Pith. sign in

REVIEW 3 major objections 6 minor 1 cited by

INP-Former++: Advancing Universal Anomaly Detection via Intrinsic Normal Prototypes and Residual Learning

T0 review · 3 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A test image carries its own normal reference: INP-Former++ extracts Intrinsic Normal Prototypes from the image and scores anomalies by reconstruction error, reaching 99.8 image-level AUROC on MVTec-AD multi-class detection.

desk verdict A clever self-referential approach to anomaly detection that works well on local defects, but the 'universal' claim rests on an untested locality prior. read the letter →

arxiv 2506.03660 v2 pith:SCPC4KIS submitted 2025-06-04 cs.CV

Wei Luo , Haiming Yao , Yunkang Cao , Qiyu Chen , Ang Gao , Weiming Shen , Wenyong Yu This is my paper
classification cs.CV
keywords anomalydetectionintrinsicnormalprototypescross-attentionfeaturereconstructionresiduallearninguniversalsemi-supervisedfew-shot
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 anomaly detection does not need a normality reference drawn from the training set: the test image itself contains enough normal information to serve as its own reference. The method builds six Intrinsic Normal Prototypes (INPs) from the image's own patch tokens, reconstructs the image from those prototypes, and scores every patch by reconstruction error. The claim is that this self-referential design solves the misalignment problem that hurts prototype-based and multi-class detectors, and that it generalizes across single-class, multi-class, few-shot, semi-supervised, and partially zero-shot settings. Reported results include 99.8 image-level AUROC on MVTec-AD multi-class detection, 97.9 in the 4-shot setting, 92.5 single-class on Real-IAD, and 96.8 AUPRO in semi-supervised localization.

What carries the argument

The central object is the Intrinsic Normal Prototype (INP): a small set, six by default, of feature vectors derived from the test image itself, each computed as a learned linear combination of that image's patch tokens through cross-attention with learnable query tokens. The INP Extractor produces these prototypes, the Soft INP Coherence Loss prevents all patch tokens from collapsing onto a single prototype, and the INP-guided Decoder consumes them as key-value pairs so that each reconstructed token is a combination of normal prototypes. The reconstruction residual between encoder and decoder features is the anomaly signal; in the extended version, a segmentation head trained on pseudo-anomaly residuals amplifies that signal. Since the number of prototypes $M$ is far smaller than the number of tokens $N$, the attention cost drops from $O(N^2 C)$ to $O(N M C)$.

What would settle it

Construct a test set in which the anomalous region covers most of the image, for example a uniform color, texture, or illumination change over the entire object surface, while training remains normal-only. INP-Former++ should fail to flag these as anomalous if the local-variation premise is load-bearing: the six INPs would absorb the global change and the decoder would reconstruct it with near-zero residual. The central claim is falsified if image-level AUROC on such a set is near chance or the residual map does not highlight the changed surface.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that intrinsic normal prototypes exist and can be extracted reliably: even in an anomalous image most patch tokens are normal, and a small set of prototypes aggregated from those tokens by cross-attention represents that normality well enough for reconstruction error to act as the anomaly score. The INP-guided decoder treats the INPs as key-value pairs, so every decoded token is a weighted combination of normal prototypes, while the first residual connection is removed and ReLU gates weak correlations to stop anomalous tokens from being copied through. Two training losses, the Soft INP Coherence Loss and the Soft Mining Loss, keep the prototypes faithful to normal regions and direct optimization toward hard normal regions. A separate residual-learning stage, trained on pseudo-anomalies while the reconstruction model is frozen, amplifies the difference between normal and anomalous features and extends the method to semi-supervised detection with a few real anomalies. This combination reaches 99.8 image-level AUROC for multi-class detection on MVTec-AD and improves localization over the conference version.

Load-bearing premise

The method depends on the assumption, stated in the abstract, that most anomalies are local variations, so even an anomalous test image still contains enough normal-looking regions to act as its own normality reference; if an anomaly is global or covers most of the image, the prototypes extracted from that same image will absorb the anomalous content and the decoder will reconstruct it, suppressing the residual signal.

Editorial extensions

If this is right

  • A single model trained jointly on 57 categories keeps nearly all of its per-dataset accuracy, so a unified industrial detector does not need to be rebuilt for each product class.
  • With only four normal samples per class in the few-shot setting, the method reaches 97.9 image-level AUROC on MVTec-AD, ahead of the SimCLIP vision-language baseline by 1.5 points.
  • Adding ten real anomalies per class in the semi-supervised setting raises localization AUPRO to 96.8 on MVTec-AD, so scarce real defects improve results without retraining the reconstruction model.
  • The INP extractor transfers across datasets: trained on Real-IAD it produces usable anomaly maps on unseen MVTec-AD classes, with an AUPRO of 79.6 in the zero-shot comparison.
  • Six INPs per image are enough, cutting attention complexity from $O(N^2 C)$ to $O(N M C)$ and letting a small variant run at 10.3 GFLOPs with competitive accuracy.

