Pith. sign in

REVIEW 4 major objections 4 minor 34 references

Quality-Aware Language-Conditioned Local Auto-Regressive Anomaly Synthesis and Detection

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

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

Pith's one-line read 99.7 AUROC: token-anchored defect synthesis beats diffusion

desk verdict New AR-based anomaly synthesis, but the quality-aware weighting that underlies the SOTA claim has a false proof and a weak ablation. read the letter →

arxiv 2508.03539 v1 pith:TLHB7K2R submitted 2025-08-05 cs.CV

Long Qian , Bingke Zhu , Yingying Chen , Ming Tang , Jinqiao Wang This is my paper
classification cs.CV
keywords industrialanomalydetectionsynthesisauto-regressiveimagegenerationlanguage-conditionededitingCLIPimage-textsimilarityquality-awarereweightingtoken-anchoredlatentMVTecAD
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 claims that industrial anomaly detection can be improved by synthesizing training defects with a frozen autoregressive image model instead of diffusion inpainting, and by re-weighting each synthetic sample according to how well it matches its textual description. The ARAS editor encodes a normal image into discrete visual tokens, freezes every token outside a user-specified defect mask, and samples only the masked tokens under a language prompt, so the surrounding material texture is preserved exactly. The QARAD framework then scales each synthetic sample's training loss by a CLIP-based image-text similarity score, down-weighting generations that deviate from their prompt. On MVTec AD, VisA, and BTAD, the authors report state-of-the-art image- and pixel-level AUROC (99.7/99.8, 98.9/99.8, 96.7/98.0) while synthesizing each defect in 1.49 seconds, about five times faster than a diffusion baseline. If correct, this makes high-fidelity synthetic defects practical for industrial inspection training without iterative denoising.

What carries the argument

The load-bearing object is the hard-gated masked-AR sampling kernel, built on a frozen VQ-token autoregressive decoder (the 8-billion-parameter Infinity model). It uses a Dirac-style gate that forces unit probability on all context tokens outside the anomaly mask, so only mask-indexed tokens are re-sampled sequentially under the text prompt and previously sampled tokens. Decoding the edited token lattice yields the locally edited image. The companion mechanism is quality-aware weighting: each sample's weight is proportional to exp(gamma times the CLIP cosine similarity between the edited image and its prompt), and the normalized weights multiply the detector loss.

What would settle it

Train QARAD on MVTec AD with the per-sample CLIP weights replaced by a random permutation of the same weight values; if image or pixel AUROC stays near 99.7/99.9, the reported gains come from the weight distribution rather than from quality awareness. A direct check would measure the correlation between CLIP similarity and leave-one-out sample utility across the synthetic dataset.

Watch

Extended reading notes

Core claim

The central claim is that a frozen, token-level autoregressive decoder can serve as a training-free local anomaly editor: gating every context token to its original value and sampling only the masked tokens under a text prompt yields defects that preserve the surrounding material's high-frequency statistics and avoid the seam artifacts of low-resolution diffusion inpainting. The paper further claims that re-weighting each synthetic sample by the CLIP cosine similarity between the edited image and its textual prompt improves detector training, softly down-weighting prompt-inconsistent generations. Combined, the ARAS + QAW pipeline, called QARAD, is claimed to reach state-of-the-art image- and pixel-level AUROC on MVTec AD (99.7/99.8), VisA (98.9/99.8), and BTAD (96.7/98.0), while synthesizing each 1024-squared defect image in 1.49 seconds, about five times faster than the diffusion-based SDAS baseline.

Load-bearing premise

The pipeline assumes that a CLIP-style image-text similarity score orders synthetic defects by how useful they are for training the detector, so that down-weighting low-scoring samples improves learning.

Editorial extensions

If this is right

  • Replacing diffusion inpainting with ARAS removes the low-resolution denoising bottleneck, so fine material statistics propagate into the synthesized defect region.
  • Synthesis latency scales with the number of masked tokens rather than a fixed denoising schedule, giving the reported five-fold speedup at 1024 by 1024 resolution.
  • Quality-aware re-weighting down-weights prompt-inconsistent synthetic samples without discarding them, and the ablations show its gains add nearly additively to the synthesis gains.
  • Because the CLIP scorer is frozen and used only during training, the detector's inference latency and learnable parameter count are unchanged.

