Pith. sign in

REVIEW 4 major objections 5 minor 57 references

UniADC: A Unified Framework for Anomaly Detection and Classification

T0 review · 4 major / 5 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read UniADC proposes that anomaly detection and classification should be a single task, and that both can be solved with only a few normal images by synthesizing defects with controllable inpainting and defining normal as the absence of every kn

desk verdict A credible one-model answer to zero/few-shot defect detection-plus-classification, worth refereeing, but the synthetic-to-real gap is under-analyzed and the 'consistently outperforms' claim over-reaches. read the letter →

arxiv 2511.06644 v3 pith:PJ4AICXA submitted 2025-11-10 cs.CV

Ximiao Zhang , Min Xu , Zheng Zhang , Yap-Peng Tan , Xiuzhuang Zhou This is my paper
classification cs.CV
keywords anomalydetectionclassificationsynthesisfew-shotlearningzero-shotdiffusioninpaintingimplicit-normaldiscriminationindustrialinspection
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

The paper introduces a unified task: detect anomalous image regions and assign them to specific defect categories in one model. It claims that with only a few normal images and either text/shape descriptions of defects (zero-shot) or one real defect sample per class (few-shot), a single model can outperform two-stage pipelines that separate detection from classification. The key move is to generate category-specific synthetic anomalies by repainting normal regions with a diffusion inpainting network, then train a discriminator that never treats 'normal' as a class, scoring pixels only by similarity to anomaly-category embeddings. If correct, factory inspection systems could be deployed for new products with almost no defect data.

What carries the argument

The load-bearing mechanism is a two-part pipeline. First, a geometric anomaly prototype library (GAP-Lib) generates candidate masks from eight shapes at three sizes; a latent diffusion model plus an inpainting control network repaints only the masked region under a text prompt, and a mask-conditioned vision-language model filters synthetic samples for category consistency. Second, the implicit-normal discriminator removes any explicit 'normal' class embedding, instead computing per-pixel similarity between fused visual features and each anomaly-category embedding, averaging those maps for detection, and taking the argmax above a threshold for classification. The exclusion-based normal defini

What would settle it

Measure the feature-space distance (e.g., maximum mean discrepancy) between synthetic anomaly crops and real anomaly crops using the same visual backbone, then train UniADC with real defect crops in place of synthetic ones and compare classification accuracy and mIoU on the same test sets; a large distance or a big performance gap would show the reported gains rest on unverified synthetic-to-real fidelity.

Watch

Extended reading notes

Core claim

The paper's central claim is that a generative-discriminative pair can jointly solve anomaly detection, localization, and classification under extreme data scarcity. A training-free controllable inpainting network synthesizes category-specific anomalies in two modes: zero-shot, guided by hand-specified shape, size, and text priors, and few-shot, by repainting pasted real defect crops for diversity. A category consistency selection step filters synthetic samples that do not match the target category. The implicit-normal discriminator then aligns fused fine-grained image features with anomaly-category text embeddings pixel-wise; the anomaly detection map is simply the average of all category m

Load-bearing premise

The load-bearing premise is that diffusion-inpainted synthetic defects live close enough to real defects in the vision-language feature space that training a discriminator on them teaches it real-world anomalies.

Editorial extensions

If this is right

  • One model can flag defective regions and assign them to specific defect classes with no real defect samples at all, using only normal images and text/shape descriptions.
  • With a single real anomaly example per category, the same model reaches classification accuracy and pixel-level mIoU far above current two-stage baselines, according to the reported tables.
  • Adding the classification loss improves anomaly detection and localization rather than hurting them, so the two tasks reinforce each other.
  • When all normal training samples are available, the model matches dedicated full-shot detection methods while still giving a large classification advantage.
  • The framework can be adapted to open-set operation by training a class-agnostic anomaly embedding and routing unseen defect types to an 'Other' category while preserving seen-category classification.