Reading between the lines

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

  • Beyond the paper, the self-referential principle extends naturally to video and medical volumes, where neighbouring frames, adjacent slices, or the contralateral structure in the same examination supply intrinsic normal patches; the paper tests only still images.
  • A testable extension is a global-anomaly stress set in which the anomalous region covers most of the image; performance should degrade gracefully as the abnormal area grows, since the framework's premise is that anomalies are local.
  • The anti-collapse Soft INP Coherence Loss is a general remedy for prototype collapse in attention-based models, and could be evaluated on other prototype-learning tasks beyond anomaly detection.
  • The paper attributes its suboptimal zero-shot results to missing language prompts; a natural next step is to inject a text prompt as an additional query alongside the learned INP tokens, a route the paper flags but does not implement.
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

3 major / 6 minor

Summary. The paper proposes INP-Former++, an extension of the conference version INP-Former, for universal anomaly detection. The key idea is to extract 'Intrinsic Normal Prototypes' (INPs) from the test image itself via cross-attention over patch tokens and to reconstruct image features under INP guidance, using reconstruction and residual errors as anomaly scores. The extensions over the conference version are a soft INP coherence loss, a soft mining loss, and a residual-learning segmentation head trained with pseudo-anomalies. The manuscript reports state-of-the-art results on MVTec-AD, VisA, Real-IAD, and Uni-Medical across single-class, multi-class, few-shot, semi-supervised, and partially zero-shot settings, with per-component ablations and computational complexity comparisons.

Significance. If the empirical claims hold, the work is a useful contribution: the self-referential normal-prototype idea is intuitive, the evaluation is broad (four datasets and five experimental regimes, including an honest re-computation of Dinomaly's pixel-level metrics), and the ablations isolate the contributions of each component. The complexity analysis is transparent. However, the central 'universal' claim is currently tied to an unvalidated locality assumption, and the reported margins over strong baselines lack uncertainty estimates, so the significance is conditional on additional stress-testing and robustness evidence.

major comments (3)
  1. [§3.2–§3.5, Eq. (1)–(9)] The claim that INPs represent normality rests on an untested locality prior. At inference, the INP Extractor aggregates all patch tokens of the test image via cross attention, and the INP coherence losses are training-only objectives; for a global anomaly (for example, a uniform illumination shift, a color cast, or a defect covering most of the image), the INPs will model the anomalous appearance, the INP-guided Decoder will reconstruct the image faithfully, and A_rec in Eq. (9) will be suppressed. This is load-bearing for the word 'universal' in the title and abstract, yet no experiment on global-anomaly benchmarks or synthesized global shifts is reported. Please add stress tests of this regime (e.g., uniform color/illumination perturbations and full-surface texture replacements on MVTec/VisA) and either qualify the universality claim or add an inference-time mechanism that cannot be satisfied by anomalously colored tokens.
  2. [§4.1.3, Tabs. 3–7, Fig. 8] The headline state-of-the-art margins are small (e.g., 0.2 I-AUROC over Dinomaly on MVTec-AD and VisA in Tab. 3) and no multiple-seed statistics or error bars are reported, so it is unclear whether the improvements are significant. In addition, the hyperparameters M, gamma, and lambda are analyzed only on MVTec-AD (Fig. 8) and the same values are used for VisA, Real-IAD, and Uni-Medical; the statement that these hyperparameters 'do not require any adjustment across the three datasets' needs at least a sensitivity check on a second dataset. Please report mean and standard deviation over at least three seeds and provide per-dataset hyperparameter sensitivity curves.
  3. [§3.4, Eq. (5)] The definition of the soft mining loss is incomplete. The gradient modification is written as ∇f̂^{l,t}_D(i,j) = ∇f^l_D(i,j) · w^l_t(i,j), but f̂^{l,t}_D is never defined as a feature; the subsequent losses L^cos_sm and L^mse_sm are written in terms of f̂^{l,cos}_D and f̂^{l,mse}_D, so the reader cannot tell whether the loss is evaluated on modified features or whether the gradient manipulation is applied only during backpropagation. Please provide the exact forward/backward formulation or pseudo-code for L_sm, since this is the central training objective and the description is not reproducible as written.
minor comments (6)
  1. [§3.5, Eq. (7)] Equation (7) contains 'f^a_res(i,j) == 1/L ...' where the double equals sign appears to be a typo for assignment; please correct it to '='.
  2. [§4.1.1, Tab. 2] The text states that Real-IAD has 36,645 normal training images, while Tab. 2 lists 36,345; please reconcile the discrepancy.
  3. [§3.5, Eq. (9)] The test-time anomaly map formula still carries the superscript 'a' on f^{l,a}_Q and f^{l,a}_D, which is left over from the pseudo-anomaly branch; since this is the deployed inference formula, the notation should be corrected to f^l_Q and f^l_D.
  4. [Fig. 2 caption] The caption contains the typo 'framwork' and should read 'framework'.
  5. [Abstract] The abstract states 'INP-Former achieves state-of-the-art performance' before introducing INP-Former++; the headline claims should be clearly attributed to the ++ version throughout.
  6. [§4.1.3 and reproducibility] No code availability statement is provided; given that the paper reports re-computed baseline results (footnote 1), releasing the evaluation code would materially strengthen reproducibility.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity; the self-referential INP mechanism is empirically grounded by normal-only training and held-out benchmarks.

full rationale

The paper's derivation chain is empirical rather than analytic: the final anomaly map A in Eq. 9 is the decoder reconstruction error plus a segmentation-head prediction, while the INPs in Eq. 1 are cross-attention aggregates of test-image tokens. This is self-referential in mechanism, but not circular in the evidentiary sense. The INP extractor and the coherence losses are applied during training on normal images only (Sec. 3.2, Eqs. 2 and 3 are training objectives), residual learning stops gradients into the reconstruction model (Sec. 3.5), and the performance claims are evaluated on held-out test images from MVTec-AD, VisA, Real-IAD, and Uni-Medical against external methods. No fitted parameter is renamed as a prediction, no load-bearing conclusion is imported from a self-citation chain, and no uniqueness theorem is invoked. The skeptic's global-anomaly concern is a scope and robustness limitation of the 'universal' claim: a full-image illumination shift or texture change could violate the paper's stated locality assumption, but that is a failure of an empirical premise, not a case where Eq. 9 reduces to its inputs by construction. Self-citations to the conference version [21] are used to report deltas and do not carry the central argument.

