Pith. sign in

REVIEW 3 major objections 5 minor 28 references

Pro-AD: Learning Comprehensive Prototypes with Prototype-based Constraint for Multi-class Unsupervised Anomaly Detection

T0 review · 3 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read Pro-AD claims a large prototype set plus a prototype-derived constraint solves the Soft Identity Mapping problem and reaches state-of-the-art multi-class anomaly detection scores on three industrial benchmarks.

desk verdict Real empirical gains on three benchmarks, but the paper's core mechanism—the Prototype-based Constraint—is an input-dependent residual, not a normal-only constraint, so the central theoretical claim doesn't hold as stated. read the letter →

arxiv 2506.13097 v3 pith:4QR44GTJ submitted 2025-06-16 cs.CV

Ziqing Zhou , Yurui Pan , Lidong Wang , Wenbing Zhu , Mingmin Chi , Dong Wu , Bo Peng This is my paper · ORCID
classification cs.CV
keywords unsupervisedanomalydetectionmulti-classprototypelearningreconstruction-basedsoftidentitymappingdynamicbidirectionaldecoderlinearcross-attentionindustrial
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

Pro-AD is a reconstruction-based method for multi-class unsupervised anomaly detection: one model trained only on normal images must flag defects across many object classes. The paper argues that prototype-based reconstruction faces a trade-off: too few learnable prototypes cannot represent normal variation, while too many let the attention mechanism reconstruct anomalies as well as normal regions, a failure it calls the "Soft Identity Mapping" problem. Pro-AD's proposal is to use a large set of learnable prototypes plus a Prototype-based Constraint that forces the final reconstruction to be composed of a weighted prototype combination and a regularized prototype term, so anomalies cannot be rebuilt by a simple weighted mix. On MVTec-AD, VisA, and Real-IAD the method reports state-of-the-art or near-top image- and pixel-level scores, with the largest gains at the pixel level.

What carries the argument

The load-bearing object is the Prototype-based Constraint, defined as $\mathbf{P}^{l+1}_{\rm reg} = \mathrm{FFN}(\mathbf{P}^{l+1})$ added to the attention reconstruction $\mathbf{f}^l_{\rm rec}$, giving $\mathbf{f}^l_D = \mathbf{f}^l_{\rm rec} + \mathbf{P}^{l+1}_{\rm reg}$. This term recasts the attention reconstruction as learning a residual complementary to a prototype-derived constant, so no single anomaly can be rebuilt by weighted combination alone. It works inside a Dynamic Bidirectional Decoder whose prototype-update and target-reconstruction branches share weights and use linear cross-attention, with an Adaptive Noisy Bottleneck keeping prototypes and query features in the same feature space.

What would settle it

Train Pro-AD on MVTec-AD, then remove the Prototype-based Constraint (set $\mathbf{P}^{l+1}_{\rm reg}=0$) and measure image-level AUROC on anomaly classes with large spatial extent, such as logical anomalies. If the gain from the constraint disappears or reverses on a broader set of classes, or if a test shows an anomaly being reconstructed through $\mathbf{P}^{l+1}_{\rm reg}$ alone, the mechanism is not doing what the paper claims.

Watch

Extended reading notes

Core claim

The central claim is that the "Soft Identity Mapping" problem, where an enlarged prototype set gives the attention mechanism enough expressive power to reconstruct anomalous features, can be broken by adding a Prototype-based Constraint inside the decoder. The reconstruction of each target feature is written as the sum of a linear-cross-attention reconstruction from prototypes and a per-position regularized term derived from the same prototypes through a feed-forward network; this changes the learning objective so that the attention part must reconstruct the encoder feature minus the regularized term, preventing the prototypes from acting as an identity map. With this constraint, an expanded set of prototypes (as many as there are image patches) can safely supply comprehensive normal information, and the paper reports the method outperforms prior multi-class anomaly detectors on three industrial benchmarks, particularly in pixel-level anomaly localization.

