acceptodds
Under review as a conference paper at ICLR 2027

AnomalyLoop: Closing the Generation–Detection Loop in Open-Set Training-Free Anomaly Generation

Abstract

Industrial visual anomaly detection has long been constrained by the scarcity of real anomalous samples and the open-ended nature of anomaly types. To alleviate this issue, open-set training-free anomaly generation has gradually emerged as a promising data synthesis paradigm, enabling the synthesis of unseen defects without anomaly specific training. However, existing methods still face two key challenges: spatial control is difficult to confine to fine-grained candidate anomaly formation regions, and generation remains separate from anomaly evaluation, such that insufficiently formed anomalies are only identified after generation and cannot be corrected through detection feedback. To address these challenges, we propose AnomalyLoop, a two-stage framework for open-set training-free anomaly generation comprising the exploration of fine-grained anomaly formation regions and generation–detection closed-loop trajectory optimization. The key idea is to transform anomaly detection from post-generation evaluation into terminal feedback for the full diffusion trajectory, thereby closing the loop between anomaly generation and detection. In the first stage, we propose an Anomaly Specific Region Exploration mechanism (ASRE) to construct fine-grained candidate anomaly formation regions. In the second stage, we use terminal detection results as the primary feedback and propagate this feedback backward through Trajectory Optimal Control (TOC) to optimize control decisions along the full diffusion trajectory, enabling detection-driven anomaly formation. Furthermore, we jointly incorporate semantic consistency and history-aware diversity as auxiliary terminal constraints to preserve target semantics while maintaining sample diversity. Experiments on MVTec AD and VisA show that AnomalyLoop achieves state-of-the-art generation quality and diversity among open-set training-free methods, while overall improving downstream anomaly detection performance.

Primary area:
generative models
Keywords:
Anomaly Generation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_SaZpOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes AnomalyLoop, a two-stage framework for open-set training-free anomaly generation. The first stage, Anomaly Specific Region Exploration (ASRE), identifies fine-grained candidate anomaly regions. The second stage employs Trajectory Optimal Control (TOC) to optimize diffusion sampling using anomaly detection feedback, with additional semantic consistency and diversity constraints. Experiments on MVTec AD and VisA demonstrate improvements in generation quality, diversity, and downstream anomaly detection performance.

Critical strengths
  1. The proposed generation–detection closed-loop framework provides an interesting approach to improving anomaly generation without fine-tuning the pretrained diffusion model.

  2. The method achieves promising improvements in generation quality and downstream anomaly detection on MVTec AD and VisA, demonstrating its potential practical value.

Critical weaknesses
  1. The proposed ASRE strategy lacks sufficient validation regarding the spatial diversity of selected anomaly regions. Its advantages over simpler random-mask generation strategies remain unclear.

  2. The downstream evaluation relies on only 100 generated anomalous images per category from a single normal reference image, raising concerns about instance-specific bias and generalization. Additionally, the lack of generation-time measurements and broader comparisons limits the assessment of practical efficiency.

Questions for authors
  1. Could you please clarify whether ASRE produces sufficiently diverse anomaly locations across repeated generations? A comparison of downstream detection performance using ASRE versus random or noise-generated binary masks would help demonstrate its effectiveness.

  2. Could you please clarify whether the downstream detectors are trained separately for each category or jointly across categories? If trained separately, would 100 synthetic images per category be sufficient for effective training? If trained jointly, how does this affect the detection performance of individual categories? Furthermore, generating all anomalies from a single normal reference image may introduce instance-specific bias due to limited object appearance diversity. Could you please clarify how this potential issue is addressed?

  3. Could you please provide the missing baseline results in Figure 5 and other qualitative comparisons, particularly those marked as N/A? These comparisons appear feasible to complete.

  4. Could you please include experimental comparisons with other trajectory optimization or anomaly-aware sampling methods, such as FAST (NeurIPS 2025) or other methods, particularly in terms of downstream detection performance and generation efficiency?

4 comments

We thank the reviewer for finding our approach interesting and for the positive comments on its promising performance.

Table 3. Controlled ablation results on MVTec AD.