Assumptions & free parameters 5 free parameters · 4 assumptions · 1 invented entities

The central method is an empirical deep learning pipeline. The only hard free choices are the hyperparameters M, gamma, and lambda (tuned on MVTec-AD), plus standard training hyperparameters. The domain assumption that anomalies are local and test-image-internal normality is sufficient is load-bearing and not validated on global anomalies. The model also assumes DINOv2 features transfer across classes and that synthetic pseudo-anomalies are a good proxy for real defects. INPs are an internal representation with no external falsifiable handle.

free parameters (5)
  • Number of INPs M = 6
    Chosen by ablation on MVTec-AD (Fig. 8a); balances reconstruction fidelity and computational cost. The SOTA claim depends on this choice.
  • Temperature gamma in Soft Mining Loss = 3.0
    Chosen by ablation on MVTec-AD (Fig. 8b); controls focus on hard normal regions in Eq. (5).
  • Loss weight lambda for L_sc = 0.2
    Chosen by ablation on MVTec-AD (Fig. 8c); balances the soft INP coherence loss and reconstruction loss in Eq. (6).
  • Input resolution = 392x392 center crop
    Selected based on Tab. 10; higher resolution improves results but increases FLOPs. This preprocessing choice affects all reported numbers.
  • Image-level anomaly score pooling = average of top 1% values
    Following [70], the image-level score is the average of the top 1% pixel values in the final anomaly map; this hand-chosen pooling affects reported I-AUROC.
assumptions (4)
  • domain assumption Most anomalies are local variations; a test image contains enough normal content to serve as prototypes.
    Stated in the Abstract and Introduction; if false, INPs may be contaminated and reconstruction will not expose anomalies.
  • domain assumption The INP Extractor, trained only with coherence losses on normal data, generalizes to unseen test images and classes without inference-time coherence enforcement.
    The coherence losses (Eqs. 2 and 3) are training objectives; at test time nothing prevents the cross-attention from including anomalous tokens. The zero-shot cross-dataset results (Tab. 13) provide partial evidence but no guarantee.
  • domain assumption Fixed DINOv2 ViT features encode semantic normal patterns that are useful across industrial and medical categories.
    All features come from a frozen ViT-Base/14 with DINOv2 weights (Section 4.1.3); the method inherits the feature quality and transferability of this encoder.
  • domain assumption Pseudo-anomalies synthesized from Perlin noise and DTD textures are a valid proxy for real anomalies in residual learning.
    Section 3.5 trains the segmentation head only on synthetic anomalies (plus 10 real anomalies in semi-supervised); if the synthetic distribution is unrepresentative, the residual amplifier may mislocalize real defects.
invented entities (1)
  • Intrinsic Normal Prototypes (INPs)
    purpose: Learned prototypes that aggregate normal tokens from the test image and guide the decoder to reconstruct only normal regions.
    INPs are an internal latent representation. The paper provides visualizations (Fig. 9) and ablations (Tab. 9) showing usefulness, but there is no external, falsifiable prediction tied to INPs independent of the proposed framework.

how reviews work

0 comments
Cite this review

Pith. "Pith review of INP-Former++: Advancing Universal Anomaly Detection via Intrinsic Normal Prototypes and Residual Learning." pith.science (2026). https://pith.science/paper/SCPC4KIS

@misc{pith2026250603660,
  author       = {Pith},
  title        = {Pith review of: INP-Former++: Advancing Universal Anomaly Detection via Intrinsic Normal Prototypes and Residual Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SCPC4KIS}},
  note         = {Machine review of arXiv:2506.03660}
}
read the original abstract

Anomaly detection (AD) is essential for industrial inspection and medical diagnosis, yet existing methods typically rely on ``comparing'' test images to normal references from a training set. However, variations in appearance and positioning often complicate the alignment of these references with the test image, limiting detection accuracy. We observe that most anomalies manifest as local variations, meaning that even within anomalous images, valuable normal information remains. We argue that this information is useful and may be more aligned with the anomalies since both the anomalies and the normal information originate from the same image. Therefore, rather than relying on external normality from the training set, we propose INP-Former, a novel method that extracts Intrinsic Normal Prototypes (INPs) directly from the test image. Specifically, we introduce the INP Extractor, which linearly combines normal tokens to represent INPs. We further propose an INP Coherence Loss to ensure INPs can faithfully represent normality for the testing image. These INPs then guide the INP-guided Decoder to reconstruct only normal tokens, with reconstruction errors serving as anomaly scores. Additionally, we propose a Soft Mining Loss to prioritize hard-to-optimize samples during training. INP-Former achieves state-of-the-art performance in single-class, multi-class, and few-shot AD tasks across MVTec-AD, VisA, and Real-IAD, positioning it as a versatile and universal solution for AD. Remarkably, INP-Former also demonstrates some zero-shot AD capability. Furthermore, we propose a soft version of the INP Coherence Loss and enhance INP-Former by incorporating residual learning, leading to the development of INP-Former++. The proposed method significantly improves detection performance across single-class, multi-class, semi-supervised, few-shot, and zero-shot settings.

