acceptodds
Under review as a conference paper at ICLR 2027

XPAD: Language-Free and Training-Free Zero-Shot Industrial Anomaly Detection via Cross-Dataset Prototypes

Abstract

Zero-shot anomaly detection (ZSAD) detects anomalies of unseen categories without target-domain supervision. Existing ZSAD methods generally fine-tune on an auxiliary dataset to learn transferable textual prompts or visual representations, incurring considerable training costs and leaving the question of what is actually transferred unresolved. Across six industrial benchmarks, we observe that a frozen encoder already organizes normal and anomalous patch features into a shared, dataset-agnostic direction, which can be estimated without any training. This suggests that an effective scoring function can be derived directly from this direction. We instantiate this idea as XPAD (Cross-dataset Prototypes for Anomaly Detection), a language-free and training-free framework. XPAD aggregates a single annotated auxiliary dataset into cross-dataset prototypes, and scores test patches by prototype similarity. To sharpen this comparison, it further separates the normal prototype into foreground and background sub-prototypes, and calibrates the aggregation-degree mismatch between dataset-level prototypes and individual query patches. Without textual prompts, additional training, or target-domain data, XPAD achieves state-of-the-art average performance on six industrial benchmarks, and runs directly on pretrained vision backbones such as the CLIP image encoder and DINOv2.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
industrial anomaly detection, zero-shot detection, language-free, training-free

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_ZADUOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

Motivation: Existing ZSAD methods typically rely on auxiliary-dataset training to learn transferable anomaly representations. However, frozen visual encoders may already possess cross-dataset anomaly-discriminative capabilities, potentially eliminating the need for additional training.

Contribution: The authors propose XPAD, which constructs normal and anomaly prototypes from an annotated auxiliary dataset and incorporates foreground–background separation and aggregation-degree calibration. The framework enables zero-shot anomaly detection without textual prompts, gradient-based training, or target-domain data.

Critical strengths
  1. The paper is reasonably well organized, and the overall framework is straightforward to understand. The motivation and methodological components are presented in a relatively clear manner.

  2. XPAD explores a relatively less-studied training-free yet annotation-dependent ZSAD paradigm. By leveraging cross-dataset prototype statistics without gradient-based learning, the work offers an alternative perspective on anomaly knowledge transfer and provides some useful insights for future training-free anomaly detection research.

Critical weaknesses
  1. The "training-free" claim is potentially misleading, and its practical significance remains questionable. Although XPAD does not involve gradient-based optimization or model fine-tuning, it relies on an annotated auxiliary dataset to construct normal and anomaly prototypes. In particular, the anomaly prototype cannot be established without pixel-level anomaly annotations, indicating that the framework still depends on supervised anomaly knowledge. This differs from more conventional zero-shot approaches, such as WinCLIP, which do not require annotated anomalies for their zero-shot inference. Therefore, XPAD is more accurately characterized as a gradient-free, annotation-dependent cross-dataset transfer method rather than a fully supervision-free approach. The authors should clarify this distinction and better justify the practical advantages of the proposed paradigm, particularly considering the cost of acquiring pixel-level anomaly annotations.

  2. The central motivation lacks sufficient logical justification and empirical validation. In Sec. 3.1, the authors infer the existence of a shared, transferable anomaly-discriminative direction from the observed similarities among dataset-level normal and anomaly prototypes. However, this inference is not sufficiently justified. First, high within-class prototype similarities across datasets do not necessarily establish consistent normal-to-anomaly discriminative directions. Second, dataset-level prototype separation does not directly imply effective anomaly discrimination at the patch level. The current analysis does not adequately examine either of these assumptions. The authors should provide more direct evidence, such as pairwise cosine similarity analysis of cross-dataset normal-to-anomaly difference vectors and patch-level discrimination experiments, to substantiate the central motivation underlying XPAD.

