Pith. sign in

REVIEW 3 major objections 4 minor 37 references

"Stones from Other Hills can Polish Jade": Zero-shot Anomaly Image Synthesis via Cross-domain Anomaly Injection

T0 review · 3 major / 4 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read The paper claims that real anomalies borrowed from other industrial domains can be injected into normal target images, without training, to produce pseudo anomalies that train better defect detectors than existing zero-shot synthesis…

desk verdict Useful training-free anomaly synthesis with a real new dataset, but the headline numbers are reported under test-set oracle epoch selection and need a fixed-epoch rerun before they are citable. read the letter →

arxiv 2501.15211 v2 pith:53NB4UTI submitted 2025-01-25 cs.CV

Siqi Wang , Yuanze Hu , Xinwang Liu , Siwei Wang , Guangpu Wang , Chuanfu Xu , Jie Liu , Ping Chen This is my paper · ORCID
classification cs.CV
keywords industrialanomalydetectionzero-shotsynthesiscross-domaininjectionPoissonimageeditingdomain-agnosticdatasetdiffusion-basedpseudoimagespixel-levelevaluation
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

Industrial defect detectors normally need examples of defects, but real defects in one factory are scarce. This paper argues that defects are actually abundant once you stop restricting yourself to one domain: cracks, scratches, stains, holes and similar patterns appear everywhere. It proposes Cross-domain Anomaly Injection (CAI), which cuts out real anomaly patterns from other domains, resizes them, and seamlessly pastes them into normal images of the target domain, with no training at all. The paper also contributes a domain-agnostic anomaly dataset of about 6,000 real anomalies and a diffusion-based extension mechanism for unlimited synthesis. Across MvTec-AD, VisA and KSDD2, the resulting pseudo anomalies improve pixel-level detection accuracy and score better on an image-fidelity measure than those of existing zero-shot synthesis methods.

What carries the argument

The central object is CAI, a four-stage, parameter-free injection pipeline. Source-Target Matching uses the Source-Target Ratio (anomaly bounding-box size over foreground size) to resize each anomaly to a small, medium or large scale instead of stretching it; Multi-scale Anomaly Synthesis keeps a balanced 4:3:3 mix of scales; Injection Location Selection restricts centers to the foreground and fully inside the image; and Poisson Editing solves a Dirichlet-boundary gradient-matching problem so the inserted pattern blends with the surrounding image. CAI is supplied by DAAD, the paper's domain-agnostic anomaly dataset of eight anomaly classes, and optionally extended by a CAI-guided diffusion mechanism that trains on CAI's outputs to generate unlimited pseudo anomalies.

What would settle it

A decisive falsifier is to run CAI on a domain whose failures are structural rather than surface-level, such as a wrong or missing component, and check whether detector gains over a simple paste baseline disappear. A cheaper version is to compute FID between CAI's pseudo anomalies and the target test set with the anomaly pool restricted to the least-similar categories; if FID is no better than random pasting or detection degrades, the transferability premise is falsified.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that real anomalies are not actually scarce once the requirement that they belong to the target domain is dropped. Because anomaly classes such as cracks, scratches, pits and stains recur across domains, a pool of roughly 6,000 real cross-domain anomaly images can serve as raw material for zero-shot synthesis. CAI selects a real anomaly, resizes it to a compatible scale, chooses a foreground location in a normal target image, and injects it with Poisson editing so the boundary is gradient-continuous instead of visibly pasted. In the paper's head-to-head evaluation with a shared UNet detector, CAI-N reaches pixel-level AUC/AP/PRO of 95.6/62.6/87.5 on MvTec-AD, while the best zero-shot baseline reaches 92.4/49.1/80.6, and CAI also leads on VisA and KSDD2 and attains lower FID on all three datasets. The paper therefore concludes that authentic and training-free zero-shot anomaly synthesis are compatible, not a trade-off.

Load-bearing premise

The load-bearing premise is that anomaly patterns from other domains resemble the target domain's own defects closely enough that injecting the real patterns produces authentic pseudo anomalies; the paper cites this similarity but never measures it directly, so a target whose defect types are absent from the pool would break the method.