VariantRandom maskStage 1LsubL_{\mathrm{sub}}LsemL_{\mathrm{sem}}LdivL_{\mathrm{div}}ISIC-LI-AUCI-F1I-APM-act
ASRE–✓–––1.21860.186370.1985.4587.6183.65
ASRE + DF–✓✓––1.26560.214775.4988.3587.8988.58
w/o LdivL_{\mathrm{div}}–✓✓✓–1.26470.204973.4786.7086.9686.67
w/o LsemL_{\mathrm{sem}}–✓✓–✓1.26510.219074.0287.0786.7783.85
w/o ASRE✓–✓✓✓1.24450.159274.7788.1688.21–
Ours–✓✓✓✓1.26310.221176.0288.1288.2689.67

Q1. Downstream detection with ASRE versus random binary masks.

In the w/o ASRE variant in Table 3, we replace the candidate region selected by ASRE with a random binary mask sampled within the object mask. We keep the area budget and subsequent generation settings unchanged and retrain the downstream U-Net for each variant under the same protocol. This random-mask comparison demonstrates the effectiveness of ASRE.

Table 13. Per-category vs. joint training for one-shot detection on MVTec AD.

MethodI-AUCI-F1P-AUCP-F1PRO
Joint training97.997.095.460.289.6
Ours98.296.997.865.993.7

Table 14. Effect of normal reference selection on one-shot detection on MVTec AD.

InstanceI-AUC (%)I-F1 (%)P-AUC (%)P-F1 (%)PRO (%)
198.196.997.866.193.8
297.396.498.266.594.5
397.896.498.166.293.8
Mean ± std97.73 ± 0.3396.57 ± 0.2498.03 ± 0.1766.27 ± 0.1794.03 ± 0.33

Q2. Per-category training, sample size, and instance-specific bias.

Table 2 summarizes the test results across categories. For a fair comparison with prior work, we follow the same protocol and train a separate detector for each category, then test it on that category. We use the AnomalyGPT-based detection framework adopted by AnomalyAny, which extracts visual features with a pretrained CLIP encoder and aligns patch features with text embeddings through a lightweight linear adapter. During training, we keep CLIP frozen and update only the adapter parameters, so the detector does not need to learn visual representations from scratch. Under our 1-shot protocol, 100 synthetic anomaly images per category were sufficient to train the downstream detector effectively. Table 13 reports the test results for a single detector trained jointly across all categories. Its overall performance was lower than that of the per-category detectors, indicating that joint training affects performance on individual categories. We plan to explore how to improve per-category performance under joint training in future work. To assess potential instance-specific bias from using a single normal reference image, we repeated the experiment three times. In each run, we randomly selected one normal reference image per category, regenerated the anomaly samples, and evaluated downstream detection performance. The detection results were similar across the three runs (Table 14), indicating that our framework is insensitive to instance-specific bias in this setting.

Q3. Missing baseline results in Figure 5.

We reproduced AnomalyDiffusion’s generation results on VisA and added the missing qualitative comparison to Figure 5 in place of the entries marked as N/A.

Table 7. Inference time per generated image, total training time, and GPU memory usage.

MethodTotal training time (h)Training GPU memory (GB)Inference time (s/image)Inference GPU memory (GB)
AnomalyDiffusion (2024)158191413
FAST (2025)9182948
Anomagic (2026)69482635
AnoStyler (2026)0–155
AnomalyAny (2025)0–7225
Ours0–3916

Table 8. One-shot detection on MVTec AD with an AnomalyGPT-based detector.

MethodI-AUCI-F1P-AUCP-F1PRO
AnoStyler (2026)90.7 ± 0.593.4 ± 0.196.6 ± 0.162.6 ± 0.290.1 ± 0.1
Anomagic (2026)90.1 ± 0.393.0 ± 0.296.7 ± 0.062.6 ± 0.091.3 ± 0.2
Ours98.2 ± 0.196.9 ± 0.297.8 ± 0.165.9 ± 0.393.7 ± 0.2

Table 15. Per-category segmentation results on MVTec AD under the FAST protocol.