Reading between the lines

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

  • If the synthetic-to-real gap is as small as the paper suggests, defect-data collection across manufacturing lines could be replaced by writing a one-line text description of each defect and a rough shape, cutting deployment cost dramatically; a direct test would be deploying UniADC on a new production line with no defect history.
  • The implicit-normal trick—defining normal as the absence of all known abnormality—is a general recipe for any dense prediction task with a dominant background class, so it could transfer to medical image segmentation or rare-event pixel labeling.
  • Because the zero-shot prior list is hand-written per category, the method's practical ceiling depends on how verbalizable defects are; an extension worth testing is auto-generating descriptions and shapes from a defect taxonomy or a language model.
  • The classification map averages all category scores, so overlapping or ambiguous defect categories could inflate detection; a testable refinement would be learning per-category thresholds or adding a calibrated uncertainty term.
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 / 5 minor

Summary. The paper introduces UniADC, a unified framework for anomaly detection and classification under zero-shot and few-shot settings. It combines a training-free controllable inpainting network (latent diffusion with BrushNet) that synthesizes category-specific anomalies from either text/shape/size priors or from few-shot anomaly examples, with an implicit-normal discriminator (IND) that aligns fine-grained visual features with anomaly-category text embeddings while treating 'normal' implicitly as the absence of all anomaly classes. The model is trained on synthetic anomalies filtered by category consistency selection (Eq. 2; SSIM-based in the few-shot case) and evaluated on MVTec-FS, MTD, and WFDD. Under both settings, UniADC reports substantial gains over existing two-stage baselines in image-level detection (I-AUC), localization (P-AUC/PRO), classification accuracy, and pixel-level mIoU. An open-set extension is also proposed.

Significance. If the reported results hold, UniADC makes a meaningful contribution: it is one of the first single models to jointly solve anomaly detection, localization, and fine-grained classification with zero or a few anomaly samples, and it outperforms prior work by a large margin on the classification metrics. The idea of an implicit normal class (IND) is a well-motivated answer to the pixel-distribution imbalance problem, and the use of a training-free, text-controllable inpainting network to generate training data is practical. The paper includes extensive ablations over components, priors, noise factor, mini-batch size, and sample count, and the code is available, which strengthens reproducibility. The open-set extension is a useful addition. However, the main claim of 'comprehensively outperforms' depends on the synthetic anomalies being representative of real defects, and this is not quantitatively established.

major comments (4)
  1. [§III-B/C, Eq. (2), Table V] Synthetic-to-real generalization is the load-bearing assumption. The discriminator is trained on 16 synthetic samples per category generated by latent diffusion and filtered by AlphaCLIP text alignment (Eq. 2). The CCS score measures alignment with the text prompt, not similarity to real defect appearance. The image-quality metrics in Table V (IS, IC-LPIPS) do not quantify the distance between synthetic and real anomaly distributions, and Fig. 8 is qualitative. Since all classification and most detection performance comes from training on these synthetic images, the reported Acc/mIoU gains could partly reflect the model recognizing prompt-rendered artifacts rather than real defects, especially for categories with highly specific hand-authored prompts (Table S1). Please provide a quantitative synthetic-to-real transfer analysis (e.g., train on synthetic and evaluate on real anomaly crops
  2. [Related Work, Sec. II-B; Table II] The closest zero-shot baseline, MultiADS [26], is absent from the experimental comparison. MultiADS is designed for zero-shot multi-type anomaly detection and classification, and the paper criticizes it in the Related Work; yet it is not included in Table II. To support 'comprehensively outperforms existing methods,' either add MultiADS to the comparison or justify its exclusion (e.g., if it requires training data not compatible with the protocol). This is particularly important for the zero-shot claims.
  3. [§IV-A.1, WFDD] The WFDD anomaly-type labels were created by the authors ('we reclassified the dataset by grouping identical defects'). No annotation protocol, label release, or inter-annotator agreement is provided. Consequently, the WFDD classification results are not comparable to any external benchmark, and the hand-crafted prior list in Table S1 may have been tuned to these self-defined categories. Please publish the labels and provide details on the grouping, or at least a comprehensive diagram showing all category assignments.
  4. [§IV, Tables II–V] No error bars or multiple seeds are reported in any table. Anomaly synthesis is stochastic (diffusion sampling, mask generation), and several hyperparameters (γ, B, τ, λ) are set by ablations. The paper reports single-run results. Given that some comparisons are close (e.g., few-shot Kn=2, Ka=2 MVTec Acc for UniADC(DINO) is 88.72 vs UniADC(CLIP) 89.88 in Table III), variance from random seeds would materially affect the conclusions. Please report mean±std over at least 3 seeds for the main tables and key ablations.
minor comments (5)
  1. [Abstract] The abstract states 'four anomaly detection and classification datasets, including MVTec-FS, MTD, WFDD and Real-IAD,' but the full text and all experiments use only three datasets (MVTec-FS, MTD, WFDD). Real-IAD is never evaluated. Correct this inconsistency.
  2. [§III-C, Table I] The claim of being the 'first method capable of synthesizing anomalies under both zero-shot and few-shot settings' should be qualified with a more exhaustive literature comparison. AnomalyPainter [29] and AnoMagic [36] are cited in Related Work and should be considered or discussed.
  3. [Eq. (2)] The Category Consistency Selection score in Eq. (2) normalizes over all anomaly categories, meaning the selection for a single category depends on the full closed-set label descriptions. This coupling should be stated explicitly, as it may affect datasets with a large number of categories.
  4. [§III-C] The SSIM-based consistency score for few-shot repainting is mentioned but not formalized (e.g., whether there is a threshold or whether the highest-scoring sample is simply kept). Provide a precise selection rule and the number of repainting candidates.
  5. [Eq. (5)] The anomaly score threshold τ is fixed at 0.5. Although Fig. 11 shows robustness, no guidance is given for selecting τ in a new domain. Consider describing a validation-free default or a simple calibration procedure.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: predictions are evaluated on external real labels.