Questions for authors
  1. How do the authors justify the "training-free" characterization of XPAD given its dependence on pixel-level anomaly annotations from an auxiliary dataset? While the method avoids gradient-based optimization, constructing anomaly prototypes still requires explicit supervised anomaly information. What are the practical advantages of this paradigm over existing training-free approaches, such as WinCLIP, that do not require annotated anomaly samples?

  2. What evidence supports the claim that dataset-level prototype similarities imply a shared, transferable anomaly-discriminative direction? High within-class similarities do not necessarily guarantee consistent normal-to-anomaly directions across datasets or effective patch-level discrimination. Could the authors provide additional analyses of cross-dataset normal-to-anomaly difference vectors and patch-level discriminative performance to substantiate this central claim?

Reviewer_NSLjOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes XPAD, a language-free and training-free framework for zero-shot industrial anomaly detection. The method identifies a shared normal–anomaly direction in frozen visual features and constructs cross-dataset prototypes from an annotated auxiliary dataset. It further introduces Foreground–Background Normal Prototype Separation (FBNPS) and Aggregation-Degree Calibration (ADC) to improve anomaly scoring. Experiments on six industrial benchmarks using CLIP and DINOv2 demonstrate competitive image-level and pixel-level performance without additional model training.

Critical strengths

S1. The observation that frozen visual features exhibit consistent cross-dataset normal–anomaly separation provides an interesting perspective on zero-shot anomaly detection.

S2. The proposed framework requires no additional training or textual prompts. Its lightweight prototype-based design is practically appealing and can be applied to different pretrained visual backbones.

S3. The paper presents strong empirical results across six industrial benchmarks. The ablation studies, auxiliary-dataset comparisons, and experiments with different backbones provide useful evidence supporting the proposed framework.

Critical weaknesses

W1. The proposed framework builds upon established techniques, including prototype-based anomaly scoring, foreground–background separation, and feature aggregation. MRAD-TF [1] has already explored training-free anomaly detection using frozen visual features and auxiliary-data memory banks, while FB-CLIP [2] investigates foreground–background disentanglement for anomaly detection. Although the cross-dataset prototype observation is interesting, the methodological novelty of the proposed components compared with existing approaches could be further clarified. Detailed analysis of why a few dataset-level prototypes are preferable to existing memory-based representations would strengthen the contribution.

W2. The performance of XPAD appears sensitive to the choice of auxiliary dataset. Although Table 6 evaluates different auxiliary datasets, substantial performance differences are observed in certain settings. For example, on MPDD, the image-level AUROC decreases from 79.0% to 64.8% when replacing MVTec-AD with DTD-Synthetic as the auxiliary dataset. This raises concerns about the robustness of the claimed dataset-agnostic normal–anomaly direction, particularly when the auxiliary dataset has limited diversity. A more systematic analysis of how auxiliary dataset size and diversity affect prototype quality would strengthen the generalization claims. W3. The authors are encouraged to provide more comprehensive comparisons with recent multi-prompt and fine-grained prompt learning methods, such as FAPrompt [3] and TokenCLIP [4].

W4. The proposed method represents normal and anomalous features using a small number of dataset-level prototypes. Although Table 5 evaluates different normal-prototype configurations, the paper does not investigate the effect of the number of prototypes on detection performance. An ablation study using different numbers of prototypes, such as multiple cluster-based prototypes for normal and anomalous features, would help clarify whether the current design is sufficient to capture diverse anomaly patterns.

W5. The proposed FBNPS relies on the assumption that the four image corners belong to the background. However, this assumption may not hold for texture images, where foreground and background are not clearly distinguishable, or for images in which foreground objects extend to the boundaries. Although the appendix evaluates FBNPS on texture datasets, the validity of the resulting foreground–background separation under these conditions remains unclear. Further analysis of the generated masks and their impact on prototype quality would strengthen the generalizability of this design.

[1] Xu, C., Lv, C., Chen, Q., Zhang, F., & Zhang, Z. (2026). MRAD: Zero-Shot Anomaly Detection with Memory-Driven Retrieval. International Conference on Learning Representations.