CategoryAnomalyDiffusion mIoUAnomalyDiffusion AccFAST mIoUFAST AccOurs mIoUOurs Acc
bottle76.3983.5486.8690.9087.1595.69
cable62.5174.4873.7177.9486.8897.35
capsule37.7344.7263.2271.1287.6798.83
carpet64.6773.5973.8483.5377.9697.01
grid38.7051.8252.4570.7086.5295.48
hazel_nut59.3367.4890.8194.7991.3898.10
leather56.4562.5166.6074.1884.4498.54
metal_nut88.0091.1094.6596.8884.7896.20
pill83.2189.0090.1794.0789.3097.47
screw38.4749.4949.9457.4888.3298.77
tile84.2989.7290.1393.7780.7493.17
toothbrush48.6864.4174.9888.6390.3995.49
transistor79.2791.7491.8094.5088.0597.47
wood60.1674.6278.7786.3185.4196.75
zipper65.3672.6672.8084.7388.2298.14
Average62.8872.0676.7283.9786.4896.96

Q4. FAST comparison and generation efficiency.

Under FAST’s downstream segmentation protocol, AnomalyLoop achieved substantially higher mean IoU and accuracy than FAST on MVTec AD (Table 15). These results demonstrate AnomalyLoop’s advantage over FAST in downstream anomaly segmentation. See Tables 7 and 8 for comparisons with other recent methods in terms of generation efficiency and downstream detection performance.

Thank you for the additional experiments and clarifications, which have addressed most of my concerns, and I will raise my rating to 3.

However, I would encourage the authors to include downstream results using anomalies generated from multiple distinct normal images per category, rather than relying solely on the one-shot setting in the final version. Since normal images are generally abundant in industrial applications, such an evaluation would provide stronger evidence of the method's generalizability.

We sincerely thank you for your positive feedback on our additional experiments and clarifications, as well as for suggesting the use of multiple distinct normal images per category as references for anomaly generation. Following your suggestion, we have added a full-shot experiment on MVTec AD to Section 4.7 of the revised manuscript. For each category, we use all normal training images as references to generate anomalous samples and report the downstream detection results for that category. Given the space constraints in the main text, we report the 2-shot downstream detection results in Appendix G, using two distinct normal reference images per category.

Reviewer_TC9YOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes AnomalyLoop, a closed-loop framework that uses anomaly detection feedback to refine anomaly generation. It combines the zero-shot and open-set generalization capabilities with the pixel-level mask controllability and fidelity to real-world industrial anomalies. AnomalyLoop consists of two stages: an Anomaly Specific Region Exploration (ASRE) mechanism for constructing fine-grained candidate anomaly formation regions, and Generation-Detection Closed-Loop for refining anomaly generation. Experiments on MVTec AD and VisA show the effectiveness of the proposed framework.

Critical strengths
  1. The motivation and overall framework design are clear and well-justified. The generation–detection closed loop is well motivated and aligns with the goal of generating anomalies to improve downstream detection.
  2. AnomalyLoop combines zero-shot and open-set generalization capabilities with pixel-level mask controllability and fidelity.
  3. The experimental results demonstrate AnomalyLoop’s strong performance in anomaly generation and its effectiveness in improving downstream anomaly detection.
Critical weaknesses
  1. On the core of AnomalyLoop: “Generation-Detection Closed-Loop”, Table 3 compares anomaly generation and downstream detection performance with and without detection feedback, while Figure 10 in Appendix E.4 analyzes the effects of closed-loop iterations and control strength. However, it remains unclear how the choice of the frozen detector in the feedback loop affects AnomalyLoop’s performance.
  2. The paper provides a qualitative analysis of how the number of closed-loop iterations affects generation quality (Figure 10 in Appendix E.4), but lacks a quantitative analysis of its effects on generation quality and downstream detection performance.
  3. The paper contains several inconsistencies in notation and capitalization. For example, Table 1 and Figure 3 use “Ano-Diff” and “Ano-Any,” while Table 2 uses “AnoDiff” and “AnoAny”; these names should be standardized. In line 82, “closed-Loop” should be “closed-loop.” In line 431, “On” following the semicolon should be “on”.
Questions for authors
  1. The full method achieves IS / IC-L scores of 2.44 / 0.41 in Table 1, but 1.2631 / 0.2211 in Table 3. What accounts for this discrepancy in generation metrics between the two experiments?
  2. The main experiments in Table 2 use an AnomalyGPT-based detection framework. Why do the ablation experiments in Table 3 use a unified lightweight U-Net instead of the same framework?
  3. Why does Table 3 not report the individual contributions of L_sem and L_div?
  4. Figure 10 shows that increasing the number of iterations makes the generated anomalies appear more natural, but the default setting of two iterations is not included in the comparison. Why were two iterations chosen, and how was the trade-off between generation quality and computational cost assessed?
5 comments

We thank the reviewer for recognizing the motivation behind our method, the design of our framework, and the strength of our experimental results.