Editorial extensions

If this is right

  • IAD detectors can be trained on normal images plus injected cross-domain defects, closing much of the gap to few-shot anomaly synthesis methods that see real target defects.
  • Synthesis costs about 0.02 seconds per pseudo anomaly image, so large training sets can be generated on commodity hardware.
  • Because the same anomaly categories recur across domains, a growing shared anomaly dataset can keep improving many target domains simultaneously.
  • The diffusion supplement means the supply of realistic pseudo anomalies is not limited by the size of the anomaly dataset.
  • Detectors trained on CAI pseudo anomalies generalize to unseen anomaly types such as wrong pill type, bent grid, and transistor misplacement.

Reading between the lines

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

  • Editorial inference: although the paper evaluates CAI only through a simple UNet trained with focal loss, the pseudo anomalies are ordinary training images, so the benefit should transfer to other anomaly-detection backbones; this is a testable claim the authors do not make.
  • Editorial inference: the paper's CDM-5-over-CDM-8 result suggests that choosing which anomaly classes to inject by their relevance to the target domain could improve performance further; the paper reports the effect but does not propose a selection rule.
  • Editorial inference: if cross-domain transferability holds broadly, the practical recipe for a new production line is just to collect its normal images and borrow from a shared, growing pool of annotated defects; the paper does not simulate this operational setting.
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 / 4 minor

Summary. The paper proposes Cross-domain Anomaly Injection (CAI), a training-free zero-shot anomaly synthesis method that pastes real anomalies from other domains into normal target-domain images using source-target matching, multi-scale synthesis, injection-location selection, and Poisson editing. The authors also construct DAAD, a domain-agnostic anomaly dataset of about 6000 real anomaly images in eight classes, and propose a diffusion-based extension (CDM) trained on CAI outputs. Evaluation uses a UNet detector trained on normal images plus pseudo anomalies and compares pixel-level AUC, AP, and PRO on MvTec-AD, VisA, and KSDD2 against five ZSAS baselines and one FSAS reference. CAI-N/CAI-M report the best scores on all datasets, e.g., 95.6/62.6/87.5 on MvTec-AD versus 92.4/49.1/80.6 for NSA-N, and CAI also reports the lowest FID values.

Significance. If the empirical claims hold, the paper makes a useful conceptual contribution: abundant real anomalies from other domains are a practical, training-free source for zero-shot anomaly synthesis. The paper ships a new dataset, DAAD, that may benefit future IAD research, and the ablations isolate the role of PE, STM, and ILS in the method. The method itself is simple and reproducible in principle, and the reported per-image synthesis cost is low. The main weaknesses are in the evaluation protocol and in the unquantified assumptions about cross-domain anomaly transferability.

major comments (3)
  1. [§4.1, Table 1] The protocol states that 'we select the best-performing model among 200 training epochs for comparison to avoid setting optimal epoch number.' This means the reported numbers are test-set-selected maxima over training trajectories, and the paper reports a single random seed for all methods. This makes the headline margins uninterpretable: the 3.2-point AUC advantage on MvTec-AD and the 17.5-point AP advantage on KSDD2 could be produced by differences in convergence speed or training-curve variability rather than by synthesis quality. Please report results under a fixed training budget (or a validation-based checkpoint selection) with mean and standard deviation over at least three seeds, and re-derive the comparisons in Tables 1 and 2 under that protocol.
  2. [§3.1 (STM) and §3.2 (DAAD)] The method's load-bearing premise is that anomaly patterns from other domains are sufficiently similar to target-domain anomalies. The paper supports this by citing prior work and by end-to-end gains, but it provides no direct test of the premise. The evaluation datasets contain anomaly classes not explicitly present in DAAD (e.g., misplaced or bent objects), and the reported gains are averaged over all categories. Please add a per-anomaly-type performance breakdown and a source-class ablation (e.g., removing one DAAD class and measuring the effect on the corresponding target class), or a quantitative measure of source-target anomaly similarity, so that the scope of the transferability assumption is explicit.
  3. [§3.3, Eq. (5), Fig. 7a] CDM is introduced as a contribution and validated in Fig. 7a, but the main text says 'Due to page limit, see supplementary material for CDM's details.' Eq. (5) alone does not specify the architecture of the denoising network, the definition of the location mask y_loc, the conditioning mechanism e, the construction of the training set from CAI outputs, or the sampling settings. Please include these details in the main text or in an appendix that accompanies the paper, so that the 'unlimited anomaly synthesis' claim can be checked and reproduced.
minor comments (4)
  1. [Table 4] The 'Balanced (4:4:3)' configuration sums to 11 and conflicts with the text's '(4:3:3)' and with the setup N_l=4, N_m=3, N_s=3 in §3.1; please correct the table entry.
  2. [Fig. 7a] The claim that CDM-5 'even surpasses CAI in terms of AUC' would be easier to verify if the exact numeric values were printed in the figure or caption; the current axis labels make the plotted values unclear.
  3. [§4.2 and Fig. 6] The generalization-to-unseen-anomalies claim is illustrated only by qualitative heatmaps; please add quantitative results (e.g., per-unseen-type AP or AUC) or explicitly state that the figure is anecdotal.
  4. [Conclusion] The limitation that the paradigm only covers appearance anomalies should appear in the introduction or experimental setup, not only in the conclusion, because it delimits the scope of the claimed 'anomaly' synthesis.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: CAI is an empirical synthesis pipeline whose outputs are evaluated against external test sets, with evaluation datasets excluded from its source anomaly pool.

