Pith. sign in

REVIEW 2 major objections 1 minor 41 references

A diffusion model acting as low-pass filter creates an auxiliary low-frequency stream that anchors cross-modal alignment for 1-shot anomaly detection.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.3

2026-06-26 17:41 UTC pith:6Z4MLVLX

load-bearing objection The unverified claim that diffusion acts as a low-pass filter is the load-bearing assumption here and lacks any supporting analysis. the 2 major comments →

arxiv 2606.20300 v2 pith:6Z4MLVLX submitted 2026-06-18 cs.CV

CMDS-AD: Cross-Modal Dual-Stream Decoupling for Few-Shot Anomaly Detection

Junhao Cai , Deyu Zeng , Junhao Pang , Junyu Chen , Qiwei Liang , Xiaopin Zhong , Zongze Wu This is my paper
classification cs.CV
keywords few-shot anomaly detectioncross-modal dual-streamdiffusion model low-pass filtermulti-modal anomaly detectionfrequency decoupling1-shot settingMVTec 3D-AD
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper argues that existing multi-modal anomaly detection mixes stable low-frequency structures with high-frequency defect signals through uniform processing, which worsens misalignment when training data is scarce. It claims that routing RGB inputs through a pre-trained diffusion model produces a separate estimated stream consisting only of low-frequency normal representations, which then guides the real stream containing mixed frequencies to isolate defects more precisely. The Coordinate-Aware Hierarchical Feature Mapper and multiplicative scoring complete the separation and noise reduction. If correct, this dual-stream decoupling would deliver large gains in the 1-shot regime by turning geometric 3D cues into reliable structural templates without losing defect signals.

Core claim

CMDS-AD shows that a pre-trained diffusion model inherently functions as a non-linear low-pass filter, generating an auxiliary estimated stream of purely low-frequency normal representations from RGB inputs that anchors robust structural templates and assists the uncompressed real stream in isolating micro-defects through adaptive cross-modal alignment and multiplicative scoring.

What carries the argument

The pre-trained diffusion model used as a non-linear low-pass filter to extract low-frequency normal representations from RGB inputs, creating the auxiliary estimated stream that decouples frequencies for cross-modal alignment.

Load-bearing premise

The pre-trained diffusion model inherently acts as a non-linear low-pass filter that directly extracts purely low-frequency normal representations from RGB inputs without losing critical defect signals or introducing artifacts that affect downstream alignment.

What would settle it

Direct inspection of the diffusion model's output on defective RGB images that reveals retention of high-frequency defect signals rather than purely low-frequency normals would falsify the claimed filtering behavior.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • The auxiliary low-frequency stream reduces cross-modal misalignment and lowers false-positive rates by providing stable structural anchors.
  • The Coordinate-Aware Hierarchical Feature Mapper enables adaptive alignment of semantics between the estimated and real streams.
  • The multiplicative scoring mechanism filters modality-specific noise after alignment.
  • These mechanisms together produce absolute gains of several percent in I-AUROC and AUPRO under the 1-shot setting on MVTec 3D-AD and EyeCandies.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same low-pass anchoring idea could be tested on other paired modalities such as RGB-infrared where frequency content also differs.
  • If the diffusion filter proves reliable, replacing it with lighter non-diffusion low-pass operators might preserve the gains at lower compute cost.
  • Extending the dual-stream logic to temporal sequences would require checking whether the same frequency separation holds across video frames.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 1 minor

Summary. The paper introduces CMDS-AD, a Cross-Modal Dual-Stream Anomaly Detection framework for few-shot multi-modal anomaly detection. It uses a LoRA-guided diffusion model to augment scarce RGB data and a pre-trained diffusion model as a normal estimator claimed to inherently function as a non-linear low-pass filter extracting purely low-frequency normal representations from RGB inputs. This creates an 'estimated stream' for stable structural templates and a 'real stream' for isolating micro-defects, aided by a Coordinate-Aware Hierarchical Feature Mapper for cross-modal alignment and multiplicative scoring to filter noise. The method reports new SOTA results under 1-shot settings: absolute gains of 5.7% I-AUROC and 2.0% AUPRO on MVTec 3D-AD, and 7.7% and 5.6% on EyeCandies.