Reading between the lines

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

  • The same hard-gated editing recipe could be applied to any token-based generative prior, not just Infinity, wherever exact context preservation and local semantic edits are needed.
  • If CLIP is replaced by a stronger text-image verifier, QAW should transfer without retraining the detector, because the weight computation is decoupled from the detector loss.
  • On near-saturated benchmarks like MVTec AD, the absolute gains are small; the most informative comparisons are the pixel-level improvements on fine-grained texture classes, where ARAS's context preservation should matter most.
  • A direct check of QAW's premise would be to see whether CLIP similarity correlates with per-sample utility: if the correlation is weak, the weighting may be acting as a fixed regularizer rather than a quality filter.
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 / 4 minor

Summary. The paper introduces ARAS, a language-conditioned, masked auto-regressive anomaly synthesis method built on the frozen Infinity-8B AR model, which edits only the VQ tokens inside a user-specified anomaly mask while hard-gating all context tokens, and integrates it with QARAD, a detection framework that reweights each synthetic sample by a CLIP-based image-text similarity score. The authors report state-of-the-art image- and pixel-level AUROC on MVTec AD (99.7/99.8), VisA (98.9/99.8), and BTAD (96.7/98.0), together with a roughly 5x synthesis speedup over diffusion inpainting (1.49s vs 7.51s per image). The paper also releases code and a synthesized dataset.

Significance. If the reported results are correct, ARAS would be a practical contribution: it is training-free, preserves context tokens exactly, is language-controllable, and is substantially faster than diffusion-based synthesis. The empirical core is clearly described and the context-invariance property in Eq. (2) is sound. The public release of code and synthesized data is a valuable asset for reproducibility. However, the formal variance-reduction guarantee in Sec. 3.3 is mathematically incorrect, and the empirical support for the QAW contribution is incomplete because the key ablation in Tab. 7 lacks a control for non-semantic reweighting. These issues affect the central claim that the quality-aware weighting, rather than other factors, produces the reported gains.

major comments (4)
  1. [Sec. 3.3, Eq. (6)] The variance-reduction guarantee does not follow from the stated assumptions. The weights w_i = phi(s_i) / ((1/N) sum_j phi(s_j)) are self-normalized ratios, not importance-sampling weights for any proposal distribution over the synthetic data, so classical importance-sampling theory [15] cannot be invoked directly. Even under Cov(ell_i, phi(s_i)) > 0, Var(w_i ell_i) can exceed Var(ell_i); for example, if ell = s ~ Uniform[0,1] and phi(s) = s, then Var(w ell) = 16/45 > 1/12 = Var(ell). Additionally, Eq. (6) normalizes by the mean rather than the sum, so sum_i w_i = N rather than 1, and the claimed unbiasedness E[w_i] = 1 is not generally true for self-normalized ratios.
  2. [Appendix C, assumptions (A3) and C.4] The sign of the correlation between loss and similarity score is reversed relative to the main text. Section 3.3 assumes Cov(ell_i, f(s_i)) > 0 to conclude variance reduction, while Appendix C states in assumption (A3) that Cov(ell_i, s_i) <= 0 and then uses this to derive d(sigma^2)/d(gamma) = 2 Cov(ell, s) < 0 in C.4. These two claims cannot both be correct. Since the main text refers the reader to Appendix C for the complete theoretical analysis, this contradiction is load-bearing and must be resolved, not merely reworded.
  3. [Sec. 4.4, Tab. 7] The QAW ablation compares only uniform, hinge, and softmax weightings of the same synthetic samples. It does not include a control in which the same non-uniform weight distribution is applied with random or permuted assignment that is independent of CLIP similarity. Such a control is necessary to attribute the observed +0.1 image-level and +0.6 pixel-level AUROC gains to CLIP's semantic ordering rather than to any non-uniform emphasis changing training dynamics. This is particularly important because Tab. 6 shows that replacing RealNet's SDAS with ARAS leaves image-level AUROC unchanged at 99.6, so the image-level SOTA gain over RealNet reported in Sec. 4.2 is entirely attributable to QAW.
  4. [Appendix B, Tab. S1] The softmax temperature gamma = 2.5 is selected using test-set performance on MVTec AD, and the reported sensitivity table Tab. S1 compares only gamma values 1, 2.5, and 5 without a validation split or multiple seeds. Because gamma directly controls the sharpness of the reweighting distribution, test-set-based selection can introduce optimistic bias into the reported SOTA numbers. Please report the selection procedure on a held-out validation split and include variance across seeds.
minor comments (4)
  1. [Tab. 6] The table heading uses 'DREAM' while the method is DRAEM; the typo appears both in the table and in the accompanying text.
  2. [Sec. 3.3] The sentence 'The proposed weighting scheme isunbiased' has a missing space; more substantively, the unbiasedness claim should be restated or removed once the normalization issue in Eq. (6) is addressed.
  3. [Sec. 3.1, Eq. (1)] The notation 'A ~ Ds' followed by 'called A' is confusing because A first denotes the distribution and then a tuple; please define the sampling measure and the triplets (x, m, tau) with consistent notation throughout.
  4. [Appendix C.5] The 'Catoni-style PAC-Bayes' bound is stated with constants and the KL term suppressed; as written it is not verifiable, so either give the full statement with its conditions or label the passage explicitly as a heuristic.