Load-bearing premise

The method assumes that the learned term added to the reconstruction truly limits how well anomalies can be rebuilt, rather than simply being another learned channel that can encode anomaly information; the paper gives no proof of this and tests it on one dataset only.

Editorial extensions

If this is right

  • If Pro-AD's mechanism is correct, pixel-level anomaly localization improves enough to matter for industrial inspection, with pixel-level AP and F1 gains of several points over the prior best on MVTec-AD.
  • A single multi-class model can cover diverse industrial categories without a per-class memory bank, since the prototypes are shared and learned rather than stored per class.
  • The Prototype-based Constraint gives a generic recipe for prototype-based reconstruction models: constrain the reconstruction output by a prototype-derived term, not just by the number of prototypes.
  • The method's robustness on Real-IAD, the largest benchmark, suggests the approach scales to realistic multi-view industrial settings.

Reading between the lines

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

  • An implication the paper leaves implicit is that the constraint deserves testing on logical-anomaly datasets like MVTec-LOCO, where the paper itself notes limited performance; such a test would show whether the mechanism mainly helps structural defects or generalizes to context-dependent anomalies.
  • Because the constraint is an additive learned residual, a natural extension is to bound or normalize $\mathbf{P}^{l+1}_{\rm reg}$ so it cannot exceed a fixed budget, turning the soft constraint into a true capacity limit.
  • The shared weights between prototype learning and reconstruction suggest the same decoder could be reused for few-shot or zero-shot anomaly detection if the initial prototypes are replaced by a small set of normal exemplars.
  • If the residual interpretation is correct, similar constraints could improve other attention-based reconstruction models, not just prototype-based ones, by forcing the attention output to be complementary to a fixed reference.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. This paper presents Pro-AD, a multi-class unsupervised anomaly detection method that learns an expanded set of learnable prototypes and reconstructs target features with a dynamic bidirectional decoder. The main novel component is a Prototype-based Constraint that adds a feed-forward transformation of the updated prototypes to the naive reconstruction, with the stated goal of preventing anomalies from being reconstructed through attention, a failure mode the authors call 'Soft Identity Mapping.' The paper reports state-of-the-art or near-state-of-the-art image- and pixel-level metrics on MVTec-AD, VisA, and Real-IAD, together with an ablation study on MVTec-AD.

Significance. The paper addresses a relevant problem in multi-class unsupervised anomaly detection and reports strong empirical results. If the proposed mechanism were established, the simple recipe of expanded prototypes plus a residual constraint would be practically useful. The main weakness is that the central mechanistic claim is not supported by the equations: the prototype update and the Prototype-based Constraint both depend on the potentially anomalous query feature. The empirical evidence is also limited by the absence of error bars, the lack of code release, per-dataset hyperparameter choices, and a one-dataset ablation. The paper does not provide machine-checked proofs or reproducible code, so the SOTA claims rest entirely on the reported tables.