[2] Hu, M., Huo, Y., Dou, M., Yin, J., Zhao, P., Wang, Y., Hu, C., Hu, B., & Wang, Q. (2026). FB-CLIP: Fine-Grained Zero-Shot Anomaly Detection with Foreground-Background Disentanglement. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition.

[3] Zhu, J., Ong, Y.-S., Shen, C., & Pang, G. (2025). Fine-grained Abnormality Prompt Learning for Zero-shot Anomaly Detection. Proceedings of the IEEE/CVF International Conference on Computer Vision.

[4] Zhou, Q., Gao, B., Pang, G., Wang, X., Chen, J., & He, S. (2025). TokenCLIP: Token-wise Prompt Learning for Zero-shot Anomaly Detection. arXiv preprint, arXiv:2510.21171.

Questions for authors
  1. What are the key methodological differences between XPAD and existing training-free or prototype-based anomaly detection approaches, particularly MRAD-TF [1]? Could the authors further clarify the advantages of dataset-level prototypes over memory-bank retrieval?

  2. Table 6 shows substantial performance variations with different auxiliary datasets. How does the diversity of the auxiliary dataset influence the quality of the prototypes and the robustness of cross-dataset anomaly detection?

  3. How does the number of normal and anomaly prototypes affect detection performance? Could the authors provide an ablation study with different prototype counts to justify the current design?

  4. How robust is FBNPS when the assumption that the four image corners belong to the background is violated, particularly for texture images or objects occupying the image boundaries? Could the authors further analyze the generated foreground masks and their influence on anomaly detection performance?

Reviewer_x572OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces XPAD, an efficient and performant zero shot anomaly detection mechanism.

  • It constructs class prototypes through foundation encoders (DINOv2, CLIP) and observes that intra-class prototype similarities exceed inter-class prototype similarities.
  • It leverages this observation to build its prototypes on an annotated auxiliary dataset.
  • It leverages two heuristics to strengthen its scoring: foreground-background prototype normal separation and aggregation-degree calibration.
  • It validates its performance across 6 industrial anomaly datasets.
Critical strengths

The observation that normal/anomalous prototypes can separate normal/anomalous patches across datasets is an interesting finding. The method is simple and gets good performance, that's also valuable.

Critical weaknesses

My main concern is that some important heuristics introduced on this paper are unjustified. They are motivated with intuitive but unsubstantiated claims and lack sufficient validation.

FBNPS is fairly justified but its foreground-background separation's performance is not validated, we don't know if it actually segments foreground-background patches correctly.

ADC's first motivating paragraph contains claims that are never tested:

  • "Such highly aggregated statistics are stable, lie in the high-density region of the feature distribution [...]". Not tested, the feature distribution could be multimodal and the mean might lie on a low-density region.
  • "Cosine similarities measured between entities of different aggregation degrees are therefore statistically biased: the distribution of similarities to a prototype is systematically shifted for individual patches compared with what would be obtained by comparing two equally aggregated features. Using the raw per patch scores directly thus yields miscalibrated scores and confidence—especially for image-level evidence, where a few outlier patches can dominate.". The claimed statistical bias isn't measured and ADC isn't shown to fix it, just that it improves downstream performance.
Questions for authors

Points that could affect my assessment:

  • Actual foreground-background separation performance should be tested and compared with at least one standard baseline.
  • Either the ADC's claimed statistical bias is measured and its solution validated, or the framing should be relaxed to a heuristic that allows image-level anomaly scores to influence patch-level anomaly scores.

Comments on the paper

1 comment
Area_Chair_tn1UOpenReview ↗

A Gentle Reminder: About One Day Left in the Discussion Period

Dear Reviewers,

This is a gentle reminder that approximately one day remains in the discussion period. Please review the authors' responses and other reviewers' comments, and engage in the discussion to clarify which concerns have been addressed and which remain unresolved.

Thank you for your time and effort.

Best,
AC

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.