acceptodds
Withdrawn from ICLR 2027

Unveiling Geometric Tension in Zero-Shot Anomaly Detection: Minimalist Calibration and Theoretical Bounds

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Zero-shot anomaly detection (ZSAD) identifies defects in unseen categories by transferring knowledge from a labeled source domain. Current CLIP-based methods compare patch features against learned tokens via cosine similarity, compressing each high-dimensional feature vector to a single scalar. We find that this compression introduces three interdependent losses: high-frequency texture, spatial neighborhoods, and distribution shape, and that they cannot be recovered independently. Suppressing high-frequency content alone significantly degrades pixel-level accuracy, and neither spatial regularization nor distributional calibration alone reverses the damage. When both are active, the same suppression becomes beneficial. We call this phenomenon directional dependency and design Complementary Feature Calibration, a lightweight three-component framework that jointly calibrates frequency responses, isotropic structural priors, and score distributions. A full factorial ablation characterizes interactions that baseline-centered, one-factor-at-a-time evaluations cannot reveal. Across five paired runs, adding FreqEnhance to SC+MS improves Pixel AUROC by 0.5 points on average; the gain is positive in four runs and its 95% confidence interval excludes zero. On zero-shot transfer from VisA to MVTec-AD, our method improves Pixel AUROC from 90.8% to 91.7% and Sample AUROC from 92.2% to 94.7%. The same dependency structure replicates across the tested architectures, backbone scales, and token paradigms, indicating that non-additive component interactions recur across several CLIP-based ZSAD settings.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Zero-Shot Anomaly Detection, Minimal Calibration, Manifold Distortion, Frequency Calibration

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_MVMROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the interactions among frequency, spatial, and distributional calibration in CLIP-based zero-shot anomaly detection. The authors identify a phenomenon termed directional dependency, where frequency suppression degrades performance when applied independently but becomes beneficial when combined with spatial regularization and distributional calibration. Based on this observation, the paper proposes Complementary Feature Calibration, consisting of FreqEnhance (FE), an Isotropic Structural Prior (SC), and Multi-Signal Classification (MS). Experiments across multiple datasets and architectures demonstrate promising results, achieving 91.7% Pixel AUROC and 94.7% Sample AUROC on MVTec-AD.

Critical strengths

S1. The paper investigates interactions among different calibration components rather than evaluating them independently. The full factorial ablation provides useful insights into the conditional effects of these components.

S2. The proposed framework is lightweight and introduces relatively few additional learnable parameters. The use of simple calibration mechanisms is practically appealing.

S3. The paper includes extensive evaluations across multiple datasets, backbone configurations, and token paradigms. The factorial ablations, cross-architecture experiments, and paired-seed analysis provide evidence supporting the observed directional dependency.

Critical weaknesses

W1. The individual components of the proposed framework are largely based on established techniques, including frequency filtering, spatial regularization, and statistical feature aggregation. Previous methods, such as FE-CLIP [1], have already explored frequency-aware representations for zero-shot anomaly detection, while VisualAD [2] provides the underlying architecture. Although the observed directional dependency is interesting, the proposed calibration modules themselves introduce limited architectural novelty.

W2. The paper provides a geometric analysis of how frequency suppression interacts with spatial regularization through changes in feature displacement and cosine similarity. However, this analysis primarily explains the interaction between FE and SC. The central empirical finding is that FE becomes beneficial only when both SC and MS are active. The current theoretical analysis does not sufficiently explain why distributional calibration through MS is also necessary for this sign reversal. A more explicit theoretical or empirical analysis of the role of MS would help establish a stronger connection between the proposed geometric explanation and the observed three-component dependency.

W3. The proposed framework is presented as a lightweight calibration mechanism that can be integrated into existing CLIP-based anomaly detection pipelines. However, FE introduces learnable parameters, SC requires training-time regularization, and MS involves trainable classification heads. The paper also states that the active calibration components are jointly optimized during training. It is therefore unclear whether the framework can be directly applied to pretrained anomaly detection models without retraining their original learnable components. If the proposed method could be applied as a general-purpose post-processing approach without joint retraining, it would significantly enhance its practical applicability and contribution. Further clarification or experiments demonstrating this capability would strengthen the paper.

[1] Gong, T., Chu, Q., Liu, B., Zhou, W., & Yu, N. (2025). FE-CLIP: Frequency Enhanced CLIP Model for Zero-Shot Anomaly Detection and Segmentation. Proceedings of the IEEE/CVF International Conference on Computer Vision.

[2] Hou, Y., Li, P., Liu, Z., Wang, Y., Ruan, Y., Qiu, J., & Xu, K. (2026). VisualAD: Language-Free Zero-Shot Anomaly Detection via Vision Transformer. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition.