major comments (3)
  1. [Section 3.4, Eqs. (4)-(9)] The Prototype-based Constraint is not a normal-only reference. In Eq. (4), the updated prototype P^{l+1} is computed by cross-attending to the target feature Q^l, so at inference P^{l+1} is a function of the potentially anomalous query. Consequently, both terms in Eq. (9), f_D^l = f_rec^l + P_reg^{l+1}, are functions of Q^l, and the statement that f_rec^l is 'implicitly tasked with reconstructing f_en^l - P_reg^{l+1}' is an arbitrary decomposition when both summands are learned from the same input. Nothing in the loss or architecture restricts P_reg^{l+1} to encode only normal information. The claimed disruption of Soft Identity Mapping is therefore not established by the presented equations; the ablation in Table 5 cannot distinguish the stated mechanism from a generic increase in model capacity or from an input-dependent residual.
  2. [Section 3.3 and Table 5] The paper's core premise—that increasing the number of prototypes causes Soft Identity Mapping and that the Prototype-based Constraint removes it—is supported only by a qualitative argument and a single ablation on MVTec-AD. No controlled experiment varies the number of prototypes N, and no direct measurement of reconstruction fidelity on anomalous regions with and without the constraint is reported. Without such evidence, the phenomenon and the proposed remedy are not demonstrated as claimed.
  3. [Section 4.1 and Tables 2-4] The SOTA claims are not supported with uncertainty estimates. The reported gains over INP-Former are small in several cases (e.g., image-level AUROC 99.8 vs. 99.7 on MVTec-AD and 91.5 vs. 90.5 on Real-IAD), and the per-dataset dropout rates (0.2, 0.3, 0.4) are given without a validation procedure. No code is released, so the results cannot be independently checked. Error bars or an explicit multi-seed evaluation protocol are needed to establish that the improvements are statistically meaningful.
minor comments (5)
  1. [Section 3.2, Eqs. (1)-(2)] The notation 'Dropout prob' is ambiguous; please specify whether 'prob' is the dropout rate or the keep probability.
  2. [Section 3.5, Eq. (11)] The loss formula uses T(·) while the text defines F(·) as the flatten operation, and the terms G_D^l(h,w) and g_d(·,·) are not defined in the text. Please make the notation consistent and define all quantities.
  3. [Table 1] The prototype parameter count (605,952) is not justified; please provide the formula used (e.g., N x C) and state the values of N and C.
  4. [Limitations and Future Works] The paragraph states the method is effective only for certain logical anomalies and acknowledges weaker performance on MVTec-LOCO; this admission tempers the broad robustness claim and should be reflected in the conclusion.
  5. [Figure 3] Figure 3 is not referenced in the text; please add an in-text citation where the anomaly localization visualizations are discussed.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: Pro-AD is a standard architecture paper evaluated on held-out benchmark splits, with no prediction that reduces to a fit or to a self-citation chain.

full rationale

The claimed derivation chain is not circular. Pro-AD proposes an expanded prototype set (P in R^{N x C}), a Dynamic Bidirectional Decoder, and a Prototype-based Constraint (Eq. 9). The 'Soft Identity Mapping' phenomenon is defined descriptively as the failure mode where anomalies are reconstructed through attention, and the Prototype-based Constraint is an architectural modification intended to limit prototype expressiveness; it is not a parameter fitted to the test labels or a quantity obtained by re-arranging the model's own loss. The empirical claims are tested on held-out test sets of MVTec-AD, VisA, and Real-IAD, and the component ablations in Table 5 are performed in the standard training/validation regime, so the reported AUROC/AP/F1/PRO numbers are not forced by construction. One may question whether the mechanism in Section 3.4 truly guarantees that the learned residual P_reg^{l+1} contains only normal information, since P^{l+1} is updated from the query Q^l via Eq. 4-5; however, that is a correctness or robustness concern about an architectural assumption, not a circularity in which an output is equivalent to an input by definition. Nor does the paper rely on load-bearing self-citations: the cited previous works (e.g., INP-Former, Dinomaly) are external baselines and building blocks, and none of the paper's central claims reduce to a uniqueness theorem or ansatz imported from the authors' own prior work. The Limitations and Future Works section honestly concedes weaker performance on MVTec-LOCO-style logical anomalies, which further indicates that the evaluation is not designed to force success. Under the stated criteria, the honest finding is no significant circularity.

