Pith. sign in

REVIEW 4 major objections 6 minor 54 references

AF-CLIP: Zero-Shot Anomaly Detection via Anomaly-Focused CLIP Adaptation

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read AF-CLIP claims that a lightweight attention adapter, multi-scale spatial aggregation, and generic "with/without defect" prompts turn a frozen CLIP into a zero-shot anomaly detector that finds and localizes defects on unseen objects.

desk verdict Solid, well-ablated extension of CLIP-based zero-shot anomaly detection, but the SOTA claim rests on single-run point estimates with 0.4–0.8% margins over strong baselines — needs error bars before it is statistically supported. read the letter →

arxiv 2507.19949 v1 pith:3X6YJG3N submitted 2025-07-26 cs.CV

Qingqing Fang , Wenxi Lv , Qinliang Su This is my paper
classification cs.CV
keywords zero-shotanomalydetectionsegmentationCLIPadaptationpromptlearningattentionadaptermulti-scalespatialaggregationfew-shotvisualinspection
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper sets out to show that zero-shot visual anomaly detection improves when the visual side of CLIP is made anomaly-focused, not just the text prompts. The authors claim that a lightweight attention adapter placed after CLIP's visual layers, combined with multi-scale spatial aggregation before the adapter and learnable 'with/without defect' textual prompts, redirects both the class token and the patch tokens onto local defective regions. With a frozen CLIP backbone and training only on an auxiliary labeled dataset, the method reports image-level and pixel-level AUROC of 92.9% and 92.3% on MVTec and 88.5% and 96.2% on VisA in the zero-shot setting, plus strong transfer to medical datasets. A reader should care because anomaly detection is often needed precisely where no target-domain training examples exist, and the paper proposes a single adapted model without per-category training.

What carries the argument

The load-bearing mechanism is a lightweight self-attention adapter inserted after each of the four hierarchical blocks of the frozen CLIP image encoder. Before the adapter, patch features are reshaped into spatial maps and aggregated under three window sizes using Gaussian weights, so each token carries neighborhood context; the adapter then applies trainable attention over the tokens, letting the class token collect information from distant anomalous patches and letting patch tokens emphasize local defects. The visual features produced this way are aligned with learnable textual prompts whose fixed suffixes are 'without defect' and 'with defect.' A patch alignment loss completes the mechanism by forcing normal-normal and anomaly-anomaly patch pairs to be more similar than normal-anomaly pairs, and the whole adapter-plus-prompt system is trained with classification, segmentation, and alignment losses while CLIP itself stays frozen.

What would settle it

Train AF-CLIP on one industrial benchmark and test on a benchmark whose defect types are absent from the training set, or remove an entire defect type such as structural deformations from the auxiliary data; if AUROC on the held-out defect type collapses to the level of an untrained CLIP baseline while in-distribution defect types keep their gains, the shared-signature transfer assumption is the binding constraint.

Watch

Extended reading notes

Core claim

The paper's central claim is that CLIP's visual encoder can be turned into an anomaly detector without modifying its weights. By training a small attention adapter on features from each of the four hierarchical blocks, aggregating patch tokens with Gaussian-weighted sliding windows of size $r\in\{1,3,5\}$, and learning generic text prompts that end in 'without defect' or 'with defect', the [CLS] token learns to gather evidence from distant anomalous patches while patch tokens learn to focus on local abnormalities. The authors argue that earlier adaptations that only learn prompts or add per-token projections cannot do this, because they do not let tokens exchange information across distant regions. The result is a model that, after optimization on an auxiliary dataset, reports zero-shot image-level/pixel-level AUROC of 92.9%/92.3% on MVTec and 88.5%/96.2% on VisA, and a few-shot extension that reports 97.2%/97.6% on MVTec and 92.2%/98.5% on VisA with four normal shots.

Load-bearing premise

The load-bearing premise is that defects in the target domain share visual signatures with defects in the auxiliary dataset used to train the adapter and prompts, so what the model learns about scratches, dents, and distortions transfers to new objects.