full rationale

The paper's central claim is that real anomalies from other domains can be injected into normal images to synthesize useful pseudo anomalies for industrial anomaly detection. This is an empirical construction-and-evaluation claim, not a derivation that reduces to its own inputs. The pseudo anomalies are produced from external cross-domain anomalies by resizing, location selection, and Poisson editing (Eq. 3), while the pseudo-anomaly masks are obtained by thresholding the difference between the synthesized image and the normal image (Eq. 4). That label-generation convention is a standard self-supervision device, not a prediction fitted to the evaluation target. The DAAD dataset is explicitly constructed so that evaluation datasets are excluded from the source pool when they are used for testing, preventing information leakage. The transferability of cross-domain anomaly patterns is stated as an assumption and then tested end-to-end through detector performance on held-out test sets, so the supporting evidence is external rather than definitional. No load-bearing self-citation appears in the derivation chain, and the paper's own stated limitation to appearance anomalies does not create circularity. The only notable concerns are experimental-protocol issues, such as oracle epoch selection over 200 test-evaluated checkpoints, but those affect the robustness of the empirical comparison and are not circularity in the derivation sense.

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

The central method is data-driven and uses no fitted predictive model, so the ledger is dominated by heuristic choices (scale thresholds, synthesis counts, binarization threshold) and domain assumptions about cross-domain transferability, mask quality, and the diffusion extension.

free parameters (4)
  • Multi-scale synthesis counts (Nl, Nm, Ns) = 4, 3, 3
    Chosen by hand as a balanced ratio for performance and computational overhead (Sec. 3.1, MAS); affects the scale distribution of pseudo anomalies.
  • Scale thresholds for STM = 0.1, 0.3, 0.7
    Used to classify anomaly patterns into trivial, small, medium, and large in Eq. 1 and Algorithm 1; chosen heuristically.
  • Mask binarization threshold T = not specified
    Used in Eq. 4 to convert the pixel-wise difference map into the anomaly mask y^(sa); the value is not reported.
  • Number of pseudo anomalies per target image = 10
    The evaluation protocol synthesizes ten pseudo anomaly images per normal image; the choice is a convention from the protocol and is not optimized per dataset.
assumptions (5)
  • domain assumption Real anomaly patterns are similar across industrial domains, so cross-domain anomalies can serve as authentic pseudo anomalies for a target domain.
    Stated in Sec. 3.1 observation (2), relying on references [29,36]; no direct validation is provided beyond the end-to-end IAD results.
  • standard math Poisson editing with a Dirichlet boundary condition produces seamless anomaly injections without artifact shortcuts.
    Poisson image editing is a well-established technique [20]; the paper uses it as a building block.
  • domain assumption The foreground mask from the rembg matting tool correctly identifies the object region in target images.
    Sec. 3.1 STM uses the matting output y^(n) to restrict injection locations; errors in the mask would misplace or truncate anomalies.
  • domain assumption Anomaly masks in DAAD are accurate enough for resizing and injection.
    Sec. 3.2 relies on EISeg annotations or existing masks; no annotation quality control is reported.
  • domain assumption The latent diffusion training objective in Eq. 5 is sufficient to learn diverse pseudo anomalies from CAI outputs.
    CDM is described as inspired by LDM and Textual Inversion, but its details are deferred to the supplementary material.

