Contamination Map Methodology

Vision / Model

WinCLIP / WinCLIP+

Organisation
Amazon AWS AI Labs / University of Wisconsin-Madison
Family
WinCLIP
Released
2023-03-26 (arXiv v1 submission date (https://arxiv.org/abs/2303.14814))
Weights
Unknown
Links
Paper

Training data disclosure: Full list

Level determined for 50% of benchmark cells.

“We emphasize that our zero-shot models do not require any tuning for individual cases, and the few-normal-only setup does not use any segmentation annotation”
arxiv.org · paper · Section 1 · retrieved 2026-10-15
“We adopt the CLIP implementation of OpenCLIP3 and its public pre-trained models in our”
arxiv.org · paper · Section 5, Implementation details · retrieved 2026-10-15

Built on

Everything these base models may have seen is inherited by WinCLIP / WinCLIP+.

Benchmark cells

Industrial anomaly detection

Unknown MVTec AD · unknown for its base model OpenCLIP (LAION-400M checkpoints): training data not disclosed in enough detail

Not seen MVTec AD 2 · absent from the complete list of training data and from its base models' data

Evidence (6)
“We emphasize that our zero-shot models do not require any tuning for individual cases, and the few-normal-only setup does not use any segmentation annotation”
arxiv.org · paper · Section 1 · retrieved 2026-10-15
“We adopt the CLIP implementation of OpenCLIP3 and its public pre-trained models in our”
arxiv.org · paper · Section 5, Implementation details · retrieved 2026-10-15
“we use the LAION-400M [37] based CLIP with ViT-B/16+ [16] unless otherwise noted.”
arxiv.org · paper · Section 5 · retrieved 2026-10-15
“The image-text-pairs have been extracted from the Common Crawl web data dump and are from random web pages crawled between 2014 and 2021.”
laion.ai · website · Concept and Content · retrieved 2026-10-15
“arXiv:2503.21622v1 [cs.CV] 27 Mar 2025”
arxiv.org · paper · header · retrieved 2026-10-15
“All image data is available at https://www.mvtec.com/company/research/datasets/mvtec-ad-2.”
arxiv.org · paper · Abstract · retrieved 2026-10-15

Not seen MVTec LOCO AD · absent from the complete list of training data and from its base models' data

Evidence (5)
“We emphasize that our zero-shot models do not require any tuning for individual cases, and the few-normal-only setup does not use any segmentation annotation”
arxiv.org · paper · Section 1 · retrieved 2026-10-15
“We adopt the CLIP implementation of OpenCLIP3 and its public pre-trained models in our”
arxiv.org · paper · Section 5, Implementation details · retrieved 2026-10-15
“we use the LAION-400M [37] based CLIP with ViT-B/16+ [16] unless otherwise noted.”
arxiv.org · paper · Section 5 · retrieved 2026-10-15
“The image-text-pairs have been extracted from the Common Crawl web data dump and are from random web pages crawled between 2014 and 2021.”
laion.ai · website · Concept and Content · retrieved 2026-10-15
“Published: 22 February 2022”
link.springer.com · paper · article metadata · retrieved 2026-10-15

Not seen VisA · absent from the complete list of training data and from its base models' data

Evidence (6)
“We emphasize that our zero-shot models do not require any tuning for individual cases, and the few-normal-only setup does not use any segmentation annotation”
arxiv.org · paper · Section 1 · retrieved 2026-10-15
“We adopt the CLIP implementation of OpenCLIP3 and its public pre-trained models in our”
arxiv.org · paper · Section 5, Implementation details · retrieved 2026-10-15
“we use the LAION-400M [37] based CLIP with ViT-B/16+ [16] unless otherwise noted.”
arxiv.org · paper · Section 5 · retrieved 2026-10-15
“The image-text-pairs have been extracted from the Common Crawl web data dump and are from random web pages crawled between 2014 and 2021.”
laion.ai · website · Concept and Content · retrieved 2026-10-15
“arXiv:2207.14315v1 [cs.CV] 28 Jul 2022”
arxiv.org · paper · header · retrieved 2026-10-15
“We release the Visual Anomaly (VisA) Dataset consisting of 10,821 high-resolution color images (9,621 normal and 1,200 anomalous samples) covering 12 objects in 3 domains”
arxiv.org · paper · Abstract · retrieved 2026-10-15

Not seen Real-IAD · absent from the complete list of training data and from its base models' data

Evidence (5)
“We emphasize that our zero-shot models do not require any tuning for individual cases, and the few-normal-only setup does not use any segmentation annotation”
arxiv.org · paper · Section 1 · retrieved 2026-10-15
“We adopt the CLIP implementation of OpenCLIP3 and its public pre-trained models in our”
arxiv.org · paper · Section 5, Implementation details · retrieved 2026-10-15
“we use the LAION-400M [37] based CLIP with ViT-B/16+ [16] unless otherwise noted.”
arxiv.org · paper · Section 5 · retrieved 2026-10-15
“The image-text-pairs have been extracted from the Common Crawl web data dump and are from random web pages crawled between 2014 and 2021.”
laion.ai · website · Concept and Content · retrieved 2026-10-15
“arXiv:2403.12580v1 [cs.CV] 19 Mar 2024”
arxiv.org · paper · header · retrieved 2026-10-15

Unknown MPDD · unknown for its base model OpenCLIP (LAION-400M checkpoints): training data not disclosed in enough detail

Unknown BTAD · unknown for its base model OpenCLIP (LAION-400M checkpoints): training data not disclosed in enough detail

Unknown KolektorSDD2 · unknown for its base model OpenCLIP (LAION-400M checkpoints): training data not disclosed in enough detail

Notes

Training-free: frozen OpenCLIP ViT-B/16+ (LAION-400M) with a compositional prompt ensemble; no learned parameters, so no extra training data. WinCLIP+ (few-shot) stores patch features of k normal reference images from the target dataset's normal training split in a memory bank at inference (no training). No official code or weights; the InCTRL authors state the official implementation is not publicly available and released a reproduction (github.com/mala-lab/WinCLIP).

Sources