Significance. If the diffusion-based decoupling mechanism holds, the approach could meaningfully advance few-shot MAD by addressing frequency conflation and cross-modal misalignment without heavy parameter fitting. The code release supports reproducibility. However, the significance is limited by the absence of verification for the core low-pass filter assumption, which underpins attribution of the reported gains to the dual-stream design rather than other factors.

major comments (2)
  1. [Abstract/Method] Abstract and method description: the claim that the pre-trained diffusion model 'inherently acts as a non-linear low-pass filter, directly extracting low-frequency normal representations from RGB inputs' without losing defect signals or introducing artifacts is load-bearing for the dual-stream decoupling and cross-modal alignment. No frequency-domain analysis, ablation removing the estimator, or pre/post-estimation defect visibility comparisons are supplied to substantiate points (a)–(c) in the stress-test note.
  2. [Experiments] Experiments section: performance gains (5.7% I-AUROC, 2.0% AUPRO on MVTec 3D-AD) are reported as absolute improvements establishing SOTA, yet no error bars, multiple random seeds, or statistical tests are mentioned, weakening confidence that the gains arise specifically from the claimed frequency separation rather than post-hoc choices or variance.
minor comments (1)
  1. [Abstract] The abstract notes code availability at the GitHub link; confirm that the released repository includes scripts and configurations to exactly reproduce the 1-shot MVTec 3D-AD and EyeCandies results.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. We address the major comments point by point below and will incorporate revisions to strengthen the empirical support for our claims and the statistical reporting of results.

read point-by-point responses

  1. Referee: [Abstract/Method] Abstract and method description: the claim that the pre-trained diffusion model 'inherently acts as a non-linear low-pass filter, directly extracting low-frequency normal representations from RGB inputs' without losing defect signals or introducing artifacts is load-bearing for the dual-stream decoupling and cross-modal alignment. No frequency-domain analysis, ablation removing the estimator, or pre/post-estimation defect visibility comparisons are supplied to substantiate points (a)–(c) in the stress-test note.

    Authors: We acknowledge that the low-pass filter property is central to the dual-stream design and that the current manuscript does not include direct empirical verification such as frequency-domain analysis or targeted ablations. In the revised version, we will add Fourier spectrum comparisons between the estimated and real streams, an ablation that removes the diffusion estimator, and qualitative pre/post-estimation visualizations of defect signals to substantiate the decoupling behavior. revision: yes

  2. Referee: [Experiments] Experiments section: performance gains (5.7% I-AUROC, 2.0% AUPRO on MVTec 3D-AD) are reported as absolute improvements establishing SOTA, yet no error bars, multiple random seeds, or statistical tests are mentioned, weakening confidence that the gains arise specifically from the claimed frequency separation rather than post-hoc choices or variance.

    Authors: We agree that the absence of variability measures and statistical tests limits confidence in attributing the gains specifically to the proposed mechanism. We will rerun the 1-shot experiments over multiple random seeds, report means with standard deviations, and include paired statistical significance tests in the updated experiments section. revision: yes

Circularity Check

0 steps flagged

No significant circularity; method relies on external pre-trained models with independent empirical claims

full rationale

The provided abstract and description contain no equations, fitted parameters, or self-citations that reduce the reported performance gains or the dual-stream decoupling to quantities defined by construction from the same inputs. The core claim invokes a pre-trained diffusion model as an external normal estimator whose low-pass property is asserted but not derived internally; this is an assumption rather than a self-referential reduction. No load-bearing step equates a prediction to a fit on the target data, renames a known result, or chains to prior author work as the sole justification. The 1-shot results on MVTec 3D-AD and EyeCandies are presented as empirical outcomes on external benchmarks, leaving the derivation self-contained against those benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the domain assumption that a pre-trained diffusion model functions as an effective non-linear low-pass filter for normal structures; no explicit free parameters or invented entities are named in the abstract.

axioms (1)
  • domain assumption Pre-trained diffusion model acts as non-linear low-pass filter extracting purely low-frequency normal representations from RGB inputs
    Invoked to justify creation of the auxiliary estimated stream that anchors structural templates.

pith-pipeline@v0.9.1-grok · 5838 in / 1304 out tokens · 48470 ms · 2026-06-26T17:41:38.718577+00:00 · methodology