Circularity Check

1 steps flagged · score 6.0 of 10

Reported MVTec image-level SOTA gain reduces to a test-set-tuned softmax temperature; ARAS's synthesis speed and other pixel-level gains remain independent.

  1. fitted input called prediction [Appendix B, Table S1; Section 4.4, Table 7]
    "Performance remains stable for γ ∈ [1, 5]; the default γ = 2.5 stays the best performance. γ Image-AUROC Pixel-AUROC 1 99.4 99.6 5 99.2 99.6 2.5 (default) 99.7 99.9"

    The central MVTec image-level gain of QARAD over RealNet is +0.1 (99.7 vs 99.6 in Tables 2 and 7), and the ablation in Table 7 attributes this entire image-level gain to softmax QAW. The softmax temperature is not a fixed prior choice: Appendix B shows γ=2.5 'stays the best performance' on the same MVTec AD test set, while γ=1 gives 99.4 and γ=5 gives 99.2 image-AUROC. The reported 99.7 softmax result is therefore the value of a hyperparameter selected by maximizing the same test metric that is later presented as evidence of improvement.

full rationale

The core ARAS contribution is not circular: the token-anchored masked sampling kernel is defined independently, context invariance follows from the hard gate by construction, and the 5x speed comparison against diffusion inpainting is an external efficiency measurement. The use of the authors' own MaPhC2F masks [17] is a self-citation, but it is a data input rather than a load-bearing argument or uniqueness claim, so it does not by itself make the derivation circular. The variance-reduction 'guarantee' in Section 3.3 and Appendix C is mathematically invalid: the normalized CLIP weights are not importance-sampling ratios, the main text assumes Cov(ℓ, f(s)) > 0 while Appendix C assumes Cov(ℓ, s) ≤ 0, and the claimed inequality does not follow from positive covariance. These are correctness flaws, not circular reductions, so they do not enter the score directly. However, the image-level SOTA improvement on MVTec AD is partial circularity: the ablation isolates the improvement to softmax QAW while the softmax temperature was tuned on the test set. Other results, including the pixel-level gains and the synthesis speedup, retain independent content, so the overall circularity is partial rather than total.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central claim rests on two external proxies (CLIP and the Qwen VLM), one self-supplied dataset (MaPhC2F masks), and a hyperparameter tuned on the test set. No new physical or mathematical entities are introduced beyond the formal notation of the masked sampling kernel. The variance-reduction 'guarantee' is an additional unsupported axiom the paper depends on.

free parameters (3)
  • softmax temperature γ = 2.5 (tuned on MVTec AD)
    Table S1 reports γ=2.5 as the best-performing value on MVTec test AUROC, so the final result depends on a hyperparameter chosen on the benchmark.
  • hinge threshold β = 0.20
    Listed as default in Table S3; the hinge variant is ablated but not used in the final QARAD model, suggesting it was chosen on the same benchmark.
  • anomaly masks per image = 6
    Section 4.1 samples 6 masks from MaPhC2F per normal image; the number is a hand-chosen design decision affecting the synthetic dataset size and diversity.
assumptions (5)
  • standard math The auto-regressive factorization in Eq. (3) with a fixed ordering π can be sampled left-to-right and the resulting distribution is a valid conditional generative model.
    Assumed by the Infinity-8B model; standard for AR transformers.
  • domain assumption The VQ-VAE encoder/decoder preserves fine material micro-structure (grain, weave, gloss) at inference resolution such that frozen context tokens yield seamless texture continuation.
    Section 3.2 claims micro-structural fidelity is inherited from frozen context tokens; this depends on the VQ-VAE's reconstruction fidelity and is not verified in the paper.
  • ad hoc to paper CLIP cosine similarity between the synthesized image and its text prompt is a reliable monotone proxy for the synthetic sample's utility in detector training.
    Section 3.3 and Fig. 1(b) assume this; no independent validation that CLIP resolves industrial defect semantics is provided, and the theoretical guarantee relies on an unproven covariance assumption.
  • domain assumption The text prompts produced by Qwen 2.5-VL are accurate, diverse, and describe plausible defects matching the mask.
    Section 4.1 feeds all prompts from Qwen; the paper shows only qualitative examples and does not audit failure cases.
  • ad hoc to paper The loss ℓ_i is correlated with the similarity score s_i, with sign differing between Sec. 3.3 (>0) and Appendix C (<0).
    The variance-reduction proof requires this covariance assumption, but the paper states opposite signs in the two sections and provides no empirical estimate of the correlation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Quality-Aware Language-Conditioned Local Auto-Regressive Anomaly Synthesis and Detection." pith.science (2026). https://pith.science/paper/TLHB7K2R