Table 2. Sensitivity to detector feedback on MVTec AD.

MethodI-AUC (%)I-F1 (%)P-AUC (%)P-F1 (%)PRO (%)Infer (s/image)GPU (GB)
DINOv2 Visual NN Feedback(2024)91.5 ± 0.193.6 ± 0.196.9 ± 0.063.6 ± 0.191.3 ± 0.04317
Ours98.2 ± 0.196.9 ± 0.297.8 ± 0.165.9 ± 0.393.7 ± 0.23916

W1. Sensitivity to detector feedback.

Since SubspaceAD is currently the latest anomaly detector for the open-set training-free setting, we select SubspaceAD by default as the frozen detector. To further analyze the impact of different detectors, we retained TOC and the frozen DINOv2 backbone, changing only the detector feedback mechanism. DINOv2 Visual NN Feedback measures the feature distance between each generated patch and its nearest normal reference patch, whereas SubspaceAD measures anomalies using the projection residual of generated features onto a normal subspace. This subspace is spanned by the principal components of normal reference patch features, capturing the dominant directions of variation in normal features. Table 2 shows that replacing SubspaceAD with Visual NN reduced downstream detection performance and increased generation time from 39 to 43 seconds per image. We hypothesize that, with limited normal references, a normal patch may differ from its nearest reference while its features remain largely explained by the normal subspace. SubspaceAD’s projection-residual scoring may therefore be less sensitive to such normal appearance variations than nearest-neighbor scoring, potentially providing more effective guidance for TOC.

Q1. Evaluation protocols for Tables 1 and 3 of our paper.

Tables 1 and 3 of our paper differ in both the number of generated images used for evaluation and how those images are matched to real test images for IC-L. Table 1 uses 73,000 generated images, with 1,000 for each of the 73 defect subtypes across 15 object categories. For IC-L, each set of 1,000 is compared only with real test images of the same object category and defect subtype. Table 3 uses 1,500 generated images, with 100 per object category. These 100 images cover all defect subtypes in that category. For IC-L, we compare the same set of 100 generated images separately with the real test images of each subtype within that category.

Table 10. AnomalyLoop ablations on MVTec AD with an AnomalyGPT-based detector.

VariantRandom maskStage 1LsubL_{\mathrm{sub}}LsemL_{\mathrm{sem}}LdivL_{\mathrm{div}}I-AUCI-F1
1–✓–––98.1 ± 0.196.8 ± 0.1
2–✓✓––98.1 ± 0.096.8 ± 0.1
3–✓✓✓–98.1 ± 0.097.0 ± 0.3
4–✓✓–✓98.1 ± 0.196.8 ± 0.2
5✓–✓✓✓98.0 ± 0.096.9 ± 0.0
6–✓✓✓✓98.2 ± 0.196.9 ± 0.2

Table 11. AnomalyAny ablations on MVTec AD with an AnomalyGPT-based detector.

VariantAttention-guided optimizationPrompt-guided optimizationI-AUCI-F1
1–✓85.992.4
2✓–86.492.4
AnomalyAny✓✓85.892.4

Q2. AnomalyGPT-based detection framework and lightweight U-Net.

Our main experiments use the same AnomalyGPT-based detection framework as AnomalyAny for comparability. We evaluated the AnomalyLoop variants (Table 10) and reproduced the AnomalyAny ablations (Table 11) using this framework. In both experiments, image-level detection scores varied little across variants. For AnomalyLoop, all variants achieved high scores with only 100 synthetic anomaly images per category. We suspect that the AnomalyGPT-based detector, which uses CLIP as a frozen backbone, can learn to recognize features that distinguish anomalies after fine-tuning on only a few examples. It may therefore be relatively insensitive to differences in the training samples generated by the AnomalyLoop ablation variants, resulting in similar detection scores across variants. To distinguish the contributions of individual modules to downstream detection, we switched to a lightweight U-Net for the ablation study in Table 3 of our paper and retrained a separate detector for each variant under the same protocol.

Table 3. Controlled ablation results on MVTec AD.