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

The method's performance depends on a handful of hand-selected hyperparameters, most notably the prototype count (tied to patch count), decoder depth, per-dataset dropout rates, and the gradient decay parameter tau. The central assumptions are that pre-trained DINOv2 features faithfully represent normality, that linear attention is superior to softmax for prototype aggregation, and that the residual 'Prototype-based Constraint' actually constrains the model rather than merely adding capacity. No new physical or mathematical entities are introduced.

free parameters (5)
  • Number of prototypes N = Equal to patch count (784 for 392x392 input with patch size 14)
    Set equal to patch count to provide comprehensive coverage; no principled derivation.
  • Number of decoder layers = 8
    Chosen without ablation or justification; likely tuned on the benchmark.
  • Bottleneck dropout probability = 0.2 (MVTec-AD), 0.3 (VisA), 0.4 (Real-IAD)
    Set separately per dataset, suggesting test-set tuning.
  • Gradient decay hyperparameter tau = 3.0
    Reported in implementation details; no sensitivity analysis or derivation.
  • Training schedule settings = LR 1e-4, weight decay 1e-5, batch size 16, 200 epochs
    Standard choices; not central but part of the reproducibility recipe.
assumptions (4)
  • domain assumption Frozen DINOv2-R-Base/14 features provide a sufficient and uniform representation of normality for all classes
    The whole pipeline operates on these features; if they fail to separate anomalies, the method fails. No analysis is given for this choice beyond citing DINO.
  • domain assumption Linear cross-attention with smooth attention better aggregates normal semantics than softmax attention
    Motivated by avoiding sharp attention maps, but no comparison to softmax is provided.
  • ad hoc to paper The 'Soft Identity Mapping' effect grows with prototype count
    Introduced as motivation (Figure 1) but never formally defined or measured.
  • ad hoc to paper Adding the prototype-derived residual term prevents the reconstruction of anomalies
    This is the central mechanism of the Prototype-based Constraint; it is asserted, not derived.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Pro-AD: Learning Comprehensive Prototypes with Prototype-based Constraint for Multi-class Unsupervised Anomaly Detection." pith.science (2026). https://pith.science/paper/4QR44GTJ

@misc{pith2026250613097,
  author       = {Pith},
  title        = {Pith review of: Pro-AD: Learning Comprehensive Prototypes with Prototype-based Constraint for Multi-class Unsupervised Anomaly Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4QR44GTJ}},
  note         = {Machine review of arXiv:2506.13097}
}
read the original abstract

Prototype-based reconstruction methods for unsupervised anomaly detection utilize a limited set of learnable prototypes which only aggregates insufficient normal information, resulting in undesirable reconstruction. However, increasing the number of prototypes may lead to anomalies being well reconstructed through the attention mechanism, which we refer to as the "Soft Identity Mapping" problem. In this paper, we propose Pro-AD to address these issues and fully utilize the prototypes to boost the performance of anomaly detection. Specifically, we first introduce an expanded set of learnable prototypes to provide sufficient capacity for semantic information. Then we employ a Dynamic Bidirectional Decoder which integrates the process of the normal information aggregation and the target feature reconstruction via prototypes, with the aim of allowing the prototypes to aggregate more comprehensive normal semantic information from different levels of the image features and the target feature reconstruction to not only utilize its contextual information but also dynamically leverage the learned comprehensive prototypes. Additionally, to prevent the anomalies from being well reconstructed using sufficient semantic information through the attention mechanism, Pro-AD introduces a Prototype-based Constraint that applied within the target feature reconstruction process of the decoder, which further improves the performance of our approach. Extensive experiments on multiple challenging benchmarks demonstrate that our Pro-AD achieve state-of-the-art performance, highlighting its superior robustness and practical effectiveness for Multi-class Unsupervised Anomaly Detection task.

Figures

Figures reproduced from arXiv: 2506.13097 by the authors.