full rationale

I find no circular step in this paper. The claimed predictions (detection, localization, classification) are measured against real ground-truth labels on MVTec-FS, MTD, and WFDD that were not used to fit any parameter or to select synthetic data. The synthetic training data are generated by a pre-trained diffusion/inpainting pipeline conditioned on hand-authored text/shape priors (Sec. III-B/C, Table S1), and the Category Consistency Selection in Eq. (2) is a data-cleaning filter, not a fitted predictor. The final metrics are computed on independent real test images, so the model could fail; the reported numbers are not forced by construction. The fact that both the AlphaCLIP filter and the inference classifier use text-anchored embeddings reflects the label space of the task, not a reduction: the synthetic-to-real generalization gap flagged by the skeptic is a genuine empirical risk, but it is a correctness/robustness concern, not a circularity, because the test labels are external and the model is not guaranteed to succeed. Self-citations (e.g., RealNet, MedICLIP) appear only as baselines or related work and are not load-bearing; no uniqueness theorem or prior claim by the same authors is imported to force the design. The core derivation is self-contained and externally benchmarked, so the appropriate circularity score is 0.

Assumptions & free parameters 6 free parameters · 6 assumptions · 0 invented entities

UniADC rests on pragmatic bets: pretrained diffusion/BrushNet can synthesize category-consistent anomalies; AlphaCLIP scoring is a valid filter; synthetic anomalies transfer to real images; and category text embeddings are available in the visual backbone's space. The first three are plausible but not quantitatively validated; the fourth is unspecified for the DINO variant. Hyperparameters gamma, B, sample count, tau, and lambda are chosen from ablations rather than derived.

free parameters (6)
  • Diffusion noise factor gamma = uniform(0.4, 0.6)
    Controls how far repainted anomalies deviate from the original image; range chosen from the Table IX ablation rather than derived.
  • Category-consistency mini-batch size B = 32
    Number of generated candidates per category in the CCS filter (Eq. 2); set from Table X ablation; larger B improves performance but costs compute.
  • Number of synthetic anomaly samples per category = 16
    Generated at 512x512 for discriminator training; set from Table XI ablation.
  • Anomaly score threshold tau = 0.5
    Threshold in Eq. 5 separating normal and anomalous pixels; fixed after the Fig. 11 robustness ablation.
  • Classification loss weight lambda = 0.5
    Weight on the cross-entropy term in Eq. 6; set manually with no dedicated ablation.
  • Scaling temperature epsilon = learned, not reported
    Temperature in Eq. 4 for the sigmoid similarity; learned during training, final value not reported.
assumptions (6)
  • domain assumption Stable Diffusion v1.5 plus BrushNet inpainting yields category-consistent anomaly images from masks and text prompts without training.
    Central to the synthesis modules in Sections III-B/C; if repainting produces off-category or incoherent images, the category consistency filter cannot fully compensate.
  • domain assumption Synthetic anomaly distribution is a sufficient proxy for real anomalies for training the discriminator.
    All discriminator training uses synthesized samples or repainted few-shot patches; generalization to real anomalies is assumed and only indirectly tested on benchmarks.
  • domain assumption AlphaCLIP matching score (Eq. 2) ranks category consistency correctly.
    The category consistency selection keeps only the top-scoring candidate; if AlphaCLIP is miscalibrated, selected samples may be mislabeled.
  • ad hoc to paper For UniADC(DINO), a text encoder exists that maps anomaly category names into the same space as DINOv3 visual features.
    DINOv3 is described as a vision-language backbone in Section IV-A.2, but no text encoder is specified; Eq. 4 requires category embeddings.
  • domain assumption The authors' regrouping of WFDD defects into anomaly categories is valid ground truth.
    Section IV-A.1 states WFDD was reclassified by the authors; no external validation or released label set is provided.
  • domain assumption The per-category anomaly prior list (Table S1) is representative and does not encode test-label information beyond category names.
    Descriptions such as 'Rotating metal nut with motion blur' are hand-written per category; the paper does not bound how much human effort is encoded in the priors.