@misc{pith2026250803539,
  author       = {Pith},
  title        = {Pith review of: Quality-Aware Language-Conditioned Local Auto-Regressive Anomaly Synthesis and Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TLHB7K2R}},
  note         = {Machine review of arXiv:2508.03539}
}
read the original abstract

Despite substantial progress in anomaly synthesis methods, existing diffusion-based and coarse inpainting pipelines commonly suffer from structural deficiencies such as micro-structural discontinuities, limited semantic controllability, and inefficient generation. To overcome these limitations, we introduce ARAS, a language-conditioned, auto-regressive anomaly synthesis approach that precisely injects local, text-specified defects into normal images via token-anchored latent editing. Leveraging a hard-gated auto-regressive operator and a training-free, context-preserving masked sampling kernel, ARAS significantly enhances defect realism, preserves fine-grained material textures, and provides continuous semantic control over synthesized anomalies. Integrated within our Quality-Aware Re-weighted Anomaly Detection (QARAD) framework, we further propose a dynamic weighting strategy that emphasizes high-quality synthetic samples by computing an image-text similarity score with a dual-encoder model. Extensive experiments across three benchmark datasets-MVTec AD, VisA, and BTAD, demonstrate that our QARAD outperforms SOTA methods in both image- and pixel-level anomaly detection tasks, achieving improved accuracy, robustness, and a 5 times synthesis speedup compared to diffusion-based alternatives. Our complete code and synthesized dataset will be publicly available.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

34 extracted references · 32 canonical work pages

  1. [15]

    Art B. Owen. Monte Carlo theory, methods and examples

    work page

  2. [1]

    Samet Akcay, Amir Atapour-Abarghouei, and Toby P. Breckon. GANomaly: Semi-supervised anomaly detection via adversarial training. In Proc. of ACCV Workshops, 2018. 1

    work page 2018

  3. [2]

    Qwen2.5-vl technical report

    Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, Humen Zhong, Yuanzhi Zhu, Mingkun Yang, Zhao- hai Li, Jianqiang Wan, Pengfei Wang, Wei Ding, Zheren Fu, Yiheng Xu, Jiabo Ye, Xi Zhang, Tianbao Xie, Zesen Cheng, Hang Zhang, Zhibo Yang, Haiyang Xu, and Jun- yang Lin. Qwen2.5-vl technical repor...

    arXiv 2025

  4. [3]

    Mvtec ad–a comprehensive real-world dataset for unsupervised anomaly detection

    Paul Bergmann, Michael Fauser, David Sattlegger, and 8 Carsten Steger. Mvtec ad–a comprehensive real-world dataset for unsupervised anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9592–9600, 2019. 5

    work page 2019

  5. [4]

    AdaCLIP: Adapt- ing CLIP with Hybrid Learnable Prompts for Zero-Shot Anomaly Detection, page 55–72

    Yunkang Cao, Jiangning Zhang, Luca Frittoli, Yuqi Cheng, Weiming Shen, and Giacomo Boracchi. AdaCLIP: Adapt- ing CLIP with Hybrid Learnable Prompts for Zero-Shot Anomaly Detection, page 55–72. Springer Nature Switzer- land, 2024. 3

    work page 2024

  6. [5]

    Trans- fusion – a transparency-based diffusion model for anomaly detection

    Matic Fu ˇcka, Vitjan Zavrtanik, and Danijel Sko ˇcaj. Trans- fusion – a transparency-based diffusion model for anomaly detection. In Proc. of ECCV (35), pages 91–108, 2024. 3

    work page 2024

  7. [6]