Questions for authors

See Weaknesses

Reviewer_2euyOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates conditional interactions among modules in CLIP-based zero-shot anomaly detection and proposes a lightweight framework consisting of frequency modulation (FE), spatial regularization (SC), and multi-signal classification (MS). The authors find that FE degrades performance when used alone but becomes beneficial when combined with SC and MS. The findings are supported by factorial ablations, cross-setting experiments, and geometric analysis, with notable performance improvements on VisA-to-MVTec-AD transfer.

Critical strengths
  • The paper identifies conditional performance interactions among calibration modules, providing useful insights into their combined effects in zero-shot anomaly detection.

  • FE introduces only four learnable scalars, SC requires no additional learnable parameters, and MS uses lightweight classification heads.

  • The paper includes full factorial ablations, cross-method and backbone evaluations, and paired-seed experiments to support the observed interactions.

Critical weaknesses
  • FE, SC, and MS rely on existing techniques, while the observed interactions are not developed into a generalizable methodological principle.

  • The geometric analysis mainly addresses FE–SC interactions, without adequately explaining MS's role or FE's performance reversal.

  • Existing experiments establish performance interactions but lack targeted evidence explaining how the three modules jointly produce the reversal.

  • The paper attributes frequency, spatial, and distributional information loss to cosine similarity, although distributional information is lost during subsequent image-level aggregation.

Questions for authors
  • What is the methodological contribution beyond combining existing techniques and observing their interactions?

  • How does MS contribute to the three-way interaction and FE's performance reversal?

  • What targeted experiments validate the mechanism underlying the three-way interaction?

  • How do the authors reconcile attributing information loss to cosine similarity when distributional information is lost during image-level aggregation?

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

The paper investigates interactions among calibration components in CLIP-based zero-shot anomaly detection. It proposes Complementary Feature Calibration, combining layer-wise frequency modulation, spatial regularization, and multi-statistic image classification. The central finding is that frequency suppression harms performance in isolation but becomes beneficial when the other two components are jointly enabled. The authors examine this dependency through full factorial ablations, a five-seed paired comparison, and experiments across datasets, backbone scales, and token paradigms, including reverse transfer. The method reports improved detection and localization over VisualAD on the main VisA-to-MVTec benchmark with a small additional parameter budget.

Critical strengths
  1. The method improves image-level and pixel-level AUROC over VisualAD on the main transfer benchmark with a compact calibration design. FE introduces only four learnable scalars, while SC adds no inference parameters. The reverse-transfer and cross-architecture experiments provide additional evidence that the approach is useful beyond the original configuration (Tables 2, 4–5).
  2. The complete factorial ablation exposes a meaningful interaction: FE is detrimental alone but beneficial when SC and MS are both present. The matched-seed comparison supports this conditional benefit (Table 7a).
Critical weaknesses
  1. The methodological novelty is limited. FE, SC, and MS largely combine standard filtering, graph regularization, and statistical classification operations. The contribution therefore rests on the interaction analysis, but the experiments establish a conditional effect without sufficiently connecting it to the claimed three interdependent information losses. The theoretical bound also does not explain MS or the complete three-way interaction.
  2. The comparison with closely related work is insufficient. FE-CLIP directly addresses frequency-based ZSAD but is absent from the main evaluation. Fixed-strength smoothing and controls separating MS’s additional supervision from its aggregation would also help establish what the proposed design contributes beyond these simpler alternatives.
  3. Inconsistencies between the figures and text make the method and its robustness claims difficult to assess. Figure 2 does not match the definition of MS in §3.3, while Figure 4 does not support the stated regularization-weight plateau. These discrepancies require correction and a clear account of the implemented configuration.
Questions for authors

Q1. What evidence distinguishes the proposed information-loss explanation from ordinary smoothing and additional classification supervision? Could the authors provide controls using a fixed-strength frequency filter and a baseline with matched total classification-loss weight?

Q2. Can the authors provide a comparison with FE-CLIP under comparable conditions?

Q3. Does source training include anomalous images and pixel-level masks? Which source dataset is used to train the checkpoint evaluated on VisA in Table 2? Please also clarify whether Tables 2, 4, and 6 use the same AnomalyCLIP implementation or different variants, and explain the corresponding baseline differences.

Q4. The appendix states that Pixel AUROC varies by only ±0.3 points for λ ∈ [0.05, 0.5]. However, Figure 4 shows approximately 91.7 at λ = 0.1 and 90.2 at λ = 0.5—a decrease of 1.5 points. Could the authors explain this discrepancy between the description in the appendix and the results shown in Figure 4?

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

Related papers

Discussion (0)

Sign in to comment.