how reviews work

0 comments
Cite this review

Pith. "Pith review of UniADC: A Unified Framework for Anomaly Detection and Classification." pith.science (2026). https://pith.science/paper/PJ4AICXA

@misc{pith2026251106644,
  author       = {Pith},
  title        = {Pith review of: UniADC: A Unified Framework for Anomaly Detection and Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PJ4AICXA}},
  note         = {Machine review of arXiv:2511.06644}
}
read the original abstract

In this paper, we introduce a novel task termed unified anomaly detection and classification, which aims to simultaneously detect anomalous regions in images and identify their specific categories. Existing methods typically treat anomaly detection and classification as separate tasks, thereby neglecting their inherent correlations and limiting information sharing, which results in suboptimal performance. To address this, we propose UniADC, a model designed to effectively perform both tasks with only a few or even no anomaly images. Specifically, UniADC consists of two key components: a training-free Controllable Inpainting Network and an Implicit-Normal Discriminator. The inpainting network can synthesize anomaly images of specific categories by repainting normal regions guided by anomaly priors, and can also repaint few-shot anomaly samples to augment the available anomaly data. The implicit-normal discriminator addresses the severe challenge of the imbalance between normal and anomalous pixel distributions by implicitly modeling the normal state, achieving precise anomaly detection and classification by aligning fine-grained image features with anomaly-category embeddings. We conduct extensive experiments on four anomaly detection and classification datasets, including MVTec-FS, MTD, WFDD and Real-IAD, and the results demonstrate that UniADC consistently outperforms existing methods in anomaly detection, localization, and classification. The code is available at https://github.com/cnulab/UniADC.

Figures

Figures reproduced from arXiv: 2511.06644 by the authors.

Figure 1
Figure 1. Task settings for unified anomaly detection and classification. (a) [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Comparison of UniADC with alternative methods on the MVTec-FS [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Overview of the proposed UniADC pipeline, which consists of a training-free controllable inpainting network and an implicit-normal discriminator. [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Examples of anomaly masks generated from the GAP-Lib. [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: Architecture of the Feature Fusion Network. [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: Examples of different anomaly categories in the WFDD dataset, with [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 7
Figure 7. Figure 7: Qualitative results of UniADC on the MVTec-FS, MTD, and WFDD datasets under zero-shot [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: Examples of synthetic anomaly samples generated by UniADC under the guidance of anomaly prior and anomaly sample on the MVTec-FS, MTD, [PITH_FULL_IMAGE:figures/full_fig_p009_8.png]
Figure 9
Figure 9. Figure 9: Qualitative results of UniADC under the setting of open-set anomaly [PITH_FULL_IMAGE:figures/full_fig_p010_9.png]
Figure 10
Figure 10. Figure 10: Examples of synthetic anomaly images generated by UniADC across different noise factors, guided by anomaly prior or anomaly sample, respectively. [PITH_FULL_IMAGE:figures/full_fig_p011_10.png]
Figure 11
Figure 11. Figure 11: Ablation study of the anomaly score threshold on the MVTec-FS [PITH_FULL_IMAGE:figures/full_fig_p012_11.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

57 extracted references · 3 linked inside Pith

  1. [54]

    Sim ´eoni, H

    O. Sim ´eoni, H. V . V o, M. Seitzer, F. Baldassarre, M. Oquab, C. Jose, V . Khalidov, M. Szafraniec, S. Yi, M. Ramamonjisoaet al., “Dinov3,” arXiv preprint arXiv:2508.10104, 2025

    arXiv 2025

  2. [26]

    Mul- tiads: Defect-aware supervision for multi-type anomaly detection and segmentation in zero-shot learning,