VariantRandom maskStage 1LsubL_{\mathrm{sub}}LsemL_{\mathrm{sem}}LdivL_{\mathrm{div}}ISIC-LI-AUCI-F1I-APM-act
ASRE–✓–––1.21860.186370.1985.4587.6183.65
ASRE + DF–✓✓––1.26560.214775.4988.3587.8988.58
w/o LdivL_{\mathrm{div}}–✓✓✓–1.26470.204973.4786.7086.9686.67
w/o LsemL_{\mathrm{sem}}–✓✓–✓1.26510.219074.0287.0786.7783.85
w/o ASRE✓–✓✓✓1.24450.159274.7788.1688.21–
Ours–✓✓✓✓1.26310.221176.0288.1288.2689.67

Q3. Individual contributions of LsemL_{\mathrm{sem}} and LdivL_{\mathrm{div}}.

The two new variants in Table 3 remove one constraint at a time and leave all other settings unchanged. Removing either constraint reduces downstream detection performance and M-act relative to the full model, demonstrating the contribution of each constraint.

Table 12. Generation time for different numbers of closed-loop iterations on a single RTX A6000.

Number of closed-loop iterations0123456
Inference time (s/image)3.221.239.357.675.093.52109.9
Estimated time for 73,000 MVTec AD images (h)64.9429.9796.91168.01520.81896.42228.5

Q4. Two iterations: trade-off between generation quality and computational cost.

In Figure 10, the generated anomalies appear more natural as the number of closed-loop iterations increases. Table 12 reports the generation time per image on a single RTX A6000 and the estimated total time for generating all 73,000 images. Each additional iteration increases the total generation time for the full set. Therefore, the default choice of two iterations is motivated by the trade-off between generation quality and computational cost.

W3. Notation and capitalization.

We standardized the baseline names in Tables 1 and 2 and Figure 3 and corrected the capitalization of “closed-loop” in line 82 and “on” in line 431.

We have incorporated all relevant revisions and additional experimental results into the revised manuscript.

We sincerely thank you again for recognizing the value of our work and for the time and effort you devoted to reviewing our manuscript. We have carefully addressed the concerns and questions you raised, conducted additional experiments, and provided detailed analyses of the results. We hope that these additional experiments and analyses help address your concerns regarding our method. We have also incorporated the relevant experimental results and discussions into the appendix of the revised manuscript. If any of your concerns remain insufficiently addressed, or if you have further questions, we would welcome the opportunity to discuss them with you and provide timely clarifications.

Reviewer_YvmGOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents AnomalyLoop, a training-free framework for open-set industrial anomaly generation that integrates anomaly detection feedback into the generation process. The framework first identifies fine-grained candidate anomaly regions through Anomaly Specific Region Exploration (ASRE), which decouples normal-conditioned responses and aggregates spatial evidence across diffusion timesteps. It then employs a frozen SubspaceAD detector to provide terminal feedback and applies Trajectory Optimal Control (TOC) to optimize cross-attention controls throughout the diffusion trajectory. Without requiring real anomaly samples or updating diffusion model parameters, AnomalyLoop generates synthetic anomaly image–mask pairs from normal images and text descriptions.