Editorial extensions

If this is right

  • Zero-shot anomaly classification and localization improve together: the paper reports image-level/pixel-level AUROC of 92.9%/92.3% on MVTec and 88.5%/96.2% on VisA with no target-domain training.
  • A model trained on industrial data transfers to medical data, with image-level AUROC of 95.2% on BrainMRI and 96.7% on Br35H and pixel-level AUROC above 83% on the colonoscopy and skin-lesion datasets.
  • In the few-shot setting, adding a memory bank of normal patch features requires no further training and yields 97.2%/97.6% AUROC on MVTec and 92.2%/98.5% on VisA with four normal shots.
  • Freezing CLIP and training only the adapter and prompts uses less training memory than prompt tuning, prefix tuning, or LoRA in the paper's measurements, with comparable inference time.
  • The patch alignment loss contributes at both image and pixel levels: without it, MVTec image-level AUROC drops from 92.9% to 92.4% and pixel-level AUROC from 92.3% to 91.8%.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • An untested corollary of the paper's transfer claim is that swapping the auxiliary and target domains, for example training on medical images and testing on industrial ones, should still show gains if the learned anomaly focus is truly generic; the paper only demonstrates the industrial-to-medical direction.
  • Because the textual prompts are object-agnostic, the same learned prompts should apply to a new object category without retraining; a cheap check is adding unseen MVTec or VisA categories and comparing AUROC with the reported values.
  • The fixed window sizes $r\in\{1,3,5\}$ are a natural place to make the method adaptive, since the spatial scale of defects differs strongly between, say, a scratch on a metal surface and a small polyp in a colonoscopy image.
  • The few-shot memory bank compares stored normal features directly against query patches, so its reliability under changes in viewpoint, lighting, or sensor noise is an open question that could be tested by introducing a lighting shift between the few shots and the test set.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper proposes AF-CLIP, a CLIP-based zero-shot anomaly detection method. The visual encoder is kept frozen, and a lightweight attention adapter is trained on an auxiliary anomaly dataset, preceded by multi-scale sliding-window spatial aggregation of patch features. Learnable textual prompts describe generic normal/abnormal states, and training uses a composite loss of image-level focal loss, segmentation focal+L1 loss, and a patch alignment loss. For few-shot detection, AF-CLIP+ adds memory banks of normal patch features and combines their nearest-neighbor scores with the zero-shot scores. The paper reports zero-shot results on MVTec, VisA, BTAD, DAGM, DTD-Synthetic, and several medical datasets, plus few-shot results on MVTec and VisA, and claims state-of-the-art performance over AdaCLIP, AnomalyCLIP, and AA-CLIP.

Significance. The method is technically coherent and the experimental design is broadly reasonable: it includes component ablations (Table 4), a comparison of fine-tuning strategies (Table 5), a loss-function analysis (Table 6), and evaluation across industrial and medical domains. The cross-dataset training protocol is disclosed: MVTec results are obtained from a VisA-trained model and vice versa, so the zero-shot claim is not circular. Code is promised, which is a reproducibility asset. If the point estimates survive a multi-seed evaluation, AF-CLIP would be a modest but useful improvement over existing CLIP-based zero-shot anomaly detectors. However, the main zero-shot comparisons are currently reported as single point estimates with printed ±0.0 values, while the few-shot rows in the same table show seed-level standard deviations that can exceed the reported zero-shot gains. The statistical support for the headline state-of-the-art claim is therefore not yet established.