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

    Zhaopeng Gu, Bingke Zhu, Guibo Zhu, Yingying Chen, Ming Tang, and Jinqiao Wang. Anomalygpt: Detecting industrial anomalies using large vision-language models,

    work page

  8. [7]

    Infinity: Scaling bit- wise autoregressive modeling for high-resolution image syn- thesis, 2024

    Jian Han, Jinlai Liu, Yi Jiang, Bin Yan, Yuqi Zhang, Zehuan Yuan, Bingyue Peng, and Xiaobing Liu. Infinity: Scaling bit- wise autoregressive modeling for high-resolution image syn- thesis, 2024. 2, 3, 6

    work page 2024

Show all 34 references
  1. [8]

    Diad: A diffusion-based framework for multi-class anomaly detection, 2023

    Haoyang He, Jiangning Zhang, Hongxu Chen, Xuhai Chen, Zhishan Li, Xu Chen, Yabiao Wang, Chengjie Wang, and Lei Xie. Diad: A diffusion-based framework for multi-class anomaly detection, 2023. 1

    2023

  2. [9]

    Anomalyd- iffusion: Few-shot anomaly image generation with diffusion model

    Teng Hu, Jiangning Zhang, Ran Yi, Yuzhen Du, Xu Chen, Liang Liu, Yabiao Wang, and Chengjie Wang. Anomalyd- iffusion: Few-shot anomaly image generation with diffusion model. In Proc. of AAAI, pages 8526–8534, 2024. 3

    2024

  3. [10]

    Cutpaste: Self-supervised learning for anomaly de- tection and localization

    Chun-Liang Li, Kihyuk Sohn, Jinsung Yoon, and Tomas Pfister. Cutpaste: Self-supervised learning for anomaly de- tection and localization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 9664–9674, 2021. 1, 3

    2021

  4. [11]

    Promptad: Learn- ing prompts with only normal samples for few-shot anomaly detection, 2024

    Xiaofan Li, Zhizhong Zhang, Xin Tan, Chengwei Chen, Yanyun Qu, Yuan Xie, and Lizhuang Ma. Promptad: Learn- ing prompts with only normal samples for few-shot anomaly detection, 2024. 1, 3

    2024

  5. [12]

    Simplenet: A simple network for image anomaly detection and localization, 2023

    Zhikang Liu, Yiming Zhou, Yuansheng Xu, and Zilei Wang. Simplenet: A simple network for image anomaly detection and localization, 2023. 1

    2023

  6. [13]

    Kevin Zhou

    Wenxin Ma, Xu Zhang, Qingsong Yao, Fenghe Tang, Chenxu Wu, Yingtai Li, Rui Yan, Zihang Jiang, and S. Kevin Zhou. Aa-clip: Enhancing zero-shot anomaly detection via anomaly-aware clip, 2025. 3

    2025

  7. [14]

    Vt-adl: A vision trans- former network for image anomaly detection and localiza- tion

    Pankaj Mishra, Riccardo Verk, Daniele Fornasier, Claudio Piciarelli, and Gian Luca Foresti. Vt-adl: A vision trans- former network for image anomaly detection and localiza- tion. In 2021 IEEE 30th International Symposium on Indus- trial Electronics (ISIE), pages 01–06, 2021. 5

    2021

  8. [16]

    Pramuditha Perera, Rajvindra Nath, Hemanth Venkateswara, Sethuraman Panchanathan, and Vishal M. Patel. OCGAN: One-class novelty detection using gans with constrained la- tent representations. In Proc. of CVPR , pages 2898–2906,

  9. [17]

    Mathphys-guided coarse-to-fine anomaly synthesis with sqe-driven bi-level optimization for anomaly detection, 2025

    Long Qian, Bingke Zhu, Yingying Chen, Ming Tang, and Jinqiao Wang. Mathphys-guided coarse-to-fine anomaly synthesis with sqe-driven bi-level optimization for anomaly detection, 2025. 6

    2025

  10. [18]

    Vcp-clip: A visual context prompting model for zero-shot anomaly segmenta- tion, 2024

    Zhen Qu, Xian Tao, Mukesh Prasad, Fei Shen, Zhengtao Zhang, Xinyi Gong, and Guiguang Ding. Vcp-clip: A visual context prompting model for zero-shot anomaly segmenta- tion, 2024. 3

    2024

  11. [19]