Critical strengths
  1. Meaningful detection-guided generation. The method uses anomaly detection feedback to update attention control variables across diffusion timesteps, strengthening anomalies that are initially insufficiently visible. Rather than merely filtering generated samples, it actively refines the generation process, which has clear practical value.
  2. The reported results demonstrate strong downstream anomaly detection performance on MVTec-AD and VisA. Moreover, the downstream evaluation uses an AnomalyGPT-based framework rather than the SubspaceAD evaluator employed during generation, partially mitigating concerns about relying on the same detector for both generation guidance and evaluation.
Critical weaknesses
  1. Insufficient validation of the independent contributions of the core designs. All quantitative variants in Table 3 retain ASRE, without separately evaluating the contributions of normal-response decoupling and cross-timestep region association. Furthermore, the paper lacks comparisons between TOC and simpler alternatives, such as gradient guidance or candidate selection, under matched computational budgets, and does not report end-to-end generation time. Consequently, the current experiments cannot clearly distinguish improvements attributable to the specific trajectory optimization design from those arising from additional inference computation, limiting the assessment of its necessity and practical value.
  2. Insufficient experimental coverage and comparative evaluation. Although the paper follows AnomalyAny’s 1-shot detection protocol, downstream evaluation is restricted to the 1-shot setting, leaving it unclear whether the reported gains persist under 2-shot and full-shot settings with more normal training data. Furthermore, the paper lacks a unified comparison with Anomagic and AnoStyler regarding downstream performance and the total computational cost of producing the same number of synthetic samples, including training where applicable.
Questions for authors
  1. Independent contribution of ASRE. Could the authors separately ablate normal-response decoupling and region-level temporal aggregation, while keeping the area budget and subsequent generation settings unchanged? A comparison with simpler candidate-region selection strategies would help establish whether the benefits arise from the proposed exploration mechanism rather than spatial confinement alone.
  2. The evaluation follows AnomalyAny’s 1-shot detection protocol, but Table 2 does not report AnomalyDiffusion results on VisA. Could the authors clarify whether the baseline results were reproduced using the same data and downstream training protocol or taken directly from the original papers? Providing the missing VisA results under a unified protocol would enable a more complete and fair comparison.
  3. Computational efficiency and comparison with other methods. Although AnomalyLoop is training-free, it involves extensive inference-time optimization. How does it compare with other open-set anomaly generation methods, such as Anomagic[1] and AnoStyler, in terms of computational cost and downstream performance? Specifically, under AnoStyler’s[2] evaluation protocol, does AnomalyLoop offer performance improvements while requiring less total generation time than the combined training and inference time of these methods? Please compare the time required to generate the same number of samples.
  4. AnomalyAny also reports results under 2-shot and full-shot settings, whereas the present paper does not include corresponding experiments. Could the authors provide results under these settings to establish whether the observed benefits extend beyond the 1-shot setting and strengthen the empirical evaluation?

[1] Jiang, Yuxin, et al. "Anomagic: Crossmodal prompt-driven zero-shot anomaly generation." Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 40. No. 7. 2026.

[2] So, Yulim, and Seokho Kang. "Anostyler: Text-driven localized anomaly generation via lightweight style transfer." Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 40. No. 18. 2026.

4 comments

We thank the reviewer for recognizing the practical value of detection-guided generation and for the positive assessment of our method’s strong downstream detection performance and our evaluation protocol.

Table 1. Comparison of TOC with simpler alternatives on MVTec AD.

MethodI-AUC (%) ↑I-F1 (%) ↑P-AUC (%) ↑P-F1 (%) ↑PRO (%) ↑Time (s/img) ↓GPU (GB) ↓
Gradient Guidance91.3 ± 0.293.6 ± 0.197.0 ± 0.163.6 ± 0.191.4 ± 0.0727
Candidate Selection90.8 ± 0.693.0 ± 0.397.0 ± 0.163.4 ± 0.191.3 ± 0.21210
Ours (TOC)98.2 ± 0.196.9 ± 0.297.8 ± 0.165.9 ± 0.393.7 ± 0.23916

W1. Comparison with simpler alternatives to TOC.

We compared TOC, gradient guidance, and candidate selection across five random seeds, using the same support mask for all three methods under each seed. Gradient guidance retains the original loss functions but directly updates the diffusion latent via loss gradients at selected denoising steps. Candidate selection generates multiple independent candidates without TOC and selects the one with the lowest final loss. Table 1 shows that TOC achieved the best overall performance. We believe this advantage may reflect how the methods use feedback: TOC uses feedback from the final image to jointly adjust attention controls across multiple denoising steps; gradient guidance updates only the current latent at selected denoising steps, while candidate selection picks the best image after generation.

Table 3. Controlled ablation results on MVTec AD.

VariantRandom MaskStage 1L_subL_semL_divIS ↑IC-L ↑I-AUC (%) ↑I-F1 (%) ↑I-AP (%) ↑M-act (%) ↑
ASRE-✓---1.21860.186370.1985.4587.6183.65
ASRE + DF-✓✓--1.26560.214775.4988.3587.8988.58
w/o L_div-✓✓✓-1.26470.204973.4786.7086.9686.67
w/o L_sem-✓✓-✓1.26510.219074.0287.0786.7783.85
w/o ASRE✓-✓✓✓1.24450.159274.7788.1688.21-
Ours-✓✓✓✓1.26310.221176.0288.1288.2689.67

Q1.Contribution of ASRE to candidate-region selection.