Figures

Figures reproduced from arXiv: 2506.03660 by the authors.

Figure 1
Figure 1. Motivation for Intrinsic Normal Prototypes (INPs). (a) Pre-stored prototypes from few-shot normal samples may fail to represent all normal patterns. (b) Pre-stored prototypes from one class can be similar to anomalies in another class. (c) The extracted INPs are concise yet well-aligned to the test image, alleviating the issues in (a) and (b). Normal Prototypes (INPs). As illustrated in [PITH_FULL_IMAGE:figures/ful… view at source ↗
Figure 2
Figure 2. Overview of our INP-Former++ framwork for universal anomaly detection. Our model consists of a pre-trained Encoder, an INP Extractor, a Bottleneck, an INP-guided decoder, and a Segmentation (Seg) head. (a) Normal Pattern Modeling. The INP Extractor dynamically extracts intrinsic normal prototypes from a single image, which the INP-guided Decoder leverages to effectively suppress anomalous features. (b) Residual Lear… view at source ↗
Figure 3
Figure 3. Qualitative results of anomaly localization on the MVTec-AD [4], VisA [5], and Real-IAD [71] datasets for multi-class anomaly detection. From top to bottom, we show the input anomalous images along with their ground truth masks, the predicted anomaly maps from INP-Former, and the results from the improved INP-Former++ model. TABLE 4 Super-multi-class anomaly detection performance on different AD datasets. ∆ represen… view at source ↗
Figures from the paper (6 more)
Figure 5
Figure 5. Figure 5: Visualization of the impact of soft INP coherence loss Lsc. (a) Input anomalous image and ground truth. (b) Distance map without Lc. (c) Distance map with Lc. (d) Distance map with Lsc [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 6
Figure 6. Figure 6: Visualization of the impact of soft mining loss Lsm. We plot the Kernel Density Estimation (KDE) for the leather and cable categories in the MVTec-AD [4] dataset to estimate the probability density of the anomaly scores [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]
Figure 7
Figure 7. Figure 7: Visualization of the impact of residual learning (“RL”). (a) Input anomalous image and ground truth. (b) Predicted anomaly map without “RL”. (c) Predicted anomaly map with “RL”. 4.3 Ablation Study 4.3.1 Overall Ablation As presented in Tab. 9, we conduct comprehensive …
Figure 8
Figure 8. Figure 8: Ablation analysis of model hyperparameters on the MVTec-AD [4] dataset. (a) The number of INPs M. (b) The temperature parameter γ in Lsm, (c) The loss weight λ. input features into a few representative INPs, which are then used by the INP-guided Decoder to reconstruct …
Figure 9
Figure 9. Figure 9: Visualizations of INPs. (a) Input anomalous image and ground truth. (b)-(g) Cross attention maps between INPs and image patches. version, INP-Former, the proposed INP-Former++ intro￾duces only an additional 2.9M parameters and 7.4G FLOPs, yet achieves a 1.4↑ performanc…
Figure 10
Figure 10. Figure 10: Zero-shot anomaly detection results. (a) Input anomalous image and ground truth. (b) Anomaly map from INP-Former (w/ Lc). (c) Anomaly map from INP-Former++ (w/ Lsc). Both INP-Former and INP-Former++ are trained on the Real-IAD [71] dataset and evaluated on MVTec-AD [4…

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. ReFP-AD: Rectified Flow Preconditioning for Energy-Based Anomaly Detection

    cs.LG 2026-08 conditional novelty 6.0 of 10

    ReFP-AD uses rectified-flow preconditioning to make finite-step MCMC stable for energy-based anomaly detection on full-dimensional DINOv2 tokens, achieving strong AUROC on MVTec-AD and VisA.

Reference graph

Works this paper leans on

79 extracted references · 64 canonical work pages · cited by 1 Pith paper

  1. [1]

    IM-IAD: Industrial image anomaly detection benchmark in manufacturing,

    G. Xie, J. Wang, J. Liu, J. Lyu, Y. Liu, C. Wang, F. Zheng, and Y. Jin, “IM-IAD: Industrial image anomaly detection benchmark in manufacturing,”IEEE Transactions on Cybernetics, pp. 1–14, 2024

    work page 2024

  2. [2]

    A comprehensive library for bench- marking multi-class visual anomaly detection,

    J. Zhang, H. He, Z. Gan, Q. He, Y. Cai, Z. Xue, Y. Wang, C. Wang, L. Xie, and Y. Liu, “A comprehensive library for bench- marking multi-class visual anomaly detection,”arXiv preprint arXiv:2406.03262, 2024

    arXiv 2024

  3. [3]

    A survey on visual anomaly detection: Challenge, approach, and prospect,

    Y. Cao, X. Xu, J. Zhang, Y. Cheng, X. Huang, G. Pang, and W. Shen, “A survey on visual anomaly detection: Challenge, approach, and prospect,”arXiv preprint arXiv:2401.16402, 2024

    arXiv 2024

  4. [4]

    The MVTec anomaly detection dataset: A comprehensive real- world dataset for unsupervised anomaly detection,