Figure 1
Figure 1. Illustration of the "Soft Identity Mapping" problem. (a) Few prototypes can only aggregation background related features with failed reconstruction and detection. (b) Expanded prototypes fail to detection certain anomalies as well and may cause "Soft Identity Mapping" problem. (c) Our Pro-AD learns a set of comprehensive prototypes and achieves both successful reconstruction and detection. deviations from normality … view at source ↗
Figure 2
Figure 2. Overview of our Pro-AD framwork for anomaly detection. (a) Our model consists of a pre-trained Encoder, an adaptive noisy bottleneck and a bidirectional Prototype-Guided decoder. (b) Detailed architecture of the bidirectional Prototype-Guided decoder layer. (c) The prototype’s ability to express normal and abnormal features. comprehensive prototypes and to dynamically utilize them for the target feature reconstructi… view at source ↗
Figure 3
Figure 3. Visualization results of anomaly localization on the MVTec-AD [3], VisA [28], and Real-IAD [23] datasets for multi-class anomaly detection. The first row presents the input images, while the second row displays the corresponding anomaly maps [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

28 extracted references · 21 canonical work pages

  1. [1]

    Probabilistic Modeling of Deep Features for Out-of-Distribution and Adversarial Detection

    Nilesh Ahuja et al. “Probabilistic Modeling of Deep Features for Out-of-Distribution and Adversarial Detection”. en-US. In:Cornell University - arXiv,Cornell University - arXiv(Sept. 2019)

    work page 2019

  2. [2]

    Beyond dents and scratches: Logical constraints in unsupervised anomaly detection and localization

    Paul Bergmann et al. “Beyond dents and scratches: Logical constraints in unsupervised anomaly detection and localization”. In:International Journal of Computer Vision130.4 (2022), pp. 947–969

    work page 2022

  3. [3]

    MVTec AD–A comprehensive real-world dataset for unsupervised anomaly detection

    Paul Bergmann et al. “MVTec AD–A comprehensive real-world dataset for unsupervised anomaly detection”. In:Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2019, pp. 9592–9600

    work page 2019

  4. [4]

    Vision transformers need registers

    Timothée Darcet et al. “Vision transformers need registers”. In:arXiv preprint arXiv:2309.16588(2023)

    arXiv 2023

  5. [5]

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

    Thomas Defard et al. “Padim: a patch distribution modeling framework for anomaly detection and localization”. In:International conference on pattern recognition. Springer. 2021, pp. 475– 489

    work page 2021

  6. [6]

    Anomaly detection via reverse distillation from one-class embedding

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

    work page 2022

  7. [7]

    Memorizing normality to detect anomaly: Memory-augmented deep autoen- coder for unsupervised anomaly detection

    Dong Gong et al. “Memorizing normality to detect anomaly: Memory-augmented deep autoen- coder for unsupervised anomaly detection”. In:Proceedings of the IEEE/CVF international conference on computer vision. 2019, pp. 1705–1714

    work page 2019

  8. [8]

    Mamba: Linear-time sequence modeling with selective state spaces

    Albert Gu and Tri Dao. “Mamba: Linear-time sequence modeling with selective state spaces”. In:arXiv preprint arXiv:2312.00752(2023)

    arXiv 2023

Show all 28 references
  1. [9]

    Dinomaly: The Less Is More Philosophy in Multi-Class Unsupervised Anomaly Detection

    Jia Guo et al. “Dinomaly: The Less Is More Philosophy in Multi-Class Unsupervised Anomaly Detection”. In:arXiv preprint arXiv:2405.14325(2024)

    2024 arXiv

  2. [10]

    A diffusion-based framework for multi-class anomaly detection

    Haoyang He et al. “A diffusion-based framework for multi-class anomaly detection”. In: Proceedings of the AAAI conference on artificial intelligence. V ol. 38. 8. 2024, pp. 8472–8480

    2024

  3. [11]

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

    Haoyang He et al. “Mambaad: Exploring state space models for multi-class unsupervised anomaly detection”. In:arXiv preprint arXiv:2404.06564(2024)

    2024 arXiv

  4. [12]

    Pixel-level anomaly detection via uncertainty-aware prototypical trans- former

    Chao Huang et al. “Pixel-level anomaly detection via uncertainty-aware prototypical trans- former”. In:Proceedings of the 30th ACM international conference on multimedia. 2022, pp. 521–530