The w/o ASRE variant in Table 3 replaces the region selected by ASRE with a random mask sampled within the object mask. We keep the area budget and subsequent generation settings unchanged. We retrained the downstream U-Net for each variant under the same protocol and updated Table 3 in the manuscript accordingly. This ablation demonstrates the effectiveness of the ASRE component, which combines normal-response decoupling with region-level temporal aggregation.

Table 6. One-shot downstream detection results for AnomalyDiffusion on VisA.

MethodI-AUC (%)I-F1 (%)P-AUC (%)P-F1 (%)PRO (%)
AnomalyDiffusion(2024)77.0 ± 0.780.0 ± 0.297.4 ± 0.141.3 ± 0.486.8 ± 0.1
Ours95.7 ± 0.291.9 ± 0.198.6 ± 0.041.3 ± 0.191.9 ± 0.1

Q2. Source of the baseline results and missing VisA results.

The baseline results in Table 2 of our paper are taken from Table 2 of AnomalyAny. We evaluated AnomalyDiffusion under the same 1-shot detection protocol and report the previously missing downstream detection results on VisA in Table 6.

Table 7. Inference time per generated image, total training time, and GPU memory usage.

MethodTotal training time (h)Training GPU memory (GB)Inference time (s/image)Inference GPU memory (GB)
AnomalyDiffusion (2024)158191413
FAST (2025)9182948
Anomagic (2026)69482635
AnoStyler (2026)0–155
AnomalyAny (2025)0–7225
Ours0–3916

The total training times in Table 7 are estimated rather than measured from complete training runs. On a single NVIDIA RTX A6000 48GB GPU, we measured the training time on a small subset of training samples using a consistent batch size and extrapolated the total time based on each method's full training configuration and training scale, to enable a comparison of training costs under the same hardware conditions.

Table 8. One-shot detection on MVTec AD with an AnomalyGPT-based detector.

MethodI-AUCI-F1P-AUCP-F1PRO
AnoStyler (2026)90.7 ± 0.593.4 ± 0.196.6 ± 0.162.6 ± 0.290.1 ± 0.1
Anomagic (2026)90.1 ± 0.393.0 ± 0.296.7 ± 0.062.6 ± 0.091.3 ± 0.2
Ours98.2 ± 0.196.9 ± 0.297.8 ± 0.165.9 ± 0.393.7 ± 0.2

Q3. Computational cost and downstream performance.

We reproduced AnoStyler and Anomagic and compared all three methods under a unified downstream detection protocol. Both baselines take less time per image than AnomalyLoop (Table 7), but AnomalyLoop performs better on downstream detection (Table 8). In addition, Anomagic requires extra training for its generator and uses real anomaly images from external datasets as references at inference time. This differs from our open-set, training-free setting. Compared with AnomalyAny, AnomalyLoop achieves better generation quality, diversity, and downstream detection performance while generating images nearly twice as fast (39 vs. 72 seconds per image).

Table 9. 2-shot and full-shot detection on MVTec AD with an AnomalyGPT-based detector.

MethodI-AUCI-F1P-AUCP-F1PRO
AnomalyAny (2025), 2-shot95.8 ± 0.195.2 ± 0.296.0 ± 0.258.8 ± 0.292.6 ± 0.1
Ours, 2-shot98.3 ± 0.197.4 ± 0.298.0 ± 0.065.3 ± 0.193.9 ± 0.1
AnomalyAny (2025), full-shot98.496.997.465.194.7
Ours, full-shot98.897.998.165.394.7
Q4. Benefits in the 2-shot and full-shot settings.

Table 9 reports 2-shot and full-shot detection results averaged across MVTec AD categories. The per-category full-shot results are presented in Table 5 of the main paper, while the per-category 2-shot results are provided in Appendix G. AnomalyLoop’s overall advantage over AnomalyAny persists in both the 2-shot and full-shot settings.

We have incorporated all relevant revisions and additional experimental results into the revised manuscript.

Thank you for the additional experiments and clarifications, which have addressed most of my concerns. I think that bridging anomaly generation with downstream detection feedback is a meaningful direction for anomaly generation. Therefore, I have decided to raise my rating to 3.