major comments (4)
  1. [Section 5.3, Table 2] The zero-shot rows report ±0.0 for all methods, but AF-CLIP's adapter and prompts are trained on an auxiliary dataset with random initialization and stochastic optimization, so these numbers have seed-level uncertainty. The few-shot rows in the same table show standard deviations of 0.9–1.3 AUROC for AF-CLIP+ (e.g., 4-shot MVTec I-AUROC 97.2±0.9, 4-shot VisA I-AUROC 92.2±1.3), which are larger than the reported zero-shot margins over the closest baselines (MVTec I-AUROC +0.8 vs AdaCLIP, MVTec P-AUROC +0.4 vs AA-CLIP, VisA P-AUROC +0.4 vs AdaCLIP in Table 1). Please report zero-shot results over multiple seeds with means and standard deviations (or confidence intervals) before claiming state-of-the-art superiority.
  2. [Section 4.1, Eq. (5)] The multi-scale spatial aggregation uses a Gaussian weight with standard deviation σ, but Section 5.1 does not report the value of σ and no sensitivity analysis is provided. Because spatial aggregation is one of the two main novel components, please specify σ, describe how it was chosen, and ideally report performance as a function of σ for the full model and for the V4 vs V3 ablation in Table 4.
  3. [Section 4.3, Eq. (14)] The patch alignment loss is undefined when there are no abnormal patches (N_a = 0), yet the auxiliary dataset includes normal images. The paper does not specify whether L_pal is computed only on anomalous images, per image or per batch, or how images with no abnormal pixels are handled. Please state the exact computation, since L_pal is a central component of the composite objective in Eq. (15).
  4. [Section 5.2, Table 1] While the text says optimized baselines are trained on the same auxiliary dataset, the table does not state, for each optimized baseline, which auxiliary dataset was used for each target dataset and whether the reported numbers were re-run under this protocol or taken from the original papers. Please make the protocol explicit (e.g., 'VisA-trained model evaluated on MVTec' vs 'MVTec-trained model evaluated on VisA') and clarify whether the same image size, learnable prompt length, and hyperparameter choices were used for all methods.
minor comments (6)
  1. [Abstract] The sentence '97.2%/97.6% on MVTec, 92.2%/98.5% on VisA for 4-shot anomaly detection' is easy to misread, with 98.5 appearing to belong to MVTec; please rephrase as 'on MVTec (97.2% image, 97.6% pixel) and VisA (92.2% image, 98.5% pixel)' for clarity.
  2. [Section 2 heading] The heading 'Releated Work' should be 'Related Work'.
  3. [Throughout] The dataset name is 'VisA', not 'Visa'; please correct this in the abstract, Tables 1–3, and Section 5.
  4. [Section 5.1] The text states '10 datasets' but then lists 11 datasets (5 industrial and 6 medical); please correct the count or the list.
  5. [Section 5.2] The name 'AnomlayCLIP' is a typo and should be 'AnomalyCLIP'.
  6. [PDF header] The running header shows 'Trovato et al.' instead of the author names; this should be corrected in the camera-ready version.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: AF-CLIP's zero-shot claims are benchmark measurements on held-out target datasets, and the disclosed auxiliary-training protocol does not make any prediction equivalent to its fitting inputs by construction.

full rationale

The paper's central claims are empirical AUROC/AP results, not derivations. AF-CLIP trains an adapter, spatial aggregator, and learnable prompts on an auxiliary anomaly dataset (MVTec or VisA) and then evaluates on the other datasets (Section 5.1), a protocol it explicitly discloses and that is standard in zero-shot AD transfer. No test-set labels or target-domain anomalies are used in fitting, so the reported zero-shot numbers are not equivalent to training objectives by construction. The few-shot extension uses a memory bank of normal target samples (Eq. 16-18), again disclosed and not a renamed fit. The only self-citations (refs [9] and [29], by overlapping authors) are used as related-work context or as comparison baselines, not as load-bearing justification of a design choice or as an imported uniqueness theorem. The ablation tables give component-level evidence that is separable from the headline comparisons. Thus there is no circular step that reduces a claimed prediction to its own inputs.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim is empirical and rests mainly on the transferability of anomaly features across domains, the availability of a labeled auxiliary dataset, and the efficacy of the proposed losses. No free parameters are used to manufacture the result, but several hyperparameters (sigma, alpha, beta, lambdas) are either unreported or chosen without sensitivity analysis.