    P . Bergmann, K. Batzner, M. Fauser, D. Sattlegger, and C. Steger, “The MVTec anomaly detection dataset: A comprehensive real- world dataset for unsupervised anomaly detection,”International Journal of Computer Vision, vol. 129, no. 4, pp. 1038–1059, 2021

    work page 2021

  5. [5]

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

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

    work page 2022

  6. [6]

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

    C. Huang, A. Jiang, J. Feng, Y. Zhang, X. Wang, and Y. Wang, “Adapting visual-language models for generalizable anomaly de- tection in medical images,” inProceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition, 2024, pp. 11 375– 11 385. 14

    work page 2024

  7. [7]

    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,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 14 318–14 328

    work page 2022

  8. [8]

    Template-based feature aggregation network for industrial anomaly detection,

    W. Luo, H. Yao, and W. Yu, “Template-based feature aggregation network for industrial anomaly detection,”Engineering Applica- tions of Artificial Intelligence, vol. 131, p. 107810, 2024

    work page 2024

Show all 79 references
  1. [9]

    Registration based few-shot anomaly detection,

    C. Huang, H. Guan, A. Jiang, Y. Zhang, M. Spratling, and Y.- F. Wang, “Registration based few-shot anomaly detection,” in European Conference on Computer Vision. Springer, 2022, pp. 303– 319

    2022

  2. [10]

    Prototypical learning guided context-aware segmentation network for few-shot anomaly detec- tion,

    Y. Jiang, Y. Cao, and W. Shen, “Prototypical learning guided context-aware segmentation network for few-shot anomaly detec- tion,”IEEE Transactions on Neural Networks and Learning Systems, pp. 1–11, 2024

    2024

  3. [11]

    A unified model for multi-class anomaly detection,

    Z. You, L. Cui, Y. Shen, K. Yang, X. Lu, Y. Zheng, and X. Le, “A unified model for multi-class anomaly detection,” inAdvances in Neural Information Processing Systems, 2022, pp. 4571–4584

    2022

  4. [12]

    Dinomaly: The less is more philosophy in multi-class unsupervised anomaly detection,

    J. Guo, S. Lu, W. Zhang, F. Chen, H. Liao, and H. Li, “Dinomaly: The less is more philosophy in multi-class unsupervised anomaly detection,”arXiv preprint arXiv:2405.14325, 2024

    2024 arXiv

  5. [13]

    Prior normality prompt transformer for multiclass industrial image anomaly detection,

    H. Yao, Y. Cao, W. Luo, W. Zhang, W. Yu, and W. Shen, “Prior normality prompt transformer for multiclass industrial image anomaly detection,”IEEE Transactions on Industrial Informatics, vol. 20, no. 10, pp. 11 866–11 876, 2024

    2024

  6. [14]

    Fo- cus your distribution: Coarse-to-fine non-contrastive learning for anomaly detection and localization,

    Y. Zheng, X. Wang, R. Deng, T. Bao, R. Zhao, and L. Wu, “Fo- cus your distribution: Coarse-to-fine non-contrastive learning for anomaly detection and localization,” in2022 IEEE International Conference on Multimedia and Expo (ICME). IEEE, 2022, pp. 1–6

    2022

  7. [15]

    Template-guided hierarchical feature restoration for anomaly detection,

    H. Guo, L. Ren, J. Fu, Y. Wang, Z. Zhang, C. Lan, H. Wang, and X. Hou, “Template-guided hierarchical feature restoration for anomaly detection,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 6447–6458

    2023

  8. [16]

    Hierarchical vector quantized transformer for multi-class un- supervised anomaly detection,

    R. Lu, Y. Wu, L. Tian, D. Wang, B. Chen, X. Liu, and R. Hu, “Hierarchical vector quantized transformer for multi-class un- supervised anomaly detection,”Advances in Neural Information Processing Systems, vol. 36, pp. 8487–8500, 2023

    2023

  9. [17]

    Hierarchical gaussian mixture normalizing flow modeling for unified anomaly detection,

    X. Yao, R. Li, Z. Qian, L. Wang, and C. Zhang, “Hierarchical gaussian mixture normalizing flow modeling for unified anomaly detection,”arXiv preprint arXiv:2403.13349, 2024

    2024 arXiv

  10. [18]

    Zero-shot versus many- shot: Unsupervised texture anomaly detection,

    T. Aota, L. T. T. Tong, and T. Okatani, “Zero-shot versus many- shot: Unsupervised texture anomaly detection,” in2023 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2023, pp. 5553–5561

    2023

  11. [19]

    Global- regularized neighborhood regression for efficient zero-shot texture anomaly detection,

    H. Yao, W. Luo, Y. Cao, Y. Zhang, W. Yu, and W. Shen, “Global- regularized neighborhood regression for efficient zero-shot texture anomaly detection,”arXiv preprint arXiv:2406.07333, 2024

    2024 arXiv

  12. [20]

    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,” in2017 IEEE International Conference on Computer Vision (ICCV), 2017, pp. 2999–3007

    2017

  13. [21]

    Exploring intrinsic normal prototypes within a single image for universal anomaly detection,

    W. Luo, Y. Cao, H. Yao, X. Zhang, J. Lou, Y. Cheng, W. Shen, and W. Yu, “Exploring intrinsic normal prototypes within a single image for universal anomaly detection,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025

    2025

  14. [22]

    Prototypical residual networks for anomaly detection and localization,

    H. Zhang, Z. Wu, Z. Wang, Z. Chen, and Y.-G. Jiang, “Prototypical residual networks for anomaly detection and localization,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 16 281–16 291

