scripts/run_winclip.py --config configs/default.yaml \
    --dataset mvtec_ad --category bottle --shot 0

Single category, 1-normal-shot:

python scripts/run_winclip.py --config configs/default.yaml \
    --dataset mvtec_ad --category bottle --shot 1

Sweep all categories on both datasets, both 0-shot and 1-shot (~1 hour on a single GPU):

python scripts/run_all.py --datasets mvtec_ad visa --shots 0 1

Full paper reproduction (108 runs, ~1 hour total):

python scripts/run_all.py --datasets mvtec_ad visa --shots 0 1 2 4

Build the comparison table after runs complete:

python scripts/aggregate_results.py --results_dir results --shots 0 1

Per-experiment results land in results/<dataset>_<category>_<shot>shot.json. The aggregated table prints to stdout, comparing your numbers against the paper's reported values per (dataset, shot) combination.

Roadmap

• Repo scaffold, configs, dataset loader (reused from patchcore-reproduced)
• CLIP ViT-B/16+ backbone integration via open_clip
• Compositional prompt ensemble (state words × prompt templates)
• Multi-scale window aggregation (window/patch/image-level)
• Zero-shot anomaly classification and segmentation
• WinCLIP+ few-normal-shot extension with reference association
• Empirical reproduction on MVTec-AD and VisA
• Walkthrough notebook with prompt-ensemble visualization
• Medium walkthrough post

Reproduction series

Part of a broader series reproducing UniVAD's full comparison set:

• patchcore-reproduced — PatchCore (CVPR 2022)
• winclip-reproduced — WinCLIP (CVPR 2023) ← this repo
• uniad-reproduced — UniAD (NeurIPS 2022)
• anomalygpt-reproduced — AnomalyGPT (AAAI 2024)
• comad-reproduced — ComAD (PR 2024)
• medclip-reproduced — MedCLIP (EMNLP 2022)
• univad-reproduced — UniVAD (CVPR 2025) ← target

References

• Jeong et al., WinCLIP: Zero-/Few-Shot Anomaly Classification and Segmentation, CVPR 2023
• Radford et al., Learning Transferable Visual Models From Natural Language Supervision (CLIP), ICML 2021
• Bergmann et al., MVTec-AD, CVPR 2019
• Zou et al., VisA, ECCV 2022

License

MIT