free parameters (5)
  • sigma
    Standard deviation of Gaussian weights in multi-scale spatial aggregation (Eq. 5); controls how much neighborhood context is merged and is never specified in Section 5.1.
  • alpha = 0.1
    Fusion weight for the few-shot image-level score in Eq. 18; chosen by hand and affects the reported few-shot results.
  • beta = 0.1
    Fusion weight for the few-shot pixel-level map in Eq. 18; chosen by hand without sensitivity analysis.
  • lambda1 = 1
    Weight of the segmentation loss in Eq. 15; set to 1 without reported sensitivity analysis.
  • lambda2 = 1
    Weight of the patch alignment loss in Eq. 15; set to 1 without reported sensitivity analysis.
assumptions (4)
  • domain assumption Anomalies across different objects and domains share visual characteristics (e.g., scratches, dents, structural distortions).
    Stated in Section 1 and used in Section 4.2 to justify training on an auxiliary dataset and transferring the learned prompts and adapter to target domains.
  • domain assumption A labeled auxiliary dataset with image-level labels and pixel-level masks is available for training.
    Section 4.3 assumes availability of the triplet {X, Y, M}; this is essential for the supervised classification and segmentation losses.
  • domain assumption Frozen CLIP features are a suitable substrate, and a lightweight adapter can redirect them without full fine-tuning.
    Implicit in the design in Section 4.1; supported by the ablation in Table 5 but not theoretically guaranteed.
  • domain assumption Focal loss plus L1 loss is an appropriate objective for imbalanced anomaly segmentation.
    Section 4.3; the choice is plausible but justified only by empirical results.

how reviews work

0 comments
Cite this review

Pith. "Pith review of AF-CLIP: Zero-Shot Anomaly Detection via Anomaly-Focused CLIP Adaptation." pith.science (2026). https://pith.science/paper/3X6YJG3N

@misc{pith2026250719949,
  author       = {Pith},
  title        = {Pith review of: AF-CLIP: Zero-Shot Anomaly Detection via Anomaly-Focused CLIP Adaptation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3X6YJG3N}},
  note         = {Machine review of arXiv:2507.19949}
}
read the original abstract

Visual anomaly detection has been widely used in industrial inspection and medical diagnosis. Existing methods typically demand substantial training samples, limiting their utility in zero-/few-shot scenarios. While recent efforts have leveraged CLIP's zero-shot recognition capability for this task, they often ignore optimizing visual features to focus on local anomalies, reducing their efficacy. In this work, we propose AF-CLIP (Anomaly-Focused CLIP) by dramatically enhancing its visual representations to focus on local defects. Our approach introduces a lightweight adapter that emphasizes anomaly-relevant patterns in visual features, simultaneously optimizing both class-level features for image classification and patch-level features for precise localization. To capture anomalies of different sizes and improve detection accuracy, prior to the adapter, we develop a multi-scale spatial aggregation mechanism to effectively consolidate neighborhood context. Complementing these visual enhancements, we design learnable textual prompts that generically characterize normal and abnormal states. After optimization on auxiliary datasets using a composite objective function, AF-CLIP demonstrates strong zero-shot detection capability. Our method is also extended to few-shot scenarios by extra memory banks. Experimental results across diverse industrial and medical datasets demonstrate the effectiveness and generalization of our proposed method. Code is available at https://github.com/Faustinaqq/AF-CLIP.

Figures

Figures reproduced from arXiv: 2507.19949 by the authors.

Figure 1
Figure 1. Illustration of attention maps of [CLS] token and [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Structure of AF-CLIP. Visual Features from different blocks are aggregated with different scale spatial information [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Visualization of anomaly localization results. From left to right, the first two [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

54 extracted references · 41 canonical work pages

  1. [1]

    Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. 2019. MVTec AD - A Comprehensive Real-World Dataset for Unsupervised Anom- aly Detection. In IEEE Conference on Computer Vision and Pattern Recognition . Computer Vision Foundation / IEEE, 9592–9600

    work page 2019

  2. [2]

    Jorge Bernal, F Javier Sánchez, Gloria Fernández-Esparrach, Debora Gil, Cristina Rodríguez, and Fernando Vilariño. 2015. WM-DOVA maps for accurate polyp highlighting in colonoscopy: Validation vs. saliency maps from physicians. Com- puterized medical imaging and graphics 43 (2015), 99–111

    work page 2015

  3. [3]

    Yunkang Cao, Jiangning Zhang, Luca Frittoli, Yuqi Cheng, Weiming Shen, and Giacomo Boracchi. 2024. Adaclip: Adapting clip with hybrid learnable prompts for zero-shot anomaly detection. In European Conference on Computer Vision . Springer, 55–72

    work page 2024

  4. [4]