    2023

  15. [24]

    Simclip: Refining image-text alignment with simple prompts for zero-/few-shot anomaly detection,

    C. Deng, H. Xu, X. Chen, H. Xu, X. Tu, X. Ding, and Y. Huang, “Simclip: Refining image-text alignment with simple prompts for zero-/few-shot anomaly detection,” inProceedings of the 32nd ACM International Conference on Multimedia, 2024, pp. 1761–1770

    2024

  16. [25]

    BiaS: Incorporating biased knowledge to boost unsupervised image anomaly localiza- tion,

    Y. Cao, X. Xu, C. Sun, L. Gao, and W. Shen, “BiaS: Incorporating biased knowledge to boost unsupervised image anomaly localiza- tion,”IEEE Transactions on Systems, Man, and Cybernetics: Systems, 2023

    2023

  17. [26]

    AMI-Net: Adaptive mask inpainting network for industrial anomaly detection and local- ization,

    W. Luo, H. Yao, W. Yu, and Z. Li, “AMI-Net: Adaptive mask inpainting network for industrial anomaly detection and local- ization,”IEEE Transactions on Automation Science and Engineering, 2024

    2024

  18. [27]

    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

  19. [28]

    Anomaly detection via reverse distillation from one-class embedding,

    H. Deng and X. Li, “Anomaly detection via reverse distillation from one-class embedding,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 9737–9746

    2022

  20. [29]

    Revisiting reverse distillation for anomaly detection,

    T. D. Tien, A. T. Nguyen, N. H. Tran, T. D. Huy, S. T. Duong, C. D. T. Nguyen, and S. Q. H. Truong, “Revisiting reverse distillation for anomaly detection,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2023, pp. 24 511–24 520

    2023

  21. [30]

    Softpatch: Unsupervised anomaly detection with noisy data,

    X. Jiang, J. Liu, J. Wang, Q. Nie, K. Wu, Y. Liu, C. Wang, and F. Zheng, “Softpatch: Unsupervised anomaly detection with noisy data,”Advances in Neural Information Processing Systems, vol. 35, pp. 15 433–15 445, 2022

    2022

  22. [31]

    Softpatch+: Fully unsupervised anomaly classification and segmentation,

    C. Wang, X. Jiang, B.-B. Gao, Z. Gan, Y. Liu, F. Zheng, and L. Ma, “Softpatch+: Fully unsupervised anomaly classification and segmentation,”Pattern Recognition, vol. 161, p. 111295, 2025

    2025

  23. [32]

    Reconpatch: Contrastive patch representation learning for indus- trial anomaly detection,

    J. Hyun, S. Kim, G. Jeon, S. H. Kim, K. Bae, and B. J. Kang, “Reconpatch: Contrastive patch representation learning for indus- trial anomaly detection,” inProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2024, pp. 2052–2061

    2024

  24. [33]

    Simplenet: A simple network for image anomaly detection and localization,

    Z. Liu, Y. Zhou, Y. Xu, and Z. Wang, “Simplenet: A simple network for image anomaly detection and localization,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 20 402–20 411

    2023

  25. [34]

    Dest- seg: Segmentation guided denoising student-teacher for anomaly detection,

    X. Zhang, S. Li, X. Li, P . Huang, J. Shan, and T. Chen, “Dest- seg: Segmentation guided denoising student-teacher for anomaly detection,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 3914–3923

    2023

  26. [35]

    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

  27. [36]

    Progressive boundary guided anomaly synthesis for industrial anomaly detec- tion,

    Q. Chen, H. Luo, H. Gao, C. Lv, and Z. Zhang, “Progressive boundary guided anomaly synthesis for industrial anomaly detec- tion,”IEEE Transactions on Circuits and Systems for Video Technology, vol. 35, no. 2, pp. 1193–1208, 2025

    2025

  28. [37]

    Center-aware residual anomaly synthesis for multi-class indus- trial anomaly detection,

    Q. Chen, H. Luo, H. Yao, W. Luo, Z. Qu, C. Lv, and Z. Zhang, “Center-aware residual anomaly synthesis for multi-class indus- trial anomaly detection,”arXiv preprint arXiv:2505.17551, 2025

    2025 arXiv

  29. [38]

    An image is worth 16x16 words: Transformers for image recognition at scale,

    A. Dosovitskiy, L. Beyer, A. Kolesnikovet al., “An image is worth 16x16 words: Transformers for image recognition at scale,” in International Conference on Learning Representations, 2021

    2021

  30. [39]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P . Abbeel, “Denoising diffusion probabilistic models,”Advances in neural information processing systems, vol. 33, pp. 6840–6851, 2020

    2020

  31. [40]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agar- wal, G. Sastry, A. Askell, P . Mishkin, J. Clarket al., “Learning transferable visual models from natural language supervision,” in International Conference on Machine Learning, 2021, pp. 8748–8763

    2021

  32. [41]

    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,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023, pp. 19 606–19 616

    2023

  33. [42]

    Anoma- lygpt: Detecting industrial anomalies using large vision-language models,

    Z. Gu, B. Zhu, G. Zhu, Y. Chen, M. Tang, and J. Wang, “Anoma- lygpt: Detecting industrial anomalies using large vision-language models,” inProceedings of the AAAI Conference on Artificial Intelli- gence, vol. 38, no. 3, 20 24, pp. 1932–1940

    1932

  34. [43]

    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,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 17 826–17 836

    2024

  35. [44]

    Adaclip: Adapting clip with hybrid learnable prompts for zero-shot anomaly detection,

    Y. Cao, J. Zhang, L. Frittoli, Y. Cheng, W. Shen, and G. Borac- chi, “Adaclip: Adapting clip with hybrid learnable prompts for zero-shot anomaly detection,” inEuropean Conference on Computer Vision. Springer, 2024

    2024

  36. [45]

    MambaAD: Exploring state space models for multi-class unsupervised anomaly detection,

    H. He, Y. Bai, J. Zhang, Q. He, H. Chen, Z. Gan, C. Wang, X. Li, G. Tian, and L. Xie, “MambaAD: Exploring state space models for multi-class unsupervised anomaly detection,” inThe Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. [Online]. Availab...

