acceptodds
Under review as a conference paper at ICLR 2027

Hierarchical Prototype-Conditioned Masked Diffusion for Multi-View Anomaly Detection

Abstract

Multi-view anomaly detection leverages complementary observations to identify anomalies that may be occluded or weakly visible from a single viewpoint. However, existing methods often rely on dense patch-level cross-view interactions, which may introduce redundant and spatially inconsistent information. To address this limitation, we propose HPMD-AD, a Hierarchical Prototype-Conditioned Masked Diffusion model for multi-view Anomaly Detection, which models complementary cross-view normal patterns through compact hierarchical prototypes to guide normal feature reconstruction. Specifically, we progressively abstract features from each view into input-dependent view-level prototypes and aggregate them into sample-level prototypes, capturing view-specific normal patterns and cross-view shared structures without explicit spatial alignment. These hierarchical prototypes jointly guide diffusion-based feature reconstruction, while prototype masking and feature masking encourage complementary cross-view reasoning and mitigate identity shortcuts during reconstruction. Furthermore, we introduce a prototype-prior adaptive fusion strategy that jointly considers inter-view score disagreement and prototype reliability for robust multi-view score aggregation. Extensive experiments on Real-IAD and the substantially larger Real-IAD Variety benchmark demonstrate that HPMD-AD achieves state-of-the-art performance, attaining sample-level AUROC scores of 94.6% and 96.5%, respectively.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Multi-view anomaly detection, Prototype learning, Diffusion models

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_51QHOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work presents HPMD-AD, a multi-view industrial anomaly detection framework that aggregates cross-view information through hierarchical dynamic prototype aggregation, performs anomaly detection and localization by coupling dual masking with prototype-conditioned diffusion-based reconstruction, and adaptively fuses view-specific anomaly scores. Without requiring explicit spatial alignment, it delivers competitive results on the Real-IAD and Real-IAD Variety datasets.

Critical strengths

This work adopts hierarchical dynamic prototypes to encode both view-specific information and cross-view shared structures. By obviating the need for explicit spatial alignment, the method is highly suitable for industrial inspection scenarios marked by significant viewpoint variations. Experiments conducted on two large-scale multi-view datasets yield competitive performance across object-level detection, single-view detection, and pixel-level localization tasks. Furthermore, ablation studies and contextual experiments — evaluating factors such as viewpoints and random noise — offer compelling evidence supporting the effectiveness of the key design elements and the successful exploitation of cross-view information.

Critical weaknesses

(1) Prototypes are generated directly from the features of the current input, and the training objective is limited exclusively to feature reconstruction loss. During inference, if the input contains defects, such defects may propagate into view-level prototypes and further affect sample-level prototypes. The paper does not include an explicit normal prototype bank, normality constraints, or mechanisms to suppress anomaly information. Furthermore, prototype masking is only applied at the sample-level aggregation stage; the original view-level prototypes remain intact and serve as reconstruction conditions for their respective views. Consequently, even if a local feature is masked, information from that region can still enter the reconstruction network via prototype conditioning. While dual masking may mitigate shortcut learning, no evidence is provided to confirm that it fully eliminates the anomaly reconstruction shortcut.

(2) PPAFS equates cross-view consistency with reliability, a strategy that carries the risk of diluting genuine defect signals originating from single views.

(3) The prototypes are essentially compressed representations derived from learnable queries via cross-attention. The paper fails to establish whether distinct prototypes assume stable functional roles, whether redundancy exists across prototypes, or whether sample-level prototypes genuinely capture structures shared across views. Moreover, the “remove prototype conditioning” experiment presented in Table 3 simultaneously reduces both information input and model capacity, which confounds the analysis and makes it impossible to isolate the specific contribution of the prototype structure itself.

(4) All key variants in Table 3 retain diffusion-based reconstruction; accordingly, these experiments only demonstrate the utility of prototypes and masking within a diffusion framework. They do not address whether comparable results can be obtained by pairing the same hierarchical prototype approach with a standard Transformer-based reconstructor. Existing methods such as Dinomaly have already achieved strong performance using simple Transformers, noise bottlenecks, and reconstruction-based designs. The authors should compare deterministic reconstruction, single-step denoising, and three-step diffusion under identical experimental conditions — reporting both performance and computational cost metrics — to clarify the specific benefits conferred by the diffusion module.