We sincerely thank you for your recognition and encouragement of our work. Following your suggestion, we have incorporated the corresponding experimental results and discussions into the revised manuscript, which help to more clearly demonstrate the specific contributions of trajectory optimization and detection-guided generation.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

General Response

We sincerely thank the Area Chair for coordinating the review process and the three reviewers for their careful evaluation and constructive feedback.

We are encouraged by the reviewers’ positive assessments of the motivation, methodological design, and potential impact of our work. Reviewer TC9Y explicitly praised the motivation and overall framework as “clear and well-justified.” Reviewer SaZp recognized the generation–detection closed-loop framework as “an interesting approach” and highlighted its “promising improvements in generation quality and downstream anomaly detection.” During the discussion, Reviewer YvmG further recognized the integration of anomaly generation with downstream detection feedback as “a meaningful direction for anomaly generation.” Collectively, these comments affirm the significance of our research motivation, the appeal of the proposed generation–detection paradigm, and its potential to advance open-set anomaly generation and downstream anomaly detection.

We have responded to the reviewers’ questions point by point and incorporated the corresponding experiments and clarifications into the revised manuscript. Reviewers YvmG and SaZp indicated that the additional experiments and explanations addressed most of their concerns, and both raised their ratings from 2 to 3. Reviewer TC9Y noted that our responses addressed some of their concerns and further strengthened their positive view of the paper. In the order YvmG, TC9Y, and SaZp, the ratings changed from (2, 3, 2) to (3, 3, 3).

The reviewers’ suggestions led to four groups of additional experiments and analyses, which we would particularly like to highlight:

  1. Controlled validation of ASRE and the auxiliary constraints. Under the same area budget and generation settings, we replaced the candidate regions selected by ASRE with random masks and separately removed the semantic and diversity constraints. These experiments provide a more controlled assessment of ASRE beyond simple spatial confinement and clarify the individual contributions of the two auxiliary constraints to generation quality and downstream detection.

  2. Downstream detection beyond the one-shot setting. On MVTec AD, AnomalyLoop achieves better overall downstream detection performance than AnomalyAny in both 2-shot and full-shot settings. We also added a sensitivity analysis of normal-reference selection and a comparison between joint and per-category detector training.

  3. Analysis of TOC and detector feedback. Across five random seeds with the same support mask per seed, TOC achieved the highest scores on all five downstream metrics. Keeping TOC and the frozen DINOv2 backbone fixed, replacing SubspaceAD’s feedback with nearest-neighbor feedback reduced downstream performance and increased generation time from 39 to 43 seconds per image. We also report generation times for different numbers of closed-loop iterations.

  4. Downstream performance and computational cost. On a single RTX A6000, AnomalyLoop requires 39 s/image and 16 GB, versus AnomalyAny's 72 s and 25 GB. Although AnoStyler and Anomagic generate faster, AnomalyLoop outperforms both on all five metrics under the same one-shot detection protocol. Under FAST's segmentation protocol, AnomalyLoop achieves 86.5% mIoU versus FAST's 76.7%. We report measured inference time separately from estimated total generator training time.

Summary of revisions. The revisions listed below are marked in blue in the revised manuscript.

  • Section 4.4, Table 2, and Figure 5: Added the missing AnomalyDiffusion one-shot detection results and a qualitative comparison on VisA.
  • Section 4.5, Appendix A.3, and Table 3: Expanded the controlled ablations with a random-mask alternative to ASRE under the same area budget and separate ablations of the semantic and diversity constraints.
  • Sections 4.6–4.7 and Appendices G–H (Tables 4, 5, 9, and 10): Added experiments on sensitivity to normal reference selection, per-category full-shot and 2-shot results, and a comparison of joint versus per-category detector training.
  • Appendices E.4 and J (Tables 8 and 14): Added generation times for different numbers of closed-loop iterations, comparisons with two simpler alternatives to TOC, and an experiment replacing the detector feedback mechanism.
  • Appendix I (Tables 11–13): Added comparisons of training and inference costs, one-shot downstream detection comparisons with AnoStyler and Anomagic under a common protocol, and a segmentation comparison under FAST’s protocol.
  • Appendix K (Tables 15–16): Added AnomalyLoop and AnomalyAny ablations using the original AnomalyGPT-based detector. Its image-level scores varied little across variants, motivating our use of a lightweight U-Net for the component ablations.

We thank the Area Chair and the reviewers again for their time, careful consideration, and constructive feedback.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.