    2022

  5. [13]

    Adaptive prototype learning and allocation for few-shot segmentation

    Gen Li et al. “Adaptive prototype learning and allocation for few-shot segmentation”. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2021, pp. 8334–8343. 10

    2021

  6. [14]

    Focal loss for dense object detection

    Tsung-Yi Lin et al. “Focal loss for dense object detection”. In:Proceedings of the IEEE international conference on computer vision. 2017, pp. 2980–2988

    2017

  7. [15]

    Simplenet: A simple network for image anomaly detection and localization

    Zhikang Liu et al. “Simplenet: A simple network for image anomaly detection and localization”. In:Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2023, pp. 20402–20411

    2023

  8. [16]

    Hierarchical vector quantized transformer for multi-class unsupervised anomaly detection

    Ruiying Lu et al. “Hierarchical vector quantized transformer for multi-class unsupervised anomaly detection”. In:Advances in Neural Information Processing Systems36 (2023), pp. 8487–8500

    2023

  9. [17]

    Exploring Intrinsic Normal Prototypes within a Single Image for Universal Anomaly Detection

    Wei Luo et al. “Exploring Intrinsic Normal Prototypes within a Single Image for Universal Anomaly Detection”. In:arXiv preprint arXiv:2503.02424(2025)

    2025 arXiv

  10. [18]

    Learning normal dynamics in videos with meta prototype network

    Hui Lv et al. “Learning normal dynamics in videos with meta prototype network”. In:Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2021, pp. 15425–15434

    2021

  11. [19]

    Learning memory-guided normality for anomaly detection

    Hyunjong Park, Jongyoun Noh, and Bumsub Ham. “Learning memory-guided normality for anomaly detection”. In:Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2020, pp. 14372–14381

    2020

  12. [20]

    Modeling the Distribution of Normal Data in Pre-Trained Deep Features for Anomaly Detection

    Oliver Rippel, Patrick Mertens, and Dorit Merhof. “Modeling the Distribution of Normal Data in Pre-Trained Deep Features for Anomaly Detection”. en-US. In:Cornell University - arXiv,Cornell University - arXiv(May 2020)

    2020

  13. [21]

    Towards total recall in industrial anomaly detection

    Karsten Roth et al. “Towards total recall in industrial anomaly detection”. In:Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2022, pp. 14318–14328

    2022

  14. [22]

    Prototypical networks for few-shot learning

    Jake Snell, Kevin Swersky, and Richard Zemel. “Prototypical networks for few-shot learning”. In:Advances in neural information processing systems30 (2017)

    2017

  15. [23]

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

    Chengjie Wang et al. “Real-iad: A real-world multi-view dataset for benchmarking versatile industrial anomaly detection”. In:Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2024, pp. 22883–22892

    2024

  16. [24]

    A unified model for multi-class anomaly detection

    Zhiyuan You et al. “A unified model for multi-class anomaly detection”. In:Advances in Neural Information Processing Systems35 (2022), pp. 4571–4584

    2022

  17. [25]

    Wide residual networks

    Sergey Zagoruyko and Nikos Komodakis. “Wide residual networks”. In:arXiv preprint arXiv:1605.07146(2016)

    2016 arXiv

  18. [26]

    Destseg: Segmentation guided denoising student-teacher for anomaly detection

    Xuan Zhang et al. “Destseg: Segmentation guided denoising student-teacher for anomaly detection”. In:Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2023, pp. 3914–3923

    2023

  19. [27]

    Omnial: A unified cnn framework for unsupervised anomaly localization

    Ying Zhao. “Omnial: A unified cnn framework for unsupervised anomaly localization”. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2023, pp. 3924–3933

    2023

  20. [28]

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

    Yang Zou et al. “Spot-the-difference self-supervised pre-training for anomaly detection and segmentation”. In:European Conference on Computer Vision. Springer. 2022, pp. 392–408. 11

    2022

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.