    2024

  37. [46]

    Mamba: Linear-time sequence modeling with selective state spaces,

    A. Gu and T. Dao, “Mamba: Linear-time sequence modeling with selective state spaces,”arXiv preprint arXiv:2312.00752, 2023

    2023 arXiv

  38. [47]

    Vision trans- formers need registers,

    T. Darcet, M. Oquab, J. Mairal, and P . Bojanowski, “Vision trans- formers need registers,”arXiv preprint arXiv:2309.16588, 2023. 15

    2023 arXiv

  39. [48]

    Deep semi-supervised anomaly detection,

    L. Ruff, R. A. Vandermeulen, N. G¨ornitz, A. Binder, E. M¨uller, K.-R. M ¨uller, and M. Kloft, “Deep semi-supervised anomaly detection,” arXiv preprint arXiv:1906.02694, 2019

    1906 arXiv

  40. [49]

    Deep one-class classification,

    L. Ruff, R. Vandermeulen, N. Goernitz, L. Deecke, S. A. Siddiqui, A. Binder, E. M¨uller, and M. Kloft, “Deep one-class classification,” inInternational conference on machine learning. PMLR, 2018, pp. 4393–4402

    2018

  41. [50]

    Rethinking assumptions in deep anomaly detection,

    L. Ruff, R. A. Vandermeulen, B. J. Franks, K.-R. M ¨uller, and M. Kloft, “Rethinking assumptions in deep anomaly detection,” arXiv preprint arXiv:2006.00339, 2020

    2006 arXiv

  42. [51]

    Explainable deep few-shot anomaly detection with deviation networks,

    G. Pang, C. Ding, C. Shen, and A. v. d. Hengel, “Explainable deep few-shot anomaly detection with deviation networks,”arXiv preprint arXiv:2108.00462, 2021

    2021 arXiv

  43. [52]

    Viral pneumonia screening on chest x-rays using confidence-aware anomaly detection,

    J. Zhang, Y. Xie, G. Pang, Z. Liao, J. Verjans, W. Li, Z. Sun, J. He, Y. Li, C. Shenet al., “Viral pneumonia screening on chest x-rays using confidence-aware anomaly detection,”IEEE transactions on medical imaging, vol. 40, no. 3, pp. 879–890, 2020

    2020

  44. [53]

    Prototypical networks for few- shot learning,

    J. Snell, K. Swersky, and R. Zemel, “Prototypical networks for few- shot learning,”Advances in neural information processing systems, vol. 30, 2017

    2017

  45. [54]

    Adaptive prototype learning and allocation for few-shot segmen- tation,

    G. Li, V . Jampani, L. Sevilla-Lara, D. Sun, J. Kim, and J. Kim, “Adaptive prototype learning and allocation for few-shot segmen- tation,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 8334–8343

    2021

  46. [55]

    Learning memory-guided normality for anomaly detection,

    H. Park, J. Noh, and B. Ham, “Learning memory-guided normality for anomaly detection,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 14 372–14 381

    2020

  47. [56]

    Learning normal dynamics in videos with meta prototype network,

    H. Lv, C. Chen, Z. Cui, C. Xu, Y. Li, and J. Yang, “Learning normal dynamics in videos with meta prototype network,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 15 425–15 434

    2021

  48. [57]

    Pixel-level anomaly detection via uncertainty-aware prototypical transformer,

    C. Huang, C. Liu, Z. Zhang, Z. Wu, J. Wen, Q. Jiang, and Y. Xu, “Pixel-level anomaly detection via uncertainty-aware prototypical transformer,” inProceedings of the 30th ACM International Conference on Multimedia, 2022, pp. 521–530

    2022

  49. [58]

    Memorizing normality to detect anomaly: Memory-augmented deep autoencoder for unsupervised anomaly detection,

    D. Gong, L. Liu, V . Le, B. Saha, M. R. Mansour, S. Venkatesh, and A. v. d. Hengel, “Memorizing normality to detect anomaly: Memory-augmented deep autoencoder for unsupervised anomaly detection,” inProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV),...

    2019

  50. [59]

    Learning to detect multi-class anomalies with just one normal image prompt,

    B.-B. Gao, “Learning to detect multi-class anomalies with just one normal image prompt,” inEuropean Conference on Computer Vision. Springer, 2024, pp. 454–470

    2024

  51. [60]

    One-for-all: Proposal masked cross-class anomaly detection,

    X. Yao, C. Zhang, R. Li, J. Sun, and Z. Liu, “One-for-all: Proposal masked cross-class anomaly detection,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 37, no. 4, 2023, pp. 4792– 4800

    2023

  52. [61]

    Resad: A simple framework for class generalizable anomaly detection,

    X. Yao, Z. Chen, C. Gao, G. Zhai, and C. Zhang, “Resad: A simple framework for class generalizable anomaly detection,”Advances in Neural Information Processing Systems, vol. 37, pp. 125 287–125 311, 2024

    2024

  53. [62]

    Learning memory-guided normality for anomaly detection,

    H. Park, J. Noh, and B. Ham, “Learning memory-guided normality for anomaly detection,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020

    2020

  54. [63]

    Recontrast: Domain- specific anomaly detection via contrastive reconstruction,

    J. Guo, S. Lu, L. Jia, W. Zhang, and H. Li, “Recontrast: Domain- specific anomaly detection via contrastive reconstruction,”Ad- vances in Neural Information Processing Systems, vol. 36, pp. 10 721– 10 740, 2023

    2023

  55. [64]

    Sparse self-attention transformer for image inpainting,

    W. Huang, Y. Deng, S. Hui, Y. Wu, S. Zhou, and J. Wang, “Sparse self-attention transformer for image inpainting,”Pattern Recogni- tion, vol. 145, p. 109897, 2024

    2024

  56. [65]

    Collaborative discrepancy optimization for reliable image anomaly localization,

    Y. Cao, X. Xu, Z. Liu, and W. Shen, “Collaborative discrepancy optimization for reliable image anomaly localization,”IEEE Trans- actions on Industrial Informatics, pp. 1–10, 2023

    2023

  57. [66]

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

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

    2021

  58. [67]

    An image synthesizer,

    K. Perlin, “An image synthesizer,”ACM Siggraph Computer Graph- ics, vol. 19, no. 3, pp. 287–296, 1985

    1985

  59. [68]

    De- scribing textures in the wild,

    M. Cimpoi, S. Maji, I. Kokkinos, S. Mohamed, and A. Vedaldi, “De- scribing textures in the wild,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2014, pp. 3606–3613

    2014

  60. [69]

    Learn to segment retinal lesions and beyond,

    Q. Wei, X. Li, W. Yu, X. Zhang, Y. Zhang, B. Hu, B. Mo, D. Gong, N. Chen, D. Dinget al., “Learn to segment retinal lesions and beyond,” in2020 25th International conference on pattern recognition (ICPR). IEEE, 2021, pp. 7403–7410

    2021

  61. [70]

    Msflow: Multi- scale flow-based framework for unsupervised anomaly detection,

    Y. Zhou, X. Xu, J. Song, F. Shen, and H. T. Shen, “Msflow: Multi- scale flow-based framework for unsupervised anomaly detection,” IEEE Transactions on Neural Networks and Learning Systems, vol. 36, no. 2, pp. 2437–2450, 2025

    2025

  62. [71]

    Real-iad: A real-world multi-view dataset for benchmarking versatile industrial anomaly detection,

    C. Wang, W. Zhu, B.-B. Gao, Z. Gan, J. Zhang, Z. Gu, S. Qian, M. Chen, and L. Ma, “Real-iad: A real-world multi-view dataset for benchmarking versatile industrial anomaly detection,” inPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, p...

    2024

  63. [72]

    Exploring plain vit reconstruction for multi-class unsupervised anomaly detection,

    J. Zhang, X. Chen, Y. Wang, C. Wang, Y. Liu, X. Li, M.-H. Yang, and D. Tao, “Exploring plain vit reconstruction for multi-class unsupervised anomaly detection,”arXiv preprint arXiv:2312.07495, 2023

    2023 arXiv

  64. [73]

    Stable and low-precision training for large-scale vision-language models,

    M. Wortsman, T. Dettmers, L. Zettlemoyer, A. Morcos, A. Farhadi, and L. Schmidt, “Stable and low-precision training for large-scale vision-language models,”Advances in Neural Information Processing Systems, vol. 36, pp. 10 271–10 298, 2023

    2023

  65. [74]

    A diffusion-based framework for multi- class anomaly detection,

    H. He, J. Zhang, H. Chen, X. Chen, Z. Li, X. Chen, Y. Wang, C. Wang, and L. Xie, “A diffusion-based framework for multi- class anomaly detection,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 38, no. 8, 2024, pp. 8472–8480

    2024

  66. [75]

    Sub-image anomaly detection with deep pyramid correspondences,

    N. Cohen and Y. Hoshen, “Sub-image anomaly detection with deep pyramid correspondences,”arXiv preprint arXiv:2005.02357, 2020

    2005 arXiv

  67. [76]

    Padim: a patch distribution modeling framework for anomaly detection and localization,

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

    2021

  68. [77]

    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 (CVPR), June 2024, pp. 16 838–16 848

    2024

  69. [78]

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

    X. Yao, R. Li, J. Zhang, J. Sun, and C. Zhang, “Explicit bound- ary 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

  70. [79]

    Supersimplenet: Unifying unsupervised and supervised learning for fast and reliable surface defect detection,

    B. Rolih, M. Fu ˇcka, and D. Sko ˇcaj, “Supersimplenet: Unifying unsupervised and supervised learning for fast and reliable surface defect detection,” inInternational Conference on Pattern Recognition. Springer, 2025, pp. 47–65

    2025

  71. [80]

    A zero-/few-shot anomaly clas- sification 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,

    X. Chen, Y. Han, and J. Zhang, “A zero-/few-shot anomaly clas- sification 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, vol. 2, no. 4, 2023

    2023 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.