- Rating
- 1
- Confidence
- 4
- Informativeness
- 1
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper proposes EAD, a CLIP-based framework for zero-shot anomaly detection. It introduces Category-Agnostic Prompt Synthesis (CAP) to generate image-conditioned prompts from visual prototypes and a shared learnable anchor, together with Hierarchical Parallel Adaptation (HiP) to integrate multi-layer features without modifying the frozen CLIP backbone. The authors argue that class-semantic priors can introduce domain-dependent bias and report competitive performance and improved inference efficiency on 14 industrial and medical benchmarks.
Critical strengths
- The paper addresses the practical limitations of category-dependent prompts and computational overhead in CLIP-based anomaly detection.
- The parallel adaptation design provides meaningful efficiency improvements over AdaCLIP, reducing computational cost while improving inference speed under the reported evaluation setup.
Critical weaknesses
1. Limited methodological novelty. The proposed framework largely follows the existing paradigm of dynamic prompt learning for CLIP-based anomaly detection. Related methods such as AdaCLIP, FAPrompt, and MRAD-CLIP have already explored learnable prompts, visual-conditioned prompt generation, or dynamic prior integration. Although the proposed layer-wise construction differs in implementation, its fundamental novelty over these methods remains unclear.
2. Insufficient evidence for the necessity of the proposed modules. Table 8 shows that removing the visual prototype component causes relatively limited performance degradation, whereas removing the learnable anchor leads to much larger drops. This raises questions about whether the main improvements come from the proposed visual prompt synthesis or simply from learnable prompt tokens. A controlled comparison with a single shared prompt across layers would help clarify the actual contribution of the layer-wise design.
3. Inadequate comparisons with recent methods. The experimental comparisons omit several recent and highly relevant CLIP-based ZSAD methods, particularly FAPrompt and MRAD-CLIP. Moreover, the improvements over AdaCLIP on industrial anomaly localization are marginal (0.3 AUROC and 0.1 max-F1). Without comparisons against stronger and more recent baselines, the claimed performance advantages remain insufficiently established.
4. Insufficient justification for the domain-dependent bias claim. The paper argues that class-semantic priors introduce domain-dependent bias, but this conclusion is primarily supported by differences in performance gains across domains. AnomalyCLIP already employs static, object-agnostic learnable prompts without relying on category-specific semantics. How does the proposed method fundamentally differ from this approach? Could the observed improvements mainly result from learnable prompt optimization rather than removing category semantics? Controlled experiments using correct, generic, and corrupted class descriptions would help substantiate the central claim.
5. Limited evidence for the accuracy-efficiency trade-off and statistical reliability. Although the proposed framework achieves notable computational savings compared with AdaCLIP, the efficiency gains largely rely on an existing parallel adaptation strategy. It remains unclear whether the method offers a superior accuracy-efficiency trade-off compared with more recent approaches. Furthermore, given the marginal improvements on several benchmarks, results across multiple random seeds are needed to establish the statistical reliability of the reported gains.
Questions for authors
Please refer to the concerns and questions raised in the Critical Weaknesses section.