    Xuhai Chen, Yue Han, and Jiangning Zhang. 2023. A zero-/fewshot anomaly classification and segmentation method for cvpr 2023 vand workshop challenge tracks 1&2: 1st place on zero-shot ad and 4th place on few-shot ad.arXiv preprint arXiv:2305.17382 2, 4 (2023)

    arXiv 2023

  5. [5]

    Noel Codella, Veronica Rotemberg, Philipp Tschandl, M Emre Celebi, Stephen Dusza, David Gutman, Brian Helba, Aadi Kalloo, Konstantinos Liopyris, Michael Marchetti, et al. 2019. Skin lesion analysis toward melanoma detection 2018: A challenge hosted by the international skin imaging collaboration (isic). arXiv preprint arXiv:1902.03368 (2019)

    arXiv 2019

  6. [6]

    Chenghao Deng, Haote Xu, Xiaolu Chen, Haodi Xu, Xiaotong Tu, Xinghao Ding, and Yue Huang. 2024. SimCLIP: Refining Image-Text Alignment with Simple Prompts for Zero-/Few-shot Anomaly Detection. In Proceedings of the 32nd ACM International Conference on Multimedia . 1761–1770

    work page 2024

  7. [7]

    Hanqiu Deng and Xingyu Li. 2022. Anomaly detection via reverse distillation from one-class embedding. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 9737–9746

    work page 2022

  8. [8]

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xi- aohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al . 2020. An Image is Worth 16x16 Words: Trans- formers for Image Recognition at Scale. In International Conference on Learning Representations

    2020

Show all 54 references
  1. [9]

    Qingqing Fang, Qinliang Su, Wenxi Lv, Wenchao Xu, and Jianxing Yu. 2025. Boost- ing Fine-Grained Visual Anomaly Detection with Coarse-Knowledge-Aware Ad- versarial Learning. In Proceedings of the AAAI Conference on Artificial Intelligence , Vol. 39. 16532–16540

    2025

  2. [10]

    Zheng Fang, Xiaoyang Wang, Haocheng Li, Jiejie Liu, Qiugui Hu, and Jimin Xiao

  3. [11]

    Zhaopeng Gu, Bingke Zhu, Guibo Zhu, Yingying Chen, Ming Tang, and Jinqiao Wang. 2024. Anomalygpt: Detecting industrial anomalies using large vision- language models. In Proceedings of the AAAI conference on artificial intelligence , Vol. 38. 1932–1940

    2024

  4. [12]

    Denis Gudovskiy, Shun Ishizaka, and Kazuki Kozuka. 2022. Cflow-ad: Real-time unsupervised anomaly detection with localization via conditional normalizing flows. In Proceedings of the IEEE/CVF winter conference on applications of computer vision. 98–107

    2022

  5. [13]

    Hamada. 2020. Br35h: Brain tumor detection 2020. (2020). https://www.kaggle. com/datasets/ahmedhamada0/brain-tumor-detection,2020

    2020

  6. [14]

    Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. 2022. LoRA: Low-Rank Adaptation of Large Language Models. In International Conference on Learning Representations , Vol. 1. 3

    2022

  7. [15]

    Chaoqin Huang, Haoyan Guan, Aofan Jiang, Ya Zhang, Michael Spratling, and Yan-Feng Wang. 2022. Registration based few-shot anomaly detection. In Euro- pean Conference on Computer Vision . Springer, 303–319

    2022

  8. [16]

    Jeeho Hyun, Sangyun Kim, Giyoung Jeon, Seung Hwan Kim, Kyunghoon Bae, and Byung Jun Kang. 2024. ReConPatch: Contrastive patch representation learning for industrial anomaly detection. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision . 2052–2061

    2024

  9. [17]

    Jongheon Jeong, Yang Zou, Taewan Kim, Dongqing Zhang, Avinash Ravichan- dran, and Onkar Dabeer. 2023. Winclip: Zero-/few-shot anomaly classification and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 19606–19616

    2023

  10. [18]

    Debesh Jha, Pia H Smedsrud, Michael A Riegler, Pål Halvorsen, Thomas De Lange, Dag Johansen, and Håvard D Johansen. 2020. Kvasir-seg: A segmented polyp dataset. In MultiMedia modeling: 26th international conference, MMM 2020, Dae- jeon, South Korea, January 5–8, 2020, proceedi...