how reviews work

0 comments
Cite this review

Pith. "Pith review of "Stones from Other Hills can Polish Jade": Zero-shot Anomaly Image Synthesis via Cross-domain Anomaly Injection." pith.science (2026). https://pith.science/paper/53NB4UTI

@misc{pith2026250115211,
  author       = {Pith},
  title        = {Pith review of: "Stones from Other Hills can Polish Jade": Zero-shot Anomaly Image Synthesis via Cross-domain Anomaly Injection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/53NB4UTI}},
  note         = {Machine review of arXiv:2501.15211}
}
read the original abstract

Industrial image anomaly detection (IAD) is a pivotal topic with huge value. Due to anomaly's nature, real anomalies in a specific modern industrial domain (i.e. domain-specific anomalies) are usually too rare to collect, which severely hinders IAD. Thus, zero-shot anomaly synthesis (ZSAS), which synthesizes pseudo anomaly images without any domain-specific anomaly, emerges as a vital technique for IAD. However, existing solutions are either unable to synthesize authentic pseudo anomalies, or require cumbersome training. Thus, we focus on ZSAS and propose a brand-new paradigm that can realize both authentic and training-free ZSAS. It is based on a chronically-ignored fact: Although domain-specific anomalies are rare, real anomalies from other domains (i.e. cross-domain anomalies) are actually abundant and directly applicable to ZSAS. Specifically, our new ZSAS paradigm makes three-fold contributions: First, we propose a novel method named Cross-domain Anomaly Injection (CAI), which directly exploits cross-domain anomalies to enable highly authentic ZSAS in a training-free manner. Second, to supply CAI with sufficient cross-domain anomalies, we build the first Domain-agnostic Anomaly Dataset within our best knowledge, which provides ZSAS with abundant real anomaly patterns. Third, we propose a CAI-guided Diffusion Mechanism, which further breaks the quantity limit of real anomalies and enable unlimited anomaly synthesis. Our head-to-head comparison with existing ZSAS solutions justifies our paradigm's superior performance for IAD and demonstrates it as an effective and pragmatic ZSAS solution.

Figures

Figures reproduced from arXiv: 2501.15211 by the authors.

Figure 1
Figure 1. ZSAS flow of Cross-domain Anomaly Injection. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Comparison of pseudo anomaly images synthesized by [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Detailed procedure of CAI for ZSAS [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Procedure to build our domain-agnostic anomaly dataset. [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 6
Figure 6. Figure 6: IAD detector trained by pseudo anomalies synthesized [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Performance (%) comparison of different configurations. [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

37 extracted references · 30 canonical work pages

  1. [1]

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

    Paul Bergmann, Michael Freauser, David Sattlegger, and 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, 6

    work page 2019

  2. [2]

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

    Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Uninformed students: Student-teacher anomaly detection with discriminative latent embeddings. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 4183–4192, 2020. 2

    work page 2020

  3. [3]

    Mixed supervision for surface-defect detection: From weakly to fully supervised learning

    Jakob Bo ˇziˇc, Domen Tabernik, and Danijel Sko ˇcaj. Mixed supervision for surface-defect detection: From weakly to fully supervised learning. Computers in Industry , 129: 103459, 2021. 6

    work page 2021

  4. [4]

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

    Qiyu Chen, Huiyuan Luo, Chengkan Lv, and Zhengtao Zhang. A unified anomaly synthesis strategy with gradient ascent for industrial anomaly detection and localization. In European Conference on Computer Vision, 2024. 6, 8

    work page 2024

  5. [5]

    Fastrecon: Few-shot indus- trial anomaly detection via fast feature reconstruction

    Zheng Fang, Xiaoyang Wang, Haocheng Li, Jiejie Liu, Qi- ugui Hu, and Jimin Xiao. Fastrecon: Few-shot indus- trial anomaly detection via fast feature reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17481–17490, 2023. 2

    work page 2023

  6. [6]

    An image is worth one word: Personalizing text-to-image generation using textual inversion

    Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit Haim Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. In The Eleventh Interna- tional Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net, 2023. 6

    work page 2023

  7. [7]

    Rembg, 2022

    Daniel Gatis. Rembg, 2022. original-date: 2020-08- 10T14:38:24Z. 4

    work page 2022

  8. [8]

