- Rating
- 3(+1)
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper proposes a new method for zero-shot anomaly detection named DeltaAD. It uses a memory bank of features gathered from a labelled auxiliary dataset. Based on the features in the memory bank, it calculates the patch features. The main idea is that the anomaly score should be calculated as the difference between the normal evidence and the anomalous evidence. The model is then tested on standard datasets.
Critical strengths
- The usage of residuals is well motivated
- I like the analysis of the method's performance based on the auxiliary dataset used for training. It also shows the importance of using the right one.
- The training-free results are good
Critical weaknesses
Major Weaknesses:
- The method is not compared to the current state of the art, AnomalyVFM [1] (with the RADIO backbone). AnomalyVFM outperforms it on almost all industrial datasets they test when paired with the correct backbone. I do not see a reason to compare it only to methods that use CLIP as a backbone; the main goal is to improve zero-shot anomaly detection overall, not just CLIP-based models. This holds even more for this method, as it is not tied down to using CLIP as a backbone.
- Only CLIP is examined as a backbone, raising the question of whether the findings about the residuals are general or specific only to CLIP. Stronger backbones, such as RADIO [2] and DINOv3 [3], should be explored (as they show the best performance in the related literature).
Minor Weaknesses:
- The U in the image-level detection differs from the U in the starting motivation, if I understood correctly? During the reading, this confused me a lot.
- What are the results if we use the maximum function in Equation 2?
- Similarly, what are the results if we take the largest patch residual (or the mean of top % residuals) in Equation 3?
- The residual learning module is complex and seems quite hand-crafted
[1] Fučka, M., et al. (2026). AnomalyVFM--Transforming Vision Foundation Models into Zero-Shot Anomaly Detectors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 35555-35566).
[2] Ranzinger, M., Heinrich, G., Kautz, J., & Molchanov, P. (2024, June). Am-radio: Agglomerative vision foundation model reduce all domains into one. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (pp. 12490-12500). IEEE.
[3] Oriane Siméoni et al. (2026). DINOv3. In Transactions on Machine Learning Research.
Questions for authors
- How does DeltaAD perform when using other backbones, such as RADIO and DINOv3? Does the motivation also hold for other backbones?
- How does the construction of Equation 2 and Equation 3 affect the final downstream performance?
7 comments
Q4: How does replacing the aggregation operations in Eq. (2) with maximum pooling affect DeltaAD-TF's performance?
A4: Following the reviewer's suggestion, we evaluated two Max-based alternatives to Eq. (2): Support-level Max and Context-level Max. The results demonstrate that the proposed aggregation outperforms both alternatives, highlighting the importance of integrating retrieval evidence across multiple shared visual contexts.
Table R3-C. Support- and context-level Max ablations of DeltaAD-TF on MVTec-AD and HeadCT (I-AUROC, %).
| Aggregation | MVTec-AD I-AUROC | HeadCT I-AUROC |
|---|---|---|
| DeltaAD (Eq. 2) | 88.5 | 89.9 |
| Support-level Max | 82.8 | 76.4 |
| Context-level Max | 65.5 | 65.2 |
Support-level Max reduces performance on both datasets, while Context-level Max causes substantially larger degradation. These results indicate that aggregating complementary evidence across multiple query-relevant visual contexts is particularly important for reliable differential response estimation, supporting our aggregation design in Eq. (2).
Q5: How does replacing the image-level aggregation in Eq. (3) with Max or Top-k% pooling of patch-level differential responses affect performance?
A5: Following the reviewer's suggestion, we compared the proposed two-state LogMeanExp aggregation with Max, Top-1%, and Top-10% pooling. The results demonstrate that our formulation achieves the best average image-level detection performance across all nine benchmarks, validating the effectiveness of aggregating normal and anomalous evidence separately before computing their difference.
Table R3-D. Image-level pooling ablation of DeltaAD-TF across nine image-level benchmarks (average metrics, %).
| Aggregation | I-AUROC (%) | I-AP (%) |
|---|---|---|
| DeltaAD (Eq. 3) | 85.3 | 84.4 |
| Max | 84.6 | 82.3 |
| Top-1% | 83.8 | 83.5 |
| Top-10% | 79.8 | 80.6 |
Unlike the alternatives, which directly pool the largest patch-level differential responses, our method separately aggregates normal and anomalous evidence using LogMeanExp () before computing their difference. This avoids relying solely on extreme patch responses and incorporates evidence from multiple patches, supporting the proposed image-level aggregation design.
Complete results are provided in Appendix E.6 (Table 23).
Q6: Is DeltaAD-LE's residual relation learning module unnecessarily complex or hand-crafted?
A6: DeltaAD-LE employs a lightweight, learnable relation module rather than hand-crafted residual scoring. Controlled ablations demonstrate its effectiveness over simpler alternatives, with only 6.99M trainable parameters.
Relation modeling. The module combines query features , normal explanations , signed and absolute residuals , and retrieval statistics . Their contributions are learned by the relation head rather than assigned fixed weights.
Table R3-E. Relation-input ablation of DeltaAD-LE on MVTec-AD (AUROC/PRO/AP, %).
| Relation Input | P-AUROC | PRO | I-AUROC | I-AP |
|---|---|---|---|---|
| Query-only | 91.6 | 82.2 | 92.4 | 96.1 |
| Residual | 91.7 | 84.1 | 92.5 | 96.7 |
| Full Relation | 93.1 | 87.2 | 94.3 | 97.8 |
The full representation leads on all four MVTec-AD metrics. Learnable ablations zero-fill removed inputs to preserve the relation-head architecture.
Model complexity. We compare DeltaAD-LE with training-based methods using the same frozen CLIP backbone:
Table R3-F. Accuracy and trainable-model size of CLIP-based methods across the 15 benchmarks.
| Method | I-AUROC | P-AUROC | Trainable Params | Model Size (MB) |
|---|---|---|---|---|
| MRAD-CLIP | 89.5 | 92.3 | 9.49M | 54 |
| VisualAD-CLIP | 88.3 | 89.5 | 31.52M | 126.12 |
| DeltaAD-LE (Ours) | 91.9 | 93.7 | 6.99M | 27.99 |
DeltaAD-LE achieves the highest average detection performance with fewer trainable parameters than MRAD-CLIP and VisualAD-CLIP. Only the retrieval projections and relation head are optimized, while the visual encoder remains frozen. These results support the effectiveness and parameter efficiency of our design (Sec. 4.3, Fig. 5; Appendix E.2, Table 15; Appendix D, Table 11).
Dear Reviewer Rb8g,
Thank you again for your time and constructive feedback. We sincerely appreciate your valuable suggestions.
We would like to briefly summarize the key findings from our response:
- Comparison with AnomalyVFM: Under matched CLIP and RADIO backbones, DeltaAD achieves competitive and, on several evaluation metrics, superior performance compared with AnomalyVFM across industrial and medical benchmarks. These results provide evidence that DeltaAD remains competitive with recent state-of-the-art approaches under comparable backbone settings.
- Backbone generalization: Evaluations with RADIO and DINOv3 demonstrate that our differential modeling framework generalizes beyond CLIP, with DeltaAD-LE maintaining strong performance across different pretrained visual representations.
- Aggregation and relation modeling: Controlled ablations validate the effectiveness of our aggregation strategies over Max/Top-k alternatives and demonstrate the benefits of learnable residual relation modeling over simpler scoring approaches. We have also clarified the notation ambiguity in the revised manuscript.
We hope these additional experiments and clarifications help address your concerns. If any questions remain, we would be more than happy to provide further clarification.
Thank you again for your time and consideration.
Best regards,
Authors
We sincerely appreciate your thoughtful and constructive feedback. Thank you for taking the time to carefully review our work and share your valuable comments and suggestions.
Q1: How does DeltaAD compare with AnomalyVFM under matched RADIO and CLIP backbones, particularly on industrial benchmarks?
A1: Following the reviewer's suggestion, we have compared DeltaAD with AnomalyVFM using matched RADIO and CLIP backbones on the same 15 industrial and medical benchmarks evaluated in our paper. The results challenge the reviewer's assessment that AnomalyVFM generally outperforms DeltaAD on industrial datasets.
Table R3-A. Backbone-matched comparison across the 15 benchmarks of this paper (average metrics, %; bold indicates the best result within each backbone).
DeltaAD-LE-RADIO achieves higher P-AUROC on five of six industrial benchmarks, while DeltaAD-TF-RADIO also achieves higher PRO. Under CLIP, DeltaAD-LE outperforms AnomalyVFM in P-AUROC across all 12 pixel-level benchmarks.
Additionally, using AnomalyVFM's released synthetic auxiliary dataset improves DeltaAD-TF's average P-AUROC from 87.5% to 90.3% without modifying the differential similarity formulation. We further validate DeltaAD's generalization across visual backbones in Q2.
AnomalyVFM results are primarily obtained from its official repository, supplemented by our evaluations on CVC-300 and HIS. Detailed DeltaAD results are provided in Sec. 4.3 (Table 2) and Appendix E.11 (Tables 29–30).
Q2: Does DeltaAD's differential modeling, particularly its learned residual relation, generalize beyond CLIP to stronger visual backbones such as RADIO and DINOv3?
A2: Following the reviewer's suggestion, we have extended both DeltaAD-TF and DeltaAD-LE to DINOv3 ViT-L/16 and RADIO v2.5-L, evaluating them on the same 15 benchmarks as our original experiments. The results demonstrate that DeltaAD's differential modeling generalizes beyond CLIP, with RADIO improving all four evaluation metrics for both variants.
Table R3-B. Backbone generalization of DeltaAD-TF and DeltaAD-LE on the same 15 benchmarks (average metrics, %; bold indicates the best result within each variant).
Notably, DeltaAD-LE maintains strong performance across all three backbones, demonstrating that its learned normal-explanation residual relation is not specific to CLIP representations. RADIO further improves both variants, while DINOv3 also improves their pixel-level performance.
These experiments retain the original differential modeling mechanisms with frozen visual encoders and further demonstrate the strong generalizability of DeltaAD across different pretrained visual backbones, supporting the transferability of its differential modeling framework beyond CLIP. Complete results are provided in Sec. 4.3 (Table 2) and Appendix E.11 (Tables 29–30).
Q3: Clarification of the notation ambiguity between the common evidence component and the image-level aggregation in DeltaAD-TF.
A3: We have resolved this notation ambiguity in the revised manuscript. In Sec. 3.1.1, Ui=(EiA+EiN)/2 denotes the common component of normal and anomalous evidence, whereas the original Uxs(T) in Eq. (3) denotes image-level evidence aggregation.
We have renamed the latter as Gxs(T) and updated Eq. (4) accordingly:
SxTF=GxA(T)−GxN(T).
This is purely a notational change and does not affect the method or experimental results.