    Y . Sadikaj, H. Zhou, L. Halilaj, S. Schmid, S. Staab, and C. Plant, “Mul- tiads: Defect-aware supervision for multi-type anomaly detection and segmentation in zero-shot learning,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2025, pp. 22 978–22 988

    2025

  3. [1]

    Deep learning for unsupervised anomaly localization in industrial images: A survey,

    X. Tao, X. Gong, X. Zhang, S. Yan, and C. Adak, “Deep learning for unsupervised anomaly localization in industrial images: A survey,”IEEE Transactions on Instrumentation and Measurement, vol. 71, pp. 1–21, 2022

    2022

  4. [2]

    Gan-based anomaly detection: A review,

    X. Xia, X. Pan, N. Li, X. He, L. Ma, X. Zhang, and N. Ding, “Gan-based anomaly detection: A review,”Neurocomputing, vol. 493, pp. 497–535, 2022

    2022

  5. [3]

    A survey of methods for automated quality control based on images,

    J. Diers and C. Pigorsch, “A survey of methods for automated quality control based on images,”International Journal of Computer Vision, vol. 131, no. 10, pp. 2553–2581, 2023

    2023

  6. [4]

    Deep industrial image anomaly detection: A survey,

    J. Liu, G. Xie, J. Wang, S. Li, C. Wang, F. Zheng, and Y . Jin, “Deep industrial image anomaly detection: A survey,”Machine Intelligence Research, vol. 21, no. 1, pp. 104–135, 2024

    2024

  7. [5]

    Mvrec: A general few-shot defect classification model using multi-view region- context,

    S. Lyu, R. Zhang, Z. Ma, F. Liao, D. Mo, and W. Wong, “Mvrec: A general few-shot defect classification model using multi-view region- context,” inProceedings of the AAAI Conference on Artificial Intelli- gence, vol. 39, no. 6, 2025, pp. 5937–5945

    2025

  8. [6]

    Anomalyncd: Towards novel anomaly class discovery in industrial scenarios,

    Z. Huang, X. Li, H. Liu, F. Xue, Y . Wang, and Y . Zhou, “Anomalyncd: Towards novel anomaly class discovery in industrial scenarios,” inPro- ceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 4755–4765

    2025

Show all 57 references
  1. [7]

    Winclip: Zero-/few-shot anomaly classification and segmentation,

    J. Jeong, Y . Zou, T. Kim, D. Zhang, A. Ravichandran, and O. Dabeer, “Winclip: Zero-/few-shot anomaly classification and segmentation,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 19 606–19 616

    2023

  2. [8]

    Promptad: Learning prompts with only normal samples for few-shot anomaly detection,

    X. Li, Z. Zhang, X. Tan, C. Chen, Y . Qu, Y . Xie, and L. Ma, “Promptad: Learning prompts with only normal samples for few-shot anomaly detection,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 16 838–16 848

    2024

  3. [9]

    Anomalygpt: Detecting industrial anomalies using large vision-language models,

    Z. Gu, B. Zhu, G. Zhu, Y . Chen, M. Tang, and J. Wang, “Anomalygpt: Detecting industrial anomalies using large vision-language models,” in Proceedings of the AAAI conference on artificial intelligence, vol. 38, no. 3, 2024, pp. 1932–1940

    2024

  4. [10]

    Kernel-aware graph prompt learning for few-shot anomaly detection,

    F. Tao, G.-S. Xie, F. Zhao, and X. Shu, “Kernel-aware graph prompt learning for few-shot anomaly detection,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 39, no. 7, 2025, pp. 7347– 7355

    2025

  5. [11]

    Medi- clip: Adapting clip for few-shot medical image anomaly detection,

    X. Zhang, M. Xu, D. Qiu, R. Yan, N. Lang, and X. Zhou, “Medi- clip: Adapting clip for few-shot medical image anomaly detection,” in International Conference on Medical Image Computing and Computer- Assisted Intervention. Springer, 2024, pp. 458–468

    2024

  6. [12]

    Adapting visual-language models for generalizable anomaly detection in medical images,

    C. Huang, A. Jiang, J. Feng, Y . Zhang, X. Wang, and Y . Wang, “Adapting visual-language models for generalizable anomaly detection in medical images,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 11 375–11 385

    2024

  7. [13]

    High- resolution image synthesis with latent diffusion models,

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer, “High- resolution image synthesis with latent diffusion models,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 10 684–10 695

    2022

  8. [14]