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

    Dong Gong, Lingqiao Liu, Vuong Le, Budhaditya Saha, Moussa Reda Mansour, Svetha Venkatesh, and Anton van den Hengel. Memorizing normality to detect anomaly: Memory-augmented deep autoencoder for unsupervised anomaly detection. In Proceedings of the IEEE/CVF inter- national conference on computer vision , pages 1705–1714,

Show all 37 references
  1. [9]

    Generative adversarial networks

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Commu- nications of the ACM, 63(11):139–144, 2020. 2

    2020

  2. [10]

    Eiseg: an efficient interactive segmentation tool based on paddlepaddle

    Yuying Hao, Yi Liu, Yizhou Chen, Lin Han, Juncai Peng, Shiyu Tang, Guowei Chen, Zewu Wu, Zeyu Chen, and Bao- hua Lai. Eiseg: an efficient interactive segmentation tool based on paddlepaddle. arXiv preprint arXiv:2210.08788 ,

    arXiv

  3. [11]

    Gans trained by a two time-scale update rule converge to a local nash equilib- rium

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilib- rium. Advances in neural information processing systems , 30, 2017. 6

    2017

  4. [12]

    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 Proceedings of the AAAI Conference on Artificial Intelligence, pages 8526–8534, 2024. 1, 2, 3, 6, 7, 8

    2024

  5. [13]

    Cagen: Controllable anomaly generator us- ing diffusion model

    Bolin Jiang, Yuqiu Xie, Jiawei Li, Naiqi Li, Yong Jiang, and Shu-Tao Xia. Cagen: Controllable anomaly generator us- ing diffusion model. In ICASSP 2024-2024 IEEE Interna- tional Conference on Acoustics, Speech and Signal Process- ing (ICASSP), pages 3110–3114. IEEE, 2024. 3

    2024

  6. [14]

    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. 2, 3, 5, 6

    2021

  7. [15]

    Eid-gan: Generative adversar- ial nets for extremely imbalanced data augmentation

    Wei Li, Jinlin Chen, Jiannong Cao, Chao Ma, Jia Wang, Xi- aohui Cui, and Ping Chen. Eid-gan: Generative adversar- ial nets for extremely imbalanced data augmentation. IEEE Transactions on Industrial Informatics , 19(3):3208–3218,

  8. [16]

    Deep indus- trial image anomaly detection: A survey

    Jiaqi Liu, Guoyang Xie, Jinbao Wang, Shangnian Li, Chengjie Wang, Feng Zheng, and Yaochu Jin. Deep indus- trial image anomaly detection: A survey. Machine Intelli- gence Research, 21(1):104–135, 2024. 1, 2

    2024

  9. [17]

    Dual-modeling decouple distillation for unsuper- vised anomaly detection

    Xinyue Liu, Jianyuan Wang, Biao Leng, and Shuo Zhang. Dual-modeling decouple distillation for unsuper- vised anomaly detection. arXiv preprint arXiv:2408.03888,

    arXiv

  10. [18]

    Simplenet: A simple network for image anomaly detection and localization

    Zhikang Liu, Yiming Zhou, Yuansheng Xu, and Zilei Wang. Simplenet: A simple network for image anomaly detection and localization. In Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition , pages 20402–20411, 2023. 2

    2023

  11. [19]

    Deep learning for anomaly detection: A review

    Guansong Pang, Chunhua Shen, Longbing Cao, and Anton Van Den Hengel. Deep learning for anomaly detection: A review. ACM computing surveys (CSUR), 54(2):1–38, 2021. 1

    2021

  12. [20]

    Pois- son image editing

    Patrick P ´erez, Michel Gangnet, and Andrew Blake. Pois- son image editing. In ACM SIGGRAPH 2003 Papers, page 313–318, New York, NY , USA, 2003. Association for Com- puting Machinery. 3, 5

    2003

  13. [21]

    Panda: Adapting pretrained features for anomaly detection and segmentation

    Tal Reiss, Niv Cohen, Liron Bergman, and Yedid Hoshen. Panda: Adapting pretrained features for anomaly detection and segmentation. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 2806–2814, 2021. 2

    2021

  14. [22]

    Gan- based defect synthesis for anomaly detection in fabrics

    Oliver Rippel, Maximilian M ¨uller, and Dorit Merhof. Gan- based defect synthesis for anomaly detection in fabrics. In 2020 25th IEEE International Conference on Emerging Technologies and Factory Automation (ETFA) , pages 534–