0 comments
read the original abstract

Few-shot anomaly detection remains challenging due to limited training data. Multi-modal anomaly detection (MAD) offers a viable solution, leveraging 3D geometric cues to enrich 2D RGB representations and compensate for this scarcity. However, existing MAD methods apply spatially uniform feature processing, conflating stable macroscopic structures with high-frequency localized defect signals, exacerbating cross-modal misalignment and inflating false-positive rates. To overcome this, we present CMDS-AD, a Cross-Modal Dual-Stream Anomaly Detection framework. A LoRA-guided diffusion model generates diverse RGB samples to mitigate extreme data scarcity. For 3D normal augmentation, we employ a pre-trained diffusion model as a normal estimator. Crucially, this estimator inherently acts as a non-linear low-pass filter, directly extracting low-frequency normal representations from RGB inputs. This establishes an auxiliary estimated stream of purely low-frequency information, anchoring robust structural templates and assisting the uncompressed real stream, containing coupled high- and low-frequency components, to precisely isolate micro-defects. A Coordinate-Aware Hierarchical Feature Mapper adaptively aligns cross-modal semantics, while a multiplicative scoring mechanism filters modality-specific noise. Under the extreme 1-shot setting, CMDS-AD achieves absolute performance gains of 5.7% (I-AUROC) and 2.0% (AUPRO) on MVTec 3D-AD, alongside 7.7% and 5.6% improvements on EyeCandies, establishing a new state-of-the-art. Code is available at https://github.com/Junhaocai27/CMDS-AD

Figures

Figures reproduced from arXiv: 2606.20300 by Deyu Zeng, Junhao Cai, Junhao Pang, Junyu Chen, Qiwei Liang, Xiaopin Zhong, Zongze Wu.

Figure 1
Figure 1. Figure 1: Few-shot 3D anomaly detection overview. (a) Illustration showing how esti￾mated streams enhance defect representation in real streams, and multi-modal 2D-3D multiplication further localizes defects. (b) Radar plots on MVTec 3D-AD and Eye￾Candies report few-shot (1, 2, 4-shot) I-AUROC and AUPRO. Our method (red) out￾performs baselines (blue) with annotated gains, demonstrating robust improvements. where man… view at source ↗
Figure 2
Figure 2. Figure 2: Proposed Cross-Modal Dual-Stream Anomaly Detection (CMDS-AD) frame￾work. (Top) Diffusion-driven augmentation synthesizes paired RGB-normal training data. (Bottom Left) Dual-stream pipeline processing Real and Estimated streams, where the purely low-frequency estimated stream acts as an auxiliary guide to enhance the coupled real stream. (Bottom Right) Coordinate-Aware Hierarchical Feature Mapper for adapti… view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative evaluation of the complementarity between 2D and 3D modali￾ties. We visualize the predicted anomaly maps across three distinct defect categories: 2D color-only anomalies, 3D shape-only anomalies, and combined anomalies. Syner￾gistically fusing these modality-specific predictions effectively captures all defect types while suppressing background noise. localization under severe data scarcity. Ov… view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparison of anomaly score overlays between our proposed method and state-of-the-art baselines on the MVTec 3D-AD and EyeCandies datasets. Under challenging few-shot constraints, our approach effectively suppresses false positives in normal regions and delineates subtle defect boundaries with significantly higher preci￾sion than existing methods [PITH_FULL_IMAGE:figures/full_fig_p013_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Parameter sensitivity analysis regarding the weights of the estimated stream (λ1, λ2) on MVTec 3D-AD and EyeCandies datasets. The plots illustrate the impact on I-AUROC and AUPRO@30% metrics as the weights vary from 0.0 to 1.0 with a step size of 0.1. From left to right: (a) I-AUROC on MVTec 3D-AD, (b) I-AUROC on EyeCandies, (c) AUPRO@30% on MVTec 3D-AD, and (d) AUPRO@30% on EyeCan￾dies [PITH_FULL_IMAGE:f… view at source ↗
Figure 6
Figure 6. Figure 6: PRO curve dynamics across different integration limits (FPR). We select four representative classes (Dowel, Rope, Bagel, Carrot) to illustrate the varying degradation behaviors of different fusion strategies. The inset panels provide a 10× magnified view specifically anchored at the strict 1% FPR threshold. Multiplicative fusion (Ψ2D ⊙Ψ3D) consistently maintains top-tier stability, effectively resisting th… view at source ↗
Figure 7
Figure 7. Figure 7: FFT-based frequency analysis of real and estimated normals. Estimated nor￾mals suppress high-frequency responses while preserving dominant structural compo￾nents [PITH_FULL_IMAGE:figures/full_fig_p022_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Typical failure cases of CMDS-AD. We highlight instances where the model misses specific structural defects or predicts broader false positive regions due to hyper￾complex high-frequency patterns. Est SN MAFR Ours GT Real SN RGB Carrot Bagel Potato Cookie Foam Cable_gland Dowel Peach Rope Tire [PITH_FULL_IMAGE:figures/full_fig_p025_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Extended qualitative results on the MVTec 3D-AD dataset (Part 1) under the 4-shot training setting. From top to bottom: RGB image, Real Normal map, Surface Normal map, Ground Truth mask, MAFR prediction, and our CMDS-AD predicted anomaly map [PITH_FULL_IMAGE:figures/full_fig_p025_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Extended qualitative results on the MVTec 3D-AD dataset (Part 2) under the 4-shot training setting. The layout of the visualizations is identical to [PITH_FULL_IMAGE:figures/full_fig_p026_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Extended qualitative results on the EyeCandies dataset (Part 1) under the 4-shot training setting. From top to bottom: RGB image, Real Normal map, Surface Normal map, Ground Truth mask, MAFR prediction, and our CMDS-AD predicted anomaly map [PITH_FULL_IMAGE:figures/full_fig_p026_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Extended qualitative results on the EyeCandies dataset (Part 2) under the 4-shot training setting. The layout of the visualizations is identical to [PITH_FULL_IMAGE:figures/full_fig_p027_12.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