    2020

  11. [19]

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al

  12. [20]

    Brian Lester, Rami Al-Rfou, and Noah Constant. 2021. The Power of Scale for Parameter-Efficient Prompt Tuning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing . 3045–3059

    2021

  13. [21]

    In Proceedings of the IEEE/CVF International Conference on Computer Vision

    Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision. 4015–4026

  14. [22]

    Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. 2023. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning . PMLR, 19730–19742

    2023

  15. [23]

    Chun-Liang Li, Kihyuk Sohn, Jinsung Yoon, and Tomas Pfister. 2021. Cutpaste: Self-supervised learning for anomaly detection and localization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 9664–9674

    2021

  16. [24]

    Xiaofan Li, Zhizhong Zhang, Xin Tan, Chengwei Chen, Yanyun Qu, Yuan Xie, and Lizhuang Ma. 2024. Promptad: Learning prompts with only normal samples for few-shot anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 16838–16848

    2024

  17. [25]

    Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. 2022. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International conference on machine learning . PMLR, 12888–12900

    2022

  18. [26]

    Yi Li, Hualiang Wang, Yiqun Duan, and Xiaomeng Li. 2023. Clip surgery for better explainability with enhancement in open-vocabulary tasks. arXiv preprint arXiv:2304.05653 (2023)

    2023 arXiv

  19. [27]

    Xiang Lisa Li and Percy Liang. 2021. Prefix-Tuning: Optimizing Continuous Prompts for Generation. In Proceedings of the 59th Annual Meeting of the Associa- tion for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: ...

    2021

  20. [28]

    Zhikang Liu, Yiming Zhou, Yuansheng Xu, and Zilei Wang. 2023. Simplenet: A simple network for image anomaly detection and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 20402– 20411

    2023

  21. [29]

    Xiao Liu, Kaixuan Ji, Yicheng Fu, Weng Tam, Zhengxiao Du, Zhilin Yang, and Jie Tang. 2022. P-Tuning: Prompt Tuning Can Be Comparable to Fine-tuning Across Scales and Tasks. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 2: Sh...

    2022

  22. [30]

    Kevin Zhou

    Wenxin Ma, Xu Zhang, Qingsong Yao, Fenghe Tang, Chenxu Wu, Yingtai Li, Rui Yan, Zihang Jiang, and S. Kevin Zhou. 2025. AA-CLIP: Enhancing Zero-shot Anomaly Detection via Anomaly-Aware CLIP. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    2025

  23. [31]

    Wenxi Lv, Qinliang Su, and Wenchao Xu. 2025. One-for-All Few-Shot Anomaly Detection via Instance-Induced Prompt Learning. In The Thirteenth International Conference on Learning Representations

    2025

  24. [32]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. 2021. Learning transferable visual models from natural language supervision. In International conference on machine learni...

    2021

  25. [33]

    Pankaj Mishra, Riccardo Verk, Daniele Fornasier, Claudio Piciarelli, and Gian Luca Foresti. 2021. VT-ADL: A vision transformer network for image anomaly detec- tion and localization. In 2021 IEEE 30th International Symposium on Industrial Electronics (ISIE). IEEE, 01–06

    2021

  26. [34]

    Karsten Roth, Latha Pemula, Joaquin Zepeda, Bernhard Schölkopf, Thomas Brox, and Peter Gehler. 2022. Towards total recall in industrial anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 14318–14328

    2022

  27. [35]

    Tal Reiss and Yedid Hoshen. 2023. Mean-shifted contrastive loss for anomaly detection. In Proceedings of the AAAI Conference on Artificial Intelligence . 2155– 2162

    2023

  28. [36]

    Mohammadreza Salehi, Niousha Sadjadi, Soroosh Baselizadeh, Mohammad H Rohban, and Hamid R Rabiee. 2021. Multiresolution knowledge distillation for anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 14902–14912