(5) Although the paper adopts only a three-step DDIM sampling scheme, the pipeline still requires processing five DINOv2 inputs and performing full reconstruction. Simply citing the use of an H100 (80 GB) GPU is insufficient to demonstrate deployment efficiency; additional metrics including object-level latency, VRAM consumption, parameter count, and inference throughput are also required.

Questions for authors

Clarification Questions:

(1) How does the method prevent anomalous features from contaminating the view-level and sample-level prototypes during inference? Since view-level prototypes remain unmasked and are used as reconstruction conditions, how does the dual-masking strategy prevent potential anomaly reconstruction shortcuts?

(2) How does PPAFS handle defects that appear in only one view? Could the cross-view consistency constraint suppress genuine view-specific anomalies, and is there evidence that the method preserves such defect signals?

(3) In Table 3, removing prototype conditioning also reduces the information available to the reconstructor and potentially changes its effective capacity. Is there a capacity-matched ablation that can isolate the contribution of the hierarchical prototype structure itself?

Suggestions for Further Evaluation:

(1) A comparison with deterministic Transformer reconstruction and single-step denoising under the same experimental settings would help clarify whether the diffusion module provides benefits beyond the hierarchical prototype design. Both detection performance and computational cost should be reported.

(2) Please report end-to-end inference latency, peak GPU memory usage, parameter count, and throughput to better assess the practical efficiency of the proposed method, particularly given its use of five DINOv2 inputs and diffusion-based reconstruction.

15 comments

We are grateful for your detailed review and the care with which you examined our work. Your questions encouraged us to look more closely at the reasoning behind our design choices. We have clarified these points in the revised manuscript. Summary of Revisions:

  • We clarify how hierarchical prototype modeling and dual-masked diffusion reconstruction mitigate anomalous information leakage and reconstruction shortcuts, supported by prototype–patch similarity visualizations. The corresponding analyses have been added to Appendix F (Figure 4).

  • We analyze the complementary functional roles of view-level and sample-level prototypes through per-prototype cross-attention statistics across DiT layers. The results and discussions have been added to Appendix G (Table 9).

  • We evaluate fusion reliability on a Real-IAD subset comprising normal samples and anomalous samples with defects visible in exactly one view. PPAFS outperforms both mean and max fusion in sample-level AUROC, AP, and F1-max. Among the anomalous samples, PPAFS assigns an average weight of 30.4% to the defective view, compared with 20.0% under mean fusion. We report these results in Appendix D (Table 7).

  • We conduct controlled comparisons of prototype configurations, showing that hierarchical prototype modeling improves sample-/image-level AUROC by 5.2/6.6 percentage points over flat cross-view prototypes. The results and analyses have been incorporated into Section 4.6 (Table 5).

  • We compare three reconstruction strategies under matched settings: deterministic Transformer reconstruction, one-step DDIM, and three-step DDIM. We report detection accuracy and inference costs in Appendix K.1 (Table 12).

  • We provide a comprehensive accuracy–efficiency trade-off analysis, reporting inference latency, throughput, peak GPU memory, and parameter counts for the full and lightweight configurations. These results are included in Section 4.5 (Table 3), with evaluation details in Appendices I and J.

All newly added and revised content is highlighted in blue in the revised PDF for ease of reference.

We provide detailed point-by-point responses and supporting experimental results below.

Q1. Prototype Contamination and Reconstruction Shortcuts

Our design mitigates anomalous information leakage and reconstruction shortcuts through two complementary mechanisms. First, VPs are generated from view-specific features via learnable-query cross-attention, while SPs are further aggregated from VPs across all five views, rather than directly from potentially anomalous spatial features. Trained exclusively on normal samples, these learnable queries encourage the aggregation of representative normal patterns. Although VPs remain unmasked to preserve view-specific structural information, they are compact representations rather than spatially aligned patch features, reducing the risk of direct feature-copying shortcuts. Second, dual-masked diffusion reconstruction further reduces reliance on local information: prototype masking (PM) regularizes SP conditioning, while spatial masking (SM) removes partial noisy spatial features, encouraging context-aware reconstruction rather than direct reproduction of local anomalous patterns.