41 extracted references · 2 canonical work pages

  1. [1]

    In: 2025 International Conference on Digital Image Computing: Techniques and Applications (DICTA)

    Ali, U., Zia, A., Rehman, A., Ramzan, U., Hassan, Z., Sattar, T., Wang, J., Xiang, W.: 2d-3d feature fusion via cross-modal latent synthesis and attention-guided restoration for industrial anomaly detection. In: 2025 International Conference on Digital Image Computing: Techniques and Applications (DICTA). pp. 1–8. IEEE (2025)

    2025

  2. [2]

    arXiv preprint arXiv:2112.09045 , year=

    Bergmann, P., Jin, X., Sattlegger, D., Steger, C.: The mvtec 3d-ad dataset for un- supervised 3d anomaly detection and localization. arXiv preprint arXiv:2112.09045 (2021)

    work page arXiv 2021

  3. [3]

    In: Proceedings of the Asian Conference on Computer Vision

    Bonfiglioli, L., Toschi, M., Silvestri, D., Fioraio, N., De Gregorio, D.: The eyecan- dies dataset for unsupervised multimodal anomaly detection and localization. In: Proceedings of the Asian Conference on Computer Vision. pp. 3586–3602 (2022)

    2022

  4. [4]

    In: Proceedings of the 31st ACM International Conference on Multimedia

    Chen, R., Xie, G., Liu, J., Wang, J., Luo, Z., Wang, J., Zheng, F.: Easynet: An easy network for 3d industrial anomaly detection. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 7038–7046 (2023)

    2023

  5. [5]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Choi, J., Lee, J., Shin, C., Kim, S., Kim, H., Yoon, S.: Perception prioritized train- ing of diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11472–11481 (2022)

    2022

  6. [6]

    Chu, Y.M., Chieh, L., Hsieh, T.I., Chen, H.T., Liu, T.L.: Shape-guided dual- memory learning for 3d anomaly detection (2023)

    2023

  7. [7]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Costanzino, A., Ramirez, P.Z., Lisanti, G., Di Stefano, L.: Multimodal indus- trial anomaly detection by crossmodal feature mapping. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17234– 17243 (2024)

    2024

  8. [8]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Deng, H., Li, X.: Anomaly detection via reverse distillation from one-class embed- ding. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9737–9746 (2022)

    2022

  9. [9]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Fang, Z., Wang, X., Li, H., Liu, J., Hu, Q., Xiao, J.: Fastrecon: Few-shot in- dustrial anomaly detection via fast feature reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17481–17490 (2023)

    2023

  10. [10]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Gu, Z., Zhang, J., Liu, L., Chen, X., Peng, J., Gan, Z., Jiang, G., Shu, A., Wang, Y., Ma, L.: Rethinking reverse distillation for multi-modal anomaly detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 8445– 8453 (2024)

    2024

  11. [11]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Horwitz, E., Hoshen, Y.: Back to the feature: classical 3d features are (almost) all you need for 3d anomaly detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2968–2977 (2023) 16 J. Cai et al

    2023

  12. [12]

    Iclr1(2), 3 (2022)

    Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W., et al.: Lora: Low-rank adaptation of large language models. Iclr1(2), 3 (2022)

    2022

  13. [13]

    In: European conference on computer vision

    Huang, C., Guan, H., Jiang, A., Zhang, Y., Spratling, M., Wang, Y.F.: Registration based few-shot anomaly detection. In: European conference on computer vision. pp. 303–319. Springer (2022)

    2022

  14. [14]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Jeong, J., Zou, Y., Kim, T., Zhang, D., Ravichandran, A., Dabeer, O.: Winclip: Zero-/few-shot anomaly classification and segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 19606– 19616 (2023)

    2023

  15. [15]

    IEEE Transactions on Pattern Analysis and Machine Intelligence (2025)

    Ke, B., Qu, K., Wang, T., Metzger, N., Huang, S., Li, B., Obukhov, A., Schindler, K.: Marigold: Affordable adaptation of diffusion-based image generators for image analysis. IEEE Transactions on Pattern Analysis and Machine Intelligence (2025)

    2025

  16. [16]

    arXiv e-prints pp

    Lee, Y., Jang, S., Yoon, H.: Anople: Few-shot anomaly detection via bi-directional prompt learning with only normal samples. arXiv e-prints pp. arXiv–2408 (2024)

    2024

  17. [17]

    In: Proceedings of the IEEE/CVF conference on com- puter vision and pattern recognition

    Li, W., Xu, X., Gu, Y., Zheng, B., Gao, S., Wu, Y.: Towards scalable 3d anomaly detection and localization: A benchmark via 3d anomaly synthesis and a self- supervised learning network. In: Proceedings of the IEEE/CVF conference on com- puter vision and pattern recognition. pp. 22207–22216 (2024)

    2024

  18. [18]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Li, Y., Liu, F., Liao, J., Tian, S., Foo, C.S., Yang, X.: Find: Few-shot anomaly inspection with normal-only multi-modal data. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 23290–23299 (2025)

    2025

  19. [19]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Liang, H., Xie, G., Hou, C., Wang, B., Gao, C., Wang, J.: Look inside for more: Internal spatial modality perception for 3d anomaly detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 39, pp. 5146–5154 (2025)

    2025

  20. [20]

    In: Proceedings of the AAAI Conference on Ar- tificial Intelligence

    Lin, Y., Yan, H., Tong, X., Chang, Y., Wang, H., Zhou, Z., Gao, S., Wang, Y., Zhang, W.: Commonality in few: Few-shot multimodal anomaly detection via hypergraph-enhanced memory. In: Proceedings of the AAAI Conference on Ar- tificial Intelligence. vol. 40, pp. 7015–7023 (2026)

    2026

  21. [21]

    In: European Conference on Computer Vision

    Liu, C., Chu, Y.M., Hsieh, T.I., Chen, H.T., Liu, T.L.: Learning diffusion models for multi-view anomaly detection. In: European Conference on Computer Vision. pp. 328–345. Springer (2024)

    2024

  22. [22]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Long, K., Xie, G., Ma, L., Liu, J., Lu, Z.: Revisiting multimodal fusion for 3d anomaly detection from an architectural perspective. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 39, pp. 12273–12281 (2025)

    2025

  23. [23]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Luo, X., Xie, Y., Qu, Y., Fu, Y.: Skipdiff: Adaptive skip diffusion model for high- fidelity perceptual image super-resolution. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 4017–4025 (2024)

    2024

  24. [24]

    Knowledge-Based Systems290, 111563 (2024)

    Lyu, S., Mo, D., keung Wong, W.: Reb: Reducing biases in representation for industrial anomaly detection. Knowledge-Based Systems290, 111563 (2024)

    2024

  25. [25]

    In: Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition

    Ristea, N.C., Madan, N., Ionescu, R.T., Nasrollahi, K., Khan, F.S., Moes- lund, T.B., Shah, M.: Self-supervised predictive convolutional attentive block for anomaly detection. In: Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition. pp. 13576–13586 (2022)

    2022

  26. [26]

    In: Proceedings of the IEEE/CVF winter conference on applications of computer vision

    Rudolph, M., Wehrbein, T., Rosenhahn, B., Wandt, B.: Asymmetric student- teacher networks for industrial anomaly detection. In: Proceedings of the IEEE/CVF winter conference on applications of computer vision. pp. 2592–2602 (2023)

    2023

  27. [27]