    Adding conditional control to text-to-image diffusion models,

    L. Zhang, A. Rao, and M. Agrawala, “Adding conditional control to text-to-image diffusion models,” inProceedings of the IEEE/CVF international conference on computer vision, 2023, pp. 3836–3847

    2023

  9. [15]

    Surface defect saliency of magnetic tile,

    Y . Huang, C. Qiu, and K. Yuan, “Surface defect saliency of magnetic tile,”The Visual Computer, vol. 36, no. 1, pp. 85–96, 2020

    2020

  10. [16]

    A unified anomaly synthesis strategy with gradient ascent for industrial anomaly detection and localization,

    Q. Chen, H. Luo, C. Lv, and Z. Zhang, “A unified anomaly synthesis strategy with gradient ascent for industrial anomaly detection and localization,” inEuropean Conference on Computer Vision. Springer, 2024, pp. 37–54

    2024

  11. [17]

    Towards total recall in industrial anomaly detection,

    K. Roth, L. Pemula, J. Zepeda, B. Sch ¨olkopf, T. Brox, and P. Gehler, “Towards total recall in industrial anomaly detection,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 14 318–14 328

    2022

  12. [18]

    Padim: a patch dis- tribution modeling framework for anomaly detection and localization,

    T. Defard, A. Setkov, A. Loesch, and R. Audigier, “Padim: a patch dis- tribution modeling framework for anomaly detection and localization,” inInternational Conference on Pattern Recognition. Springer, 2021, pp. 475–489

    2021

  13. [19]

    Anomaly detection and localization via reverse distillation with latent anomaly suppression,

    G. Wang, Y . Zou, S. He, Y . Wang, and R. Dai, “Anomaly detection and localization via reverse distillation with latent anomaly suppression,” IEEE Transactions on Circuits and Systems for Video Technology, 2025

    2025

  14. [20]

    Ura-net: Uncertainty-integrated anomaly perception and restoration attention network for unsupervised anomaly detection,

    W. Luo, P. Xing, Y . Cao, H. Yao, W. Shen, and Z. Li, “Ura-net: Uncertainty-integrated anomaly perception and restoration attention network for unsupervised anomaly detection,”IEEE Transactions on Circuits and Systems for Video Technology, 2025

    2025

  15. [21]

    Towards high-resolution industrial image anomaly detection,

    X. Zhang, M. Xu, and X. Zhou, “Towards high-resolution industrial image anomaly detection,”arXiv preprint arXiv:2508.12931, 2025

    2025 arXiv

  16. [22]

    Few-shot anomaly- driven generation for anomaly classification and segmentation,

    G. Gui, B.-B. Gao, J. Liu, C. Wang, and Y . Wu, “Few-shot anomaly- driven generation for anomaly classification and segmentation,” in European Conference on Computer Vision. Springer, 2024, pp. 210– 226

    2024

  17. [23]

    Explicit boundary guided semi-push-pull contrastive learning for supervised anomaly detection,

    X. Yao, R. Li, J. Zhang, J. Sun, and C. Zhang, “Explicit boundary guided semi-push-pull contrastive learning for supervised anomaly detection,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 24 490–24 499

    2023

  18. [24]

    Normal-abnormal guided generalist anomaly detection,

    Y . Wang, X. Wang, Y . Gong, and J. XIAO, “Normal-abnormal guided generalist anomaly detection,” inThe Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025

    2025

  19. [25]

    Alpha-clip: A clip model focusing on wherever you want,

    Z. Sun, Y . Fang, T. Wu, P. Zhang, Y . Zang, S. Kong, Y . Xiong, D. Lin, and J. Wang, “Alpha-clip: A clip model focusing on wherever you want,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024, pp. 13 019–13 029. JOURNAL OF LATEX CLASS ...

    2024

  20. [27]

    A noise robust method based on completed local binary patterns for hot-rolled steel strip surface defects,

    K. Song and Y . Yan, “A noise robust method based on completed local binary patterns for hot-rolled steel strip surface defects,”Applied Surface Science, vol. 285, pp. 858–864, 2013

    2013

  21. [28]

    Fabric defect classification using proto- typical network of few-shot learning algorithm,

    Z. Zhan, J. Zhou, and B. Xu, “Fabric defect classification using proto- typical network of few-shot learning algorithm,”Computers in Industry, vol. 138, p. 103628, 2022

    2022

  22. [29]

    Anoma- lypainter: Vision-language-diffusion synergy for zero-shot realistic and diverse industrial anomaly synthesis,

    Z. Lai, Y . Lu, X. Li, J. Lin, Y . Qu, L. Cao, M. Li, and R. Ji, “Anoma- lypainter: Vision-language-diffusion synergy for zero-shot realistic and diverse industrial anomaly synthesis,”arXiv preprint arXiv:2503.07253, 2025

    2025 arXiv

  23. [30]

    Cutpaste: Self-supervised learning for anomaly detection and localization,

    C.-L. Li, K. Sohn, J. Yoon, and T. Pfister, “Cutpaste: Self-supervised learning for anomaly detection and localization,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 9664–9674

    2021

  24. [31]

    Natural synthetic anomalies for self-supervised anomaly detection and localization,

    H. M. Schl ¨uter, J. Tan, B. Hou, and B. Kainz, “Natural synthetic anomalies for self-supervised anomaly detection and localization,” in European Conference on Computer Vision. Springer, 2022, pp. 474– 489

    2022

  25. [32]

    Normal image guided segmentation framework for unsupervised anomaly detection,

    P. Xing, Y . Sun, D. Zeng, and Z. Li, “Normal image guided segmentation framework for unsupervised anomaly detection,”IEEE Transactions on Circuits and Systems for Video Technology, vol. 34, no. 6, pp. 4639– 4652, 2023

    2023

  26. [33]

    Revisiting reverse distillation for anomaly detection,

    T. D. Tien, A. T. Nguyen, N. H. Tran, T. D. Huy, S. Duong, C. D. T. Nguyen, and S. Q. Truong, “Revisiting reverse distillation for anomaly detection,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 24 511–24 520

    2023

  27. [34]

    Realnet: A feature selection network with realistic synthetic anomaly for anomaly detection,

    X. Zhang, M. Xu, and X. Zhou, “Realnet: A feature selection network with realistic synthetic anomaly for anomaly detection,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024, pp. 16 699–16 708

    2024

  28. [35]

    Unseen visual anomaly gener- ation,

    H. Sun, Y . Cao, H. Dong, and O. Fink, “Unseen visual anomaly gener- ation,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 25 508–25 517

    2025

  29. [36]

    Anomagic: Crossmodal prompt-driven zero-shot anomaly generation,

    Y . Jiang, W. Luo, H. Zhang, Q. Chen, H. Yao, W. Shen, and Y . Cao, “Anomagic: Crossmodal prompt-driven zero-shot anomaly generation,” arXiv preprint arXiv:2511.10020, 2025

    2025

  30. [37]

    Few-shot defect image generation via defect-aware feature manipulation,

    Y . Duan, Y . Hong, L. Niu, and L. Zhang, “Few-shot defect image generation via defect-aware feature manipulation,” inProceedings of the AAAI conference on artificial intelligence, vol. 37, no. 1, 2023, pp. 571–578

    2023

  31. [38]

    Anomalydiffusion: Few-shot anomaly image generation with diffusion model,

    T. Hu, J. Zhang, R. Yi, Y . Du, X. Chen, L. Liu, Y . Wang, and C. Wang, “Anomalydiffusion: Few-shot anomaly image generation with diffusion model,” inProceedings of the AAAI conference on artificial intelligence, vol. 38, no. 8, 2024, pp. 8526–8534

    2024

  32. [39]

    Dual-interrelated diffusion model for few-shot anomaly image generation,

    Y . Jin, J. Peng, Q. He, T. Hu, J. Wu, H. Chen, H. Wang, W. Zhu, M. Chi, J. Liuet al., “Dual-interrelated diffusion model for few-shot anomaly image generation,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 30 420–30 429

    2025

  33. [40]

    Defectfill: Realistic defect generation with inpainting diffusion model for visual inspection,

    J. Song, D. Park, K. Baek, S. Lee, J. Choi, E. Kim, and S. Yoon, “Defectfill: Realistic defect generation with inpainting diffusion model for visual inspection,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 18 718–18 727

    2025

  34. [41]

    Brushnet: A plug-and-play image inpainting model with decomposed dual-branch diffusion,

    X. Ju, X. Liu, X. Wang, Y . Bian, Y . Shan, and Q. Xu, “Brushnet: A plug-and-play image inpainting model with decomposed dual-branch diffusion,” inEuropean Conference on Computer Vision. Springer, 2024, pp. 150–168

    2024

  35. [42]

    Image quality assessment: from error visibility to structural similarity,

    Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli, “Image quality assessment: from error visibility to structural similarity,”IEEE transactions on image processing, vol. 13, no. 4, pp. 600–612, 2004

    2004

  36. [43]

    Draem-a discriminatively trained reconstruction embedding for surface anomaly detection,

    V . Zavrtanik, M. Kristan, and D. Sko ˇcaj, “Draem-a discriminatively trained reconstruction embedding for surface anomaly detection,” in Proceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 8330–8339

    2021

  37. [44]

    AnomalyCLIP: Object- agnostic prompt learning for zero-shot anomaly detection,

    Q. Zhou, G. Pang, Y . Tian, S. He, and J. Chen, “AnomalyCLIP: Object- agnostic prompt learning for zero-shot anomaly detection,” inThe Twelfth International Conference on Learning Representations, 2024

    2024

  38. [45]

    Focal loss for dense object detection,

    T.-Y . Lin, P. Goyal, R. Girshick, K. He, and P. Doll ´ar, “Focal loss for dense object detection,” inProceedings of the IEEE international conference on computer vision, 2017, pp. 2980–2988

    2017

  39. [46]

    V-net: Fully convolutional neural networks for volumetric medical image segmentation,

    F. Milletari, N. Navab, and S.-A. Ahmadi, “V-net: Fully convolutional neural networks for volumetric medical image segmentation,” in2016 fourth international conference on 3D vision (3DV). Ieee, 2016, pp. 565–571

    2016

  40. [47]

    Toward generalist anomaly detection via in-context residual learning with few-shot sample prompts,

    J. Zhu and G. Pang, “Toward generalist anomaly detection via in-context residual learning with few-shot sample prompts,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024, pp. 17 826–17 836

    2024

  41. [48]

    Anomalydino: Boosting patch-based few-shot anomaly detection with dinov2,

    S. Damm, M. Laszkiewicz, J. Lederer, and A. Fischer, “Anomalydino: Boosting patch-based few-shot anomaly detection with dinov2,” in 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). IEEE, 2025, pp. 1319–1329

    2025

  42. [49]

    Mvtec-ad: A comprehensive real-world dataset for unsupervised anomaly detection,

    P. Bergmann, M. Fauser, D. Sattlegger, and C. Steger, “Mvtec-ad: A comprehensive real-world dataset for unsupervised anomaly detection,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 9592–9600

    2019

  43. [50]

    Spot-the- difference self-supervised pre-training for anomaly detection and seg- mentation,

    Y . Zou, J. Jeong, L. Pemula, D. Zhang, and O. Dabeer, “Spot-the- difference self-supervised pre-training for anomaly detection and seg- mentation,” inEuropean Conference on Computer Vision. Springer, 2022, pp. 392–408

    2022

  44. [51]

    Denoising diffusion implicit models,

    J. Song, C. Meng, and S. Ermon, “Denoising diffusion implicit models,” inInternational Conference on Learning Representations, 2021

    2021

  45. [52]

    Bilateral reference for high-resolution dichotomous image segmentation,

    P. Zheng, D. Gao, D.-P. Fan, L. Liu, J. Laaksonen, W. Ouyang, and N. Sebe, “Bilateral reference for high-resolution dichotomous image segmentation,”CAAI Artificial Intelligence Research, vol. 3, p. 9150038, 2024

    2024

  46. [53]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clarket al., “Learning transferable visual models from natural language supervision,” inInternational conference on machine learning. PmLR, 2021, pp. 8748–8763

    2021

  47. [55]

    Uninformed students: Student-teacher anomaly detection with discriminative latent embeddings,

    P. Bergmann, M. Fauser, D. Sattlegger, and C. Steger, “Uninformed students: Student-teacher anomaly detection with discriminative latent embeddings,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 4183–4192

    2020

  48. [56]

    Anomaly detection via reverse distillation from one-class embedding,

    H. Deng and X. Li, “Anomaly detection via reverse distillation from one-class embedding,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 9737–9746

    2022

  49. [57]

    Catching both gray and black swans: Open-set supervised anomaly detection,

    C. Ding, G. Pang, and C. Shen, “Catching both gray and black swans: Open-set supervised anomaly detection,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 7388–7398. SUPPLEMENTARY MATERIAL 14 TABLE S1 ANOMALY PRIOR LIST PROVIDED ...

    2022

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.