    Schl ¨uter, Jeremy Tan, Benjamin Hou, and Bern- hard Kainz

    Hannah M. Schl ¨uter, Jeremy Tan, Benjamin Hou, and Bern- hard Kainz. Natural synthetic anomalies for self-supervised anomaly detection and localization. In Proc. of ECCV, pages 474–489, 2022. 1

    2022

  12. [20]

    Unseen visual anomaly generation, 2025

    Han Sun, Yunkang Cao, Hao Dong, and Olga Fink. Unseen visual anomaly generation, 2025. 3

    2025

  13. [21]

    Visual autoregressive modeling: Scalable image generation via next-scale prediction, 2024

    Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Li- wei Wang. Visual autoregressive modeling: Scalable image generation via next-scale prediction, 2024. 3

    2024

  14. [22]

    Duong, Chanh D

    Tran Dinh Tien, Anh Tuan Nguyen, Nguyen Hoang Tran, Ta Duc Huy, Soan T.M. Duong, Chanh D. Tr. Nguyen, and Steven Q. H. Truong. Revisiting reverse distillation for anomaly detection. In Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR) ,...

    2023

  15. [23]

    Leach, Sebastian M

    Julian Wyatt, Adam D. Leach, Sebastian M. Schmon, and Chris G. Willcocks. AnoDDPM: Anomaly detection with de- noising diffusion probabilistic models using simplex noise. In Proc. of CVPR Workshops, pages 650–656, 2022. 1

    2022

  16. [24]

    Glad: Towards better reconstruction with global and local adaptive diffusion mod- els for unsupervised anomaly detection, 2024

    Hang Yao, Ming Liu, Haolin Wang, Zhicun Yin, Zifei Yan, Xiaopeng Hong, and Wangmeng Zuo. Glad: Towards better reconstruction with global and local adaptive diffusion mod- els for unsupervised anomaly detection, 2024. 3

    2024

  17. [25]

    Arik, and Tomas Pfister

    Jinsung Yoon, Kihyuk Sohn, Chun-Liang Li, Sercan O. Arik, and Tomas Pfister. Spade: Semi-supervised anomaly detec- tion under distribution mismatch, 2022. 3

    2022

  18. [26]

    DRÆM: A discriminatively trained reconstruction embed- ding for surface anomaly detection

    Vitjan Zavrtanik, Matej Kristan, and Danijel Sko ˇcaj. DRÆM: A discriminatively trained reconstruction embed- ding for surface anomaly detection. In Proc. of ICCV, pages 8330–8339, 2021. 1, 3

    2021

  19. [27]

    Dsr – a dual subspace re-projection network for surface anomaly detection

    Vitjan Zavrtanik, Matej Kristan, and Danijel Sko ˇcaj. Dsr – a dual subspace re-projection network for surface anomaly detection. In Computer Vision – ECCV 2022 , pages 539– 554, Cham, 2022. Springer Nature Switzerland. 1

    2022

  20. [28]

    Diffusionad: Norm-guided one-step denoising diffusion for anomaly detection

    Hui Zhang, Zheng Wang, Zuxuan Wu, and Yu-Gang Jiang. Diffusionad: Norm-guided one-step denoising diffusion for anomaly detection. arXiv preprint arXiv:2303.08730, 2023. 3

    2023 arXiv

  21. [29]

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

    Ximiao Zhang, Min Xu, and Xiuzhuang Zhou. Realnet: A feature selection network with realistic synthetic anomaly for anomaly detection. In Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR) ,

  22. [30]

    Anomalyclip: Object-agnostic prompt learning for zero-shot anomaly detection, 2025

    Qihang Zhou, Guansong Pang, Yu Tian, Shibo He, and Jim- ing Chen. Anomalyclip: Object-agnostic prompt learning for zero-shot anomaly detection, 2025. 3

    2025

  23. [31]

    bottle”, “capsule

    Yang Zou, Jongheon Jeong, Latha Pemula, Dongqing Zhang, and Onkar Dabeer. Spot-the-difference self-supervised pre- training for anomaly detection and segmentation. In Eu- ropean Conference on Computer Vision , pages 392–408. Springer, 2022. 5 9 Appendix A. Reproducibility Hype...

    2022

  24. [32]

    [First step in your logical reasoning process]

  25. [33]

    [Second step in your logical reasoning process]

  26. [34]

    from 1/4 to 3/4 horizontally, and from 1/2 to top edge vertically

    [Third step in your logical reasoning process] Description: [A concise and clear description of the single identified defect, specifying location, appearance, color, shape description, coordinate range, and approximate size relative to the object. less than 35 tokens. Must men...

Pith tools

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