    In: Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Si, C., Huang, Z., Jiang, Y., Liu, Z.: Freeu: Free lunch in diffusion u-net. In: Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4733–4743 (2024) CMDS-AD for Few-Shot Anomaly Detection 17

    2024

  28. [28]

    Information Fusion p

    Sui, W., Lichau, D., Lefèvre, J., Phelippeau, H.: Incomplete multimodal indus- trial anomaly detection via cross-modal distillation. Information Fusion p. 103572 (2025)

    2025

  29. [29]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Tang, N., Luo, X., Cheng, Z., Zhou, L., Zhang, D., Qu, Y.: Diffusion once and done: Degradation-aware lora for all-in-one image restoration. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 40, pp. 9448–9456 (2026)

    2026

  30. [30]

    In: Proceedings of the 32nd ACM International Conference on Multi- media

    Tian, L., Zhao, H., Lu, R., Wang, R., Wu, Y., Wang, L., He, X., Liu, X.: Foct: Few-shot industrial anomaly detection with foreground-aware online conditional transport. In: Proceedings of the 32nd ACM International Conference on Multi- media. pp. 6241–6249 (2024)

    2024

  31. [31]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Tien, T.D., Nguyen, A.T., Tran, N.H., Huy, T.D., Duong, S., Nguyen, C.D.T., Truong, S.Q.: Revisiting reverse distillation for anomaly detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 24511–24520 (2023)

    2023

  32. [32]

    In: International Conference on Learning Represen- tations

    Trabucco, B., Doherty, K., Gurinas, M., Salakhutdinov, R.: Effective data augmen- tation with diffusion models. In: International Conference on Learning Represen- tations. vol. 2024, pp. 14590–14612 (2024)

    2024

  33. [33]

    In: European conference on computer vision

    Tu, Y., Zhang, B., Liu, L., Li, Y., Zhang, J., Wang, Y., Wang, C., Zhao, C.: Self- supervised feature adaptation for 3d industrial anomaly detection. In: European conference on computer vision. pp. 75–91. Springer (2024)

    2024

  34. [34]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Wang, Y., Peng, J., Zhang, J., Yi, R., Wang, Y., Wang, C.: Multimodal industrial anomaly detection via hybrid fusion. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8032–8041 (2023)

    2023

  35. [35]

    In: Proceed- ings of the IEEE/CVF conference on computer vision and pattern recognition

    Wyatt, J., Leach, A., Schmon, S.M., Willcocks, C.G.: Anoddpm: Anomaly detec- tion with denoising diffusion probabilistic models using simplex noise. In: Proceed- ings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 650–656 (2022)

    2022

  36. [36]

    Advances in Neural Information Processing Systems35, 32398–32411 (2022)

    Yan, X., Zhan, H., Zheng, C., Gao, J., Zhang, R., Cui, S., Li, Z.: Let images give you more: Point cloud cross-modal training for shape analysis. Advances in Neural Information Processing Systems35, 32398–32411 (2022)

    2022

  37. [37]

    In: Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition

    Yang, X., Zhou, D., Feng, J., Wang, X.: Diffusion probabilistic model made slim. In: Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition. pp. 22552–22562 (2023)

    2023

  38. [38]

    In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision

    Zavrtanik, V., Kristan, M., Skočaj, D.: Cheating depth: Enhancing 3d surface anomaly detection via depth simulation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 2164–2172 (2024)

    2024

  39. [39]

    IEEE transactions on pattern analysis and machine intelligence (2025)

    Zhang, H., Wang, Z., Zeng, D., Wu, Z., Jiang, Y.G.: Diffusionad: Norm-guided one-step denoising diffusion for anomaly detection. IEEE transactions on pattern analysis and machine intelligence (2025)

    2025

  40. [40]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Zhang, X., Li, S., Li, X., Huang, P., Shan, J., Chen, T.: Destseg: Segmentation guided denoising student-teacher for anomaly detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3914–3923 (2023)

    2023

  41. [41]

    Zhou, Z., Wang, L., Fang, N., Wang, Z., Qiu, L., Zhang, S.: R3d-ad: Reconstruction viadiffusionfor3danomalydetection.In:Europeanconferenceoncomputervision. pp. 91–107. Springer (2024) 18 J. Cai et al. Supplementary Material for CMDS-AD: Cross-Modal Dual-Stream Decoupling for Few-Shot Anomaly Detection A Extended Analysis on Strict FPR Thresholds A.1 Evalu...

    work page arXiv 2024