As shown in Figure 4 of the revised manuscript, the similarity maps averaged over eight VPs exhibit stronger responses in normal regions and weaker responses in annotated anomalous regions across representative Real-IAD categories. This suggests that VPs predominantly capture normal patterns even from anomalous inputs. Since SPs are further aggregated from VPs across all five views, rather than directly from the original spatial features, this hierarchical structure further reduces the risk of anomalous information propagation. Together, these mechanisms and observations support that our method mitigates anomalous information leakage and reconstruction shortcuts.

To further investigate how VPs and SPs contribute to reconstruction, we conduct cross-attention analysis and cross-view context evaluation.

Each view is reconstructed using eight view prototypes derived from its own features and four sample prototypes aggregated from the prototypes of all five views. We analyze attention across all 12 DiT blocks, dividing each group's total attention by its number of tokens to obtain the average attention per prototype. The analysis covers all 30 Real-IAD categories, with 20 normal and 20 anomalous objects per category, totaling 1,200 objects and 6,000 views.

The results are as follows.

DiT blocksNormal patches: VP (%)Normal patches: SP (%)Anomalous regions: VP (%)Anomalous regions: SP (%)
1–48.368.288.048.93
5–86.6411.735.9413.13
9–127.959.107.3810.25
Average7.659.707.1110.78

The attention analysis reveals two key observations. First, anomalous regions consistently assign higher attention to sample prototypes (SPs) than normal regions, with average weights of 10.78% and 9.70%, respectively. This suggests that reconstruction of anomalous regions relies more heavily on cross-view contextual information. Second, attention shifts toward SPs in intermediate blocks. Specifically, in blocks 5–8, attention to SPs reaches 11.73% for normal patches and 13.13% for anomalous regions, compared with 6.64% and 5.94% for view prototypes (VPs), respectively. Although this difference narrows in later blocks, SPs continue to receive higher attention than VPs. These observations support the complementary roles of view-specific representations and cross-view contextual information during reconstruction.

The paired context experiment in Table 2 of the original manuscript further examines whether the aggregated cross-view information benefits reconstruction. With evaluated views and annotations held fixed, adding one, two, and three context views yields mAD gains of 0.46, 0.63, and 0.87 points, respectively. Together with the attention results, this supports the use of complementary information from other views through sample-prototype conditioning.

Q2. Preserving Single-View Anomaly Evidence in PPAFS

To investigate whether prototype-consistency weighting suppresses genuine view-specific anomalies, we construct an evaluation subset comprising normal samples and anomalous samples with defects visible in exactly one of the five views. The toy and zipper categories contain no such single-view anomalous samples and are therefore excluded, leaving 28 Real-IAD categories for evaluation. We compare PPAFS with mean and max fusion, reporting sample-level detection performance and the average fusion weight assigned to the anomalous view.

The results are as follows.

FusionSample AUROC (%)Sample AP (%)Sample F1-max (%)Anomalous-view weight (%)
Max84.872.671.058.3
Mean85.769.971.920.0
PPAFS86.773.873.630.4

As shown above, PPAFS assigns an average weight of 30.4% to the anomalous view, exceeding the 20.0% under mean fusion, indicating that prototype-consistency weighting does not systematically suppress anomalous-view contributions. Unlike mean fusion with fixed uniform weights or max fusion that favors the strongest response, PPAFS dynamically adjusts view contributions based on prototype reliability and cross-view disagreement, balancing single-view anomaly evidence with information from other views. Consequently, PPAFS achieves the best sample-level AUROC/AP/F1-max of 86.7%/73.8%/73.6%, outperforming both mean and max fusion. Although max fusion assigns a higher average weight (58.3%) to the anomalous view, its inferior detection performance suggests that indiscriminately emphasizing the strongest response may also amplify spurious high responses in normal samples, potentially increasing false positives. These results demonstrate that PPAFS preserves single-view anomaly evidence on average while achieving more reliable anomaly detection.