    2020

  15. [23]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022. 6

    2022

  16. [24]

    Focal loss for dense ob- ject detection

    T-YLPG Ross and GKHP Doll ´ar. Focal loss for dense ob- ject detection. In proceedings of the IEEE conference on computer vision and pattern recognition, pages 2980–2988,

  17. [25]

    Same same but differnet: Semi-supervised defect detection with normalizing flows

    Marco Rudolph, Bastian Wandt, and Bodo Rosenhahn. Same same but differnet: Semi-supervised defect detection with normalizing flows. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 1907– 1916, 2021. 2

    1907

  18. [26]

    Fully convolutional cross-scale-flows for image- based defect detection

    Marco Rudolph, Tom Wehrbein, Bodo Rosenhahn, and Bas- tian Wandt. Fully convolutional cross-scale-flows for image- based defect detection. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , pages 1088–1097, 2022. 2

    2022

  19. [27]

    Natural synthetic anomalies for self-supervised anomaly detection and localization

    Hannah M Schl ¨uter, Jeremy Tan, Benjamin Hou, and Bern- hard Kainz. Natural synthetic anomalies for self-supervised anomaly detection and localization. In European Conference on Computer Vision, pages 474–489. Springer, 2022. 1, 2, 3, 5, 6

    2022

  20. [28]

    Generative modeling by esti- mating gradients of the data distribution.Advances in Neural Information Processing Systems, 32, 2019

    Yang Song and Stefano Ermon. Generative modeling by esti- mating gradients of the data distribution.Advances in Neural Information Processing Systems, 32, 2019. 2

    2019

  21. [29]

    Defect transfer gan: Diverse defect synthesis for data augmentation

    Ruyu Wang, Sabrina Hoppe, Eduardo Monari, and Marco F Huber. Defect transfer gan: Diverse defect synthesis for data augmentation. arXiv preprint arXiv:2302.08366, 2023. 3

    2023 arXiv

  22. [30]

    Patch svdd: Patch-level svdd for anomaly detection and segmentation

    Jihun Yi and Sungroh Yoon. Patch svdd: Patch-level svdd for anomaly detection and segmentation. In Proceedings of the Asian conference on computer vision, 2020. 2

    2020

  23. [31]

    Draem- a discriminatively trained reconstruction embedding for sur- face anomaly detection

    Vitjan Zavrtanik, Matej Kristan, and Danijel Skoˇcaj. Draem- a discriminatively trained reconstruction embedding for sur- face anomaly detection. In Proceedings of the IEEE/CVF international conference on computer vision , pages 8330– 8339, 2021. 2, 3, 5, 6

    2021

  24. [32]

    Defect-gan: High-fidelity defect synthesis for automated defect inspection

    Gongjie Zhang, Kaiwen Cui, Tzu-Yi Hung, and Shijian Lu. Defect-gan: High-fidelity defect synthesis for automated defect inspection. In Proceedings of the IEEE/CVF Win- ter Conference on Applications of Computer Vision , pages 2524–2534, 2021. 3

    2021

  25. [33]

    Con- textual affinity distillation for image anomaly detection

    Jie Zhang, Masanori Suganuma, and Takayuki Okatani. Con- textual affinity distillation for image anomaly detection. In Proceedings of the IEEE/CVF Winter Conference on Appli- cations of Computer Vision, pages 149–158, 2024. 8

    2024

  26. [34]

    Destseg: Segmentation guided denoising student-teacher for anomaly detection

    Xuan Zhang, Shiyu Li, Xi Li, Ping Huang, Jiulong Shan, and Ting Chen. Destseg: Segmentation guided denoising student-teacher for anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3914–3923, 2023. 2

    2023

  27. [35]

    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 Con- ference on Computer Vision and Pattern Recognition, pages 16699–16708, 2024. 2, 3, 6

    2024

  28. [36]

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

    Qihang Zhou, Guansong Pang, Yu Tian, Shibo He, and Jim- ing Chen. Anomalyclip: Object-agnostic prompt learning for zero-shot anomaly detection. In The Twelfth International Conference on Learning Representations, 2024. 3

    2024

  29. [37]

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

    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. 3, 5, 6

    2022

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.