    2021

  29. [37]

    Marco Rudolph, Bastian Wandt, and Bodo Rosenhahn. 2021. Same same but differnet: Semi-supervised defect detection with normalizing flows. InProceedings of the IEEE/CVF winter conference on applications of computer vision . 1907–1916

    2021

  30. [38]

    Justin Tebbe and Jawad Tayyub. 2024. Dynamic Addition of Noise in a Diffusion Model for Anomaly Detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops . 3940–3949

    2024

  31. [39]

    Nima Tajbakhsh, Suryakanth R Gurudu, and Jianming Liang. 2015. Automated polyp detection in colonoscopy videos using shape and context information.IEEE transactions on medical imaging 35, 2 (2015), 630–644

    2015

  32. [40]

    A Vaswani. 2017. Attention is all you need. Advances in Neural Information Processing Systems (2017). MM ’25, October 27–31, 2025, Dublin, Ireland Trovato et al

    2017

  33. [41]

    Tran Dinh Tien, Anh Tuan Nguyen, Nguyen Hoang Tran, Ta Duc Huy, Soan Duong, Chanh D Tr Nguyen, and Steven QH Truong. 2023. Revisiting reverse distillation for anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 24511–24520

    2023

  34. [42]

    Jihun Yi and Sungroh Yoon. 2020. Patch svdd: Patch-level svdd for anomaly detection and segmentation. In Proceedings of the Asian conference on computer vision. 375–390

    2020

  35. [43]

    Matthias Wieler and Tobias Hahn. 2007. Weakly supervised learning for industrial optical inspection. In DAGM symposium in, Vol. 6. 11

    2007

  36. [44]

    Vitjan Zavrtanik, Matej Kristan, and Danijel Skočaj. 2021. Reconstruction by inpainting for visual anomaly detection. Pattern Recognition 112 (2021), 107706

    2021

  37. [45]

    Vitjan Zavrtanik, Matej Kristan, and Danijel Skočaj. 2021. Draem-a discrimi- natively trained reconstruction embedding for surface anomaly detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision . 8330– 8339

    2021

  38. [46]

    Xuan Zhang, Shiyu Li, Xi Li, Ping Huang, Jiulong Shan, and Ting Chen. 2023. DeSTSeg: Segmentation Guided Denoising Student-Teacher for Anomaly Detec- tion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 3914–3923

    2023

  39. [47]

    Jiaqiang Zhang, Senzhang Wang, and Songcan Chen. 2022. Reconstruction Enhanced Multi-View Contrastive Learning for Anomaly Detection on Attributed Networks. In Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence, IJCAI 2022, Vienna, Austri...

    2022

  40. [48]

    Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. 2022. Learning to prompt for vision-language models. International Journal of Computer Vision 130, 9 (2022), 2337–2348

    2022

  41. [49]

    Yuzhong Zhao, Qiaoqiao Ding, and Xiaoqun Zhang. 2023. AE-FLOW: Autoen- coders with Normalizing Flows for Medical Images Anomaly Detection. In The Eleventh International Conference on Learning Representations

    2023

  42. [50]

    Jiawen Zhu and Guansong Pang. 2024. Toward generalist anomaly detection via in-context residual learning with few-shot sample prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 17826–17836

    2024

  43. [51]

    Qihang Zhou, Guansong Pang, Yu Tian, Shibo He, and Jiming Chen. 2024. Anoma- lyCLIP: Object-agnostic Prompt Learning for Zero-shot Anomaly Detection. In The Twelfth International Conference on Learning Representations

    2024

  44. [53]

    Yang Zou, Jongheon Jeong, Latha Pemula, Dongqing Zhang, and Onkar Dabeer

  45. [2022]

    In European Conference on Computer Vision

    Spot-the-difference self-supervised pre-training for anomaly detection and segmentation. In European Conference on Computer Vision . Springer, 392–408

  46. [2023]

    In Proceedings of the IEEE/CVF International Conference on Computer Vision

    Fastrecon: Few-shot industrial anomaly detection via fast feature recon- struction. In Proceedings of the IEEE/CVF International Conference on Computer Vision. 17481–17490

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.