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

This paper proposes HPMD-AD, a two-level hierarchy of input-dependent prototypes to condition diffusion-based feature reconstruction for multi-view industrial anomaly detection. Dual masking is used during training, while reconstruction discrepancies and adaptive score fusion provide anomaly scores at inference. Experimental results show a sample-level AUROC of 94.6% on Real-IAD and 96.5% on Real-IAD Variety, together with component ablations, fusion comparisons, and paired cross-view context experiments that fix the evaluated views. (Tables 1–4; Appendix E)

Critical strengths

S1. View- and sample-level prototypes incorporate cross-view context without requiring explicit local spatial correspondence. In Table 1, sample-level AUROC reaches 94.6% on Real-IAD, 4.4 points above MVAD, and 96.5% on Real-IAD Variety, 4.0 points above the strongest baseline listed there, Dinomaly. Appendix Table 6 reports the same sample-level AUROC as MVEAD but a higher image-level AUROC (89.9% versus 87.8%). These results demonstrate practical value on the evaluated multi-view benchmarks (see Tables 1 and 6).

S2. Table 3 systematically evaluates view-level prototypes, sample-level prototypes, dual-masking mechanisms, and adaptive fusion. As these components are progressively introduced, sample-level AUROC increases from 90.8% for the baseline to 92.4%, 93.7%, 94.1%, and finally 94.6%. Meanwhile, image-level AUROC improves from 85.9% to 89.9%, and pixel-level AUPRO increases from 89.6% to 92.1%. These results demonstrate the positive contributions of the main components to the overall detection performance. Table 2 further evaluates the method under different view-availability settings and compares it with MVAD. With two additional context views, HPMD-AD achieves 91.9% sample-level AUROC and 80.1% mAD, outperforming MVAD's 88.5% and 77.3%, respectively. Consistent performance advantages are also observed under the other evaluated view configurations, further supporting the effectiveness of the proposed framework across different multi-view settings.

Critical weaknesses
  1. INP-Former (Luo et al., CVPR 2025) also extracts input-dependent prototypes to guide feature reconstruction and reports multi-class Real-IAD results. HPMD-AD adds cross-view hierarchical aggregation and dual-masked diffusion, but Section 2.2 does not clearly distinguish these mechanisms from the closest prior work, and Tables 1 and 6 do not discuss this relevant comparison. As the input-view settings and evaluation protocols may differ, a careful account of what is directly comparable, and which advantages arise from additional cross-view information, would clarify the contribution.

  2. Avoiding dense cross-view feature interaction is a central motivation, yet the paper does not report end-to-end inference latency, peak memory, or a corresponding complexity comparison. Since the system also performs multi-view feature extraction and diffusion reconstruction, a small number of prototype tokens does not by itself establish the computational cost of the complete pipeline. A concise matched-cost analysis would clarify the practical implications of this design.

Questions for authors

Q1. Compared with INP-Former, what distinctive modeling or detection capability is enabled by cross-view hierarchical conditioning? Please clarify the key differences in view inputs, preprocessing, and evaluation protocols, and discuss the Real-IAD results only where they are comparable.

Q2. Could the authors provide a brief comparison of end-to-end inference latency, peak memory, or complexity (e.g., against MVAD with matched hardware and inputs), including feature extraction and DDIM reconstruction, to characterize the compact design’s cost trade-off?

4 comments

We are grateful for your careful reading of our manuscript and for your constructive suggestions. Your feedback helped us present our contribution more clearly, and we have revised the manuscript accordingly.

Summary of Revisions:

  • We expand Section 2.2 (Prototype-Based Representation Learning) to clarify how HPMD-AD differs from INP-Former. INP-Former extracts prototypes from individual images, whereas HPMD-AD uses a hierarchy of view-level and sample-level prototypes to incorporate cross-view information into dual-masked diffusion reconstruction.

  • We provide a performance comparison with INP-Former on Real-IAD Variety (160 categories), where HPMD-AD outperforms INP-Former across all ten evaluation metrics, improving sample-/image-level AUROC by 4.7/4.6 percentage points and mAD by 1.81 points. The corresponding results and analyses are presented in Table 1 and the accompanying discussion.

  • We conduct an accuracy–efficiency trade-off analysis against MVAD, reporting inference latency, object throughput, peak GPU memory, and parameter counts, including backbone feature extraction and DDIM reconstruction. HPMD-AD-S reduces peak GPU memory by 43.3% while achieving comparable inference latency and superior detection accuracy. The corresponding experiments and analyses have been added to Section 4.5 (Table 3), with further details in Appendices I and J.

  • We compare hierarchical prototype interaction (HDPM) with dense cross-view attention under matched backbone features, inputs, and reconstruction settings, with dual masking disabled in both configurations. HDPM improves sample-, image-, and pixel-level AUROC by 10.5, 12.1, and 1.9 percentage points, respectively, reduces end-to-end inference latency by 43.5%, and increases object throughput by 77.0%, with comparable parameter counts and peak GPU memory usage. We report the results and analysis in Appendix K.2 (Table 13) and describe the inference efficiency evaluation protocol in Appendix J.

All newly added and revised content is highlighted in blue in the revised PDF for ease of reference.

We provide detailed point-by-point responses below.

Q1. Comparison with INP-Former

We thank the reviewer for highlighting INP-Former (Luo et al., CVPR 2025) as a closely related work. To address this concern, we have expanded the discussion in Section 2.2 (Related Work) to clarify the distinctions between INP-Former's single-image prototype learning and our hierarchical cross-view prototype modeling, dual-masked diffusion reconstruction, and prototype-guided adaptive fusion (PPAFS). We have also updated Table 1 and its accompanying discussion to provide a more explicit performance comparison with INP-Former.

(1) Hierarchical cross-view prototype modeling.

INP-Former dynamically extracts intrinsic normal prototypes (INPs) from individual images and employs an INP coherence loss to constrain their normal representations. Specifically, this loss minimizes the cosine distance between each query feature and its nearest INP, encouraging the extracted prototypes to represent normal patterns within a single image.

In contrast, HPMD-AD learns hierarchical dynamic prototypes without requiring an additional prototype coherence loss. Specifically, view-level prototypes (VPs) capture view-specific normal patterns, while sample-level prototypes (SPs) aggregate complementary normal information across different views of the same object.

This hierarchical design enables HPMD-AD to exploit cross-view complementary normal information unavailable to single-image prototype modeling, while preserving view-specific characteristics. Furthermore, both prototype levels jointly condition diffusion-based feature reconstruction, providing richer normal priors for distinguishing normal and anomalous features.

(2) Dual-Masked Prototype-Conditioned Diffusion (DMPCD).

Unlike INP-Former's prototype-guided feature reconstruction, HPMD-AD introduces DMPCD, which employs hierarchical cross-view prototypes as conditions for diffusion-based feature reconstruction and incorporates dual masking to encourage normal pattern recovery. During training, prototype masking masks the prototypes of one view before sample-level aggregation, encouraging the model to exploit complementary information from the remaining views. Spatial feature masking removes a subset of noisy feature tokens, encouraging the denoiser to recover normal features using hierarchical prototype conditions and spatial context. Together, these mechanisms promote complementary cross-view reasoning and mitigate reconstruction shortcuts, improving anomaly detection and localization.

In addition, PPAFS adaptively integrates anomaly evidence across views based on prototype reliability and cross-view disagreement, further distinguishing HPMD-AD from INP-Former's single-image reconstruction framework.

Overall, compared with INP-Former's single-image prototype-guided reconstruction, these comparisons highlight the distinctive capabilities of HPMD-AD in hierarchical cross-view prototype modeling and dual-masked conditional diffusion reconstruction.

The results on Real-IAD Variety (160 categories) are as follows.

Metric (%)INP-FormerHPMD-ADGain (points)
Sample AUROC91.896.5+4.7
Sample AP98.899.4+0.6
Sample F1-max96.798.2+1.5
Image AUROC87.592.1+4.6
Image AP97.698.4+0.8
Image F1-max94.796.1+1.4
Pixel AUROC93.694.3+0.7
Pixel AP46.447.7+1.3
Pixel F1-max47.949.3+1.4
Pixel AUPRO78.679.6+1.0
mAD83.3685.17+1.81

As shown above, HPMD-AD outperforms INP-Former across all ten evaluation metrics on Real-IAD Variety (160 categories). Specifically, HPMD-AD improves sample-level and image-level AUROC by 4.7 and 4.6 percentage points, respectively, while increasing the overall mAD from 83.36 to 85.17 (+1.81 points). These results provide further evidence supporting the effectiveness of hierarchical cross-view prototype modeling.

We have further clarified these methodological distinctions in Section 2.2 (Related Work) and expanded the quantitative comparisons and their interpretation in Table 1 and its accompanying discussion of the revised manuscript.

Q2. End-to-End Inference Efficiency and Cost Trade-off

Balancing detection accuracy and inference efficiency is particularly important for practical industrial anomaly detection, where deployment scenarios involve different accuracy requirements and computational constraints. To address these demands, we provide two variants: the full HPMD-AD for accuracy-critical applications and the lightweight HPMD-AD-S for efficiency-sensitive deployment. HPMD-AD-S employs DINOv2-S/14 with 224 × 224 inputs, whereas the full HPMD-AD uses DINOv2-B/14 with 448 × 448 inputs.

To evaluate their practical computational costs, we conduct an end-to-end accuracy–efficiency comparison with MVAD on a single NVIDIA H100 GPU, using a fixed batch size of 30 multi-view objects (150 images). The measured inference latency includes backbone feature extraction, reconstruction, anomaly map and score generation, and data transfer. Both HPMD-AD variants employ three DDIM denoising steps. We report the mean and standard deviation over 10 timed measurements.

The results are as follows.

DatasetMethodSample AUROC (%)Image AUROC (%)Pixel AUROC (%)Latency (ms/batch)Objects/sMemory (GiB)Parameters (M)
Real-IADMVAD90.286.697.9530.04 ± 9.6656.602.3117.80
Real-IADHPMD-AD-S93.486.798.2524.59 ± 1.7457.191.3142.11
Real-IADHPMD-AD94.689.999.13374.30 ± 27.528.8919.53119.99
Real-IAD VarietyMVAD71.369.286.5535.49 ± 11.5256.022.3117.80
Real-IAD VarietyHPMD-AD-S95.288.893.9525.92 ± 2.7357.041.3142.11
Real-IAD VarietyHPMD-AD96.592.194.33415.83 ± 33.008.7819.53119.99

As shown above, the full HPMD-AD achieves the highest detection accuracy on both benchmarks, at the cost of increased inference latency and memory consumption. Meanwhile, HPMD-AD-S outperforms MVAD in sample-/image-/pixel-level AUROC by 3.2/0.1/0.3 percentage points on Real-IAD, respectively. Notably, despite the substantially expanded category space of Real-IAD Variety, HPMD-AD-S maintains strong detection performance, surpassing MVAD by 23.9/19.6/7.4 points on the corresponding metrics.

In terms of efficiency, HPMD-AD-S achieves comparable inference latency and throughput to MVAD while reducing peak GPU memory by 43.3% (1.31 vs. 2.31 GiB). These results demonstrate the flexibility of HPMD-AD in balancing detection accuracy and inference efficiency across different deployment requirements.

The comparison is conducted under matched hardware, batch size, and number of views, while retaining method-specific backbone architectures and input resolutions. The corresponding analysis has been incorporated into Section 4.5 (Table 3) of the revised manuscript, with further implementation details and evaluation protocols provided in Appendices I and J.

W2. Computational Efficiency of Hierarchical Prototype Interaction

To verify that compact prototype modeling translates into practical end-to-end computational efficiency, we conduct a controlled comparison between hierarchical prototype interaction (HDPM) and dense cross-view attention under matched backbone features, input configurations, and diffusion reconstruction settings. To isolate the contribution of hierarchical prototype interaction, both configurations are evaluated without dual masking. Experiments are conducted on a single NVIDIA H100 PCIe GPU with 30 multi-view objects (150 images) per batch. The measured end-to-end inference latency includes backbone feature extraction, cross-view interaction, diffusion reconstruction, anomaly map and score generation, and data transfer. We report the mean and standard deviation over 10 timed measurements.

The results are as follows.

MethodSample AUROC (%)Image AUROC (%)Pixel AUROC (%)Latency (ms/batch, mean ± SD)Objects/sPeak memory (GiB)Parameters (M)
Dense cross-view attention83.277.097.16048.32 ± 126.414.9619.51120.25
HDPM93.789.199.03416.18 ± 13.558.7819.53119.99

As shown above, compared with dense cross-view attention, HDPM improves sample-/image-/pixel-level AUROC by 10.5/12.1/1.9 percentage points, respectively, while reducing end-to-end inference latency by 43.5% and increasing object throughput by 77.0%. With comparable parameter counts and peak GPU memory consumption, these results demonstrate that hierarchical prototype interaction achieves both higher detection accuracy and faster inference through more effective and efficient cross-view information modeling. The corresponding controlled comparison has been incorporated into Table 13 of the revised manuscript, with the detailed inference evaluation protocol provided in Appendix J.

We sincerely appreciate your careful review and constructive suggestions. The additional clarifications and experimental analyses strengthen our evaluation of HPMD-AD and provide a fuller account of its modeling capabilities, detection performance, and practical efficiency. If you have any further questions, we would be happy to discuss them with you during the discussion phase.

Reviewer_aW4MOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes HPMD-AD for unsupervised multi-class multi-view anomaly detection. The method compresses frozen DINOv2 features into input-dependent view-level prototypes and subsequently aggregates them into sample-level prototypes. These hierarchical representations condition diffusion-based feature reconstruction, with prototype masking and feature masking encouraging complementary cross-view reasoning and reducing reconstruction shortcuts. A prototype-prior adaptive fusion strategy combines view-wise anomaly scores using score disagreement and prototype consistency. Experiments on Real-IAD and Real-IAD Variety report sample-level AUROC scores of 94.6% and 96.5%, respectively. Component ablations and a paired evaluation under partial observations provide additional evidence for the effectiveness of the proposed framework.

Critical strengths
  1. Strong empirical results across two substantial benchmarks. HPMD-AD achieves competitive detection and localization performance on both the Real-IAD and Real-IAD Variety benchmark. The improvements across multiple evaluation levels make the contribution more convincing.
  2. Useful evidence that the model exploits cross-view information. Table 3 shows progressive improvements from view-level and sample-level prototype conditioning, followed by additional benefits from dual masking. The paired experiment in Table 2 keeps evaluated views and labels fixed while varying the available context. Together with the matched-noise protocol in Appendix E, this provides meaningful evidence that additional views improve reconstruction.
Critical weaknesses
  1. The central architectural advantage is not fully isolated. The ablations establish that prototype conditioning helps the proposed diffusion baseline, but they do not directly compare the hierarchy against a flat prototype representation or dense cross-view conditioning under matched features and model capacity. Although Appendix F includes additional strong methods, the reporting does not fully clarify comparability in backbone features and preprocessing. Consequently, the results support the effectiveness of the complete system more strongly than the specific claim that hierarchical prototypes provide a superior mechanism for cross-view interaction.
  2. The interpretation of prototype consistency as reliability needs stronger validation. PPAFS reduces the contribution of views whose prototype centers disagree with the cross-view consensus. However, a genuine defect visible in only one view may itself cause such disagreement, as the paper acknowledges. Table 4 supports the combined fusion rule empirically, but aggregate results do not establish when this weighting preserves or suppresses valid anomaly evidence.
Questions for authors
  1. Clarification and controlled comparison: Which baseline results use the same backbone features, input preprocessing, and data splits as HPMD-AD? Under an otherwise identical reconstruction setup, how does hierarchical conditioning compare with flat prototype conditioning or dense cross-view attention?
  2. Evidence concerning fusion reliability: For samples with defects visible in only one view, does prototype-consistency weighting reduce the contribution of the defective view? A comparison of PPAFS, mean fusion, and max fusion on this subset, together with the corresponding fusion weights, would materially clarify this concