- Rating
- 3
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
The paper proposes a new approach to continual zero-shot anomaly detection, called 0-CLIP. The main hypothesis is that CLIP already has normality and abnormality baked into its projection layer. To check this, they construct a “new” projection layer from the existing one by performing an SVD and gradually building it using the first test sample. The method is evaluated across multiple standard industrial and medical datasets, on which it achieves good results.
Critical strengths
- The overall idea is quite simple
- The paper is well written and easy to follow.
Critical weaknesses
Major Weaknesses:
- The method is not compared to the current state of the art for continual zero-shot anomaly detection, ReMem [1]. This method outperforms it on the two datasets they test. It is unclear how it works on others. Since they have released the code publicly, I do not see a reason why 0-CLIP shouldn’t be compared to it. This might overturn the claim about SOTA results.
- The method works only with the final-layer features of the vision encoder, neglecting information from earlier layers.
Minor Weaknesses:
- How does the method work when using a 0-shot method that also fine-tunes the projection head? For example, AdaCLIP?
- How are the std’s for the 5 runs calculated? The way it is calculated might underestimate the actual performance jumps based on the first image in the test stream.
- I would mention the latest language-free zero-shot methods [2,3] in the related work, as they also achieve better results on some datasets.
Remarks:
- A mistake in the citations: DRAEM: A discriminatively trained model for surface anomaly detection -> DRÆM – A discriminatively trained reconstruction embedding for surface anomaly detection
[1] Yi, L., Chen, Z., Wu, G., Ding, J., Wang, X., & Ning, Z. (2026). ReMem: A Dynamic Memory Evolution Detector for Zero-Shot Anomaly Detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Findings (pp. 7697-7705).
[2] 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).
[3] Hou, Y., et al. (2026). VisualAD: Language-Free Zero-Shot Anomaly Detection via Vision Transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 21346-21356).
Questions for authors
- How does 0-CLIP perform in comparison to ReMem on datasets besides MVTec AD and VisA?
- Could the method be expanded to work with earlier layers as well?
- How does the method work when applied to a 0-shot model that tunes the projection head, e.g. AdaCLIP?
- How are the Standard Deviations for the multiple seed runs calculated?
6 comments
W2/Q2. Could the method be expanded to work with earlier layers as well?
Thank you for your question. 0-CLIP already uses earlier layers: it takes patch tokens from layers 12, 16, 20 and 24 (Eq. (1) and Sec. 4.1) and builds a separate operator for each layer.
We compared this default with the final layer alone (P-AUROC / PRO / I-AUROC / I-AP):
| Layers | MVTec-AD | VisA |
|---|---|---|
| 12/16/20/24 (default) | 95.2 / 88.1 / 94.7 / 97.3 | 95.3 / 86.1 / 85.1 / 87.0 |
| final layer (24 only) | 92.2 / 82.6 / 93.6 / 96.8 | 92.1 / 82.7 / 82.9 / 85.7 |
Using only the final layer lowers all eight metrics. Combining multiple layers is therefore highly beneficial and indispensable to 0-CLIP.
Revision: Sec. 3.1 now states explicitly that 0-CLIP constructs these operators for each layer.
W3/Q3. How does the method work when applied to a 0-shot model that tunes the projection head, e.g. AdaCLIP?
Using AdaCLIP's official code and weights, we applied our construction to its trained patch heads and kept everything else unchanged (P-AUROC / PRO / I-AUROC / I-AP):
| AdaCLIP patch heads | MVTec-AD | VisA |
|---|---|---|
| Original | 86.8 / 33.8 / 92.0 / 96.4 | 95.1 / 71.3 / 83.0 / 84.9 |
| With our construction | 43.9 / 12.8 / 76.8 / 88.2 | 57.3 / 32.1 / 62.9 / 68.9 |
The second row is our run of the official code with only the patch heads changed. AdaCLIP's heads are new layers trained on labeled data, and this training has already aligned their polarity and singular-value weights with AD. Flipping and reweighting them therefore undoes what training learned. 0-CLIP is meant to replace this training.
Revision: Added as Table 11 (Appendix E.6) and noted in the Discussion (Sec. 5).
W4/Q4. How are the Standard Deviations for the multiple seed runs calculated?
Each seed shuffles the test stream of every class, which changes the calibration image. For each seed we average the metrics over classes, and report the mean and standard deviation over the 5 seeds (Sec. 4.1 and Appendix E.7). Below we also compute the standard deviation over the 5 seeds for each class first and then average over classes (mean ± dataset std / per-class std; identical on 1-class datasets):
| Dataset (#cls) | Setting | P-AUROC | PRO | I-AUROC | I-AP |
|---|---|---|---|---|---|
| MVTec-AD (15) | 0-CLIP$^\dagger$ | 97.83 ± 0.25 / 0.46 | 93.95 ± 0.21 / 0.57 | 98.22 ± 0.12 / 0.60 | 99.34 ± 0.07 / 0.26 |
| 0-CLIP | 94.91 ± 0.20 / 0.47 | 87.30 ± 0.47 / 0.66 | 94.44 ± 0.32 / 0.69 | 97.35 ± 0.21 / 0.48 | |
| VisA (12) | 0-CLIP$^\dagger$ | 98.90 ± 0.03 / 0.08 | 93.73 ± 0.23 / 0.47 | 92.92 ± 0.31 / 0.67 | 93.73 ± 0.23 / 0.58 |
| 0-CLIP | 95.32 ± 0.10 / 0.19 | 86.16 ± 0.04 / 0.26 | 85.29 ± 0.16 / 0.33 | 86.98 ± 0.11 / 0.27 | |
| BTAD (3) | 0-CLIP$^\dagger$ | 98.03 ± 0.03 / 0.09 | 83.89 ± 0.78 / 0.99 | 96.54 ± 0.36 / 0.60 | 97.90 ± 0.30 / 0.42 |
| 0-CLIP | 96.48 ± 0.23 / 0.26 | 81.74 ± 0.57 / 0.66 | 94.98 ± 0.64 / 1.15 | 94.59 ± 0.35 / 0.65 | |
| MPDD (6) | 0-CLIP$^\dagger$ | 97.47 ± 0.36 / 0.51 | 93.54 ± 0.68 / 1.02 | 73.27 ± 2.12 / 5.18 | 77.48 ± 1.60 / 3.32 |
| 0-CLIP | 96.17 ± 0.07 / 0.11 | 91.10 ± 0.19 / 0.28 | 76.79 ± 0.26 / 0.70 | 81.84 ± 0.13 / 0.44 | |
| DTD-Syn (12) | 0-CLIP$^\dagger$ | 98.33 ± 0.01 / 0.14 | 95.19 ± 0.10 / 0.34 | 98.03 ± 0.14 / 0.47 | 99.25 ± 0.03 / 0.21 |
| 0-CLIP | 97.99 ± 0.05 / 0.10 | 93.99 ± 0.17 / 0.27 | 98.01 ± 0.05 / 0.24 | 99.28 ± 0.01 / 0.07 | |
| SDD (1) | 0-CLIP$^\dagger$ | 93.51 ± 0.27 | 77.84 ± 0.64 | 85.56 ± 0.89 | 76.24 ± 1.98 |
| 0-CLIP | 91.88 ± 0.24 | 77.97 ± 0.26 | 84.52 ± 0.16 | 78.41 ± 0.29 | |
| DAGM (10) | 0-CLIP$^\dagger$ | 98.19 ± 0.03 / 0.13 | 96.38 ± 0.07 / 0.19 | 99.56 ± 0.08 / 0.08 | 96.72 ± 0.23 / 0.29 |
| 0-CLIP | 96.91 ± 0.05 / 0.08 | 94.08 ± 0.12 / 0.17 | 99.53 ± 0.02 / 0.07 | 98.30 ± 0.12 / 0.24 | |
| KSDD2 (1) | 0-CLIP$^\dagger$ | 98.52 ± 0.29 | 96.17 ± 0.32 | 96.42 ± 0.18 | 91.78 ± 0.23 |
| 0-CLIP | 97.28 ± 0.30 | 94.81 ± 0.25 | 95.48 ± 0.09 | 89.57 ± 0.15 | |
| Real-IAD (30) | 0-CLIP$^\dagger$ | 97.38 ± 0.05 / 0.14 | 88.59 ± 0.24 / 0.58 | 80.84 ± 0.06 / 0.39 | 78.99 ± 0.11 / 0.48 |
| 0-CLIP | 95.34 ± 0.05 / 0.12 | 87.37 ± 0.08 / 0.29 | 76.29 ± 0.12 / 0.21 | 74.23 ± 0.11 / 0.25 | |
| Liver CT (1) | 0-CLIP$^\dagger$ | 97.53 ± 0.13 | 92.01 ± 0.54 | – | – |
| 0-CLIP | 95.85 ± 0.07 | 90.87 ± 0.17 | – | – | |
| Brain MRI (1) | 0-CLIP$^\dagger$ | 96.88 ± 0.06 | 81.16 ± 0.36 | 73.96 ± 0.71 | 91.13 ± 0.14 |
| 0-CLIP | 93.74 ± 0.06 | 76.28 ± 0.18 | 80.42 ± 0.21 | 94.72 ± 0.05 | |
| Retina RESC (1) | 0-CLIP$^\dagger$ | 95.11 ± 0.24 | 81.66 ± 0.80 | – | – |
| 0-CLIP | 87.36 ± 0.14 | 59.01 ± 0.63 | – | – | |
| TN3K (1) | 0-CLIP$^\dagger$ | 80.01 ± 0.45 | 57.87 ± 0.34 | – | – |
| 0-CLIP | 82.44 ± 0.18 | 52.99 ± 1.04 | – | – | |
| Br35H (1) | 0-CLIP$^\dagger$ | – | – | 89.49 ± 0.73 | 83.75 ± 0.79 |
| 0-CLIP | – | – | 96.06 ± 0.05 | 94.59 ± 0.05 | |
| Retina OCT (1) | 0-CLIP$^\dagger$ | – | – | 88.07 ± 1.32 | 95.92 ± 0.45 |
| 0-CLIP | – | – | 79.29 ± 0.16 | 91.99 ± 0.06 |
For offline 0-CLIP, even the per-class std stays small across datasets (mostly below 0.7%), showing minimal sensitivity to the calibration image. Larger online fluctuations such as on MPDD come from stream order in RareCLIP's memory.
We also tested the two extreme cases of the first image directly. In each seed's shuffled test stream, we calibrated every class once on its first normal image and once on its first anomalous image, and then scored the whole stream as usual.
| Dataset | Calibration image | P-AUROC | PRO | I-AUROC | I-AP |
|---|---|---|---|---|---|
| MVTec-AD | First image (default) | 94.91 ± 0.20 / 0.47 | 87.30 ± 0.47 / 0.66 | 94.44 ± 0.32 / 0.69 | 97.35 ± 0.21 / 0.48 |
| Normal | 94.92 ± 0.14 / 0.50 | 87.41 ± 0.28 / 0.52 | 94.40 ± 0.33 / 0.65 | 97.34 ± 0.23 / 0.48 | |
| Anomalous | 94.91 ± 0.20 / 0.47 | 87.28 ± 0.47 / 0.66 | 94.41 ± 0.33 / 0.69 | 97.33 ± 0.22 / 0.49 | |
| VisA | First image (default) | 95.32 ± 0.10 / 0.19 | 86.16 ± 0.04 / 0.26 | 85.29 ± 0.16 / 0.33 | 86.98 ± 0.11 / 0.27 |
| Normal | 95.25 ± 0.11 / 0.19 | 86.08 ± 0.09 / 0.25 | 85.25 ± 0.20 / 0.38 | 86.96 ± 0.14 / 0.31 | |
| Anomalous | 95.33 ± 0.11 / 0.16 | 86.19 ± 0.09 / 0.25 | 85.31 ± 0.16 / 0.30 | 87.00 ± 0.10 / 0.23 |
In both cases, all metrics stay within 0.15 points of the default. Eq. (7) averages prototype contrast over all patches of the calibration image. Even in an anomalous image, most patches are normal, so defects barely affect which directions are selected.
Revision: The calibration-image comparison is added as Table 13 (Appendix E.8) and referenced in the ablation study (Sec. 4.3).
Thank you for pointing us to ReMem. In short, ReMem is stronger on MVTec-AD and VisA, while 0-CLIP$^\dagger$ is higher on most other datasets and on average.
We ran its official code, here are results (pixel P-AUROC / PRO; image I-AUROC / I-AP):
Two differences in setting between 0-CLIP and ReMem:
Revision: Added as Table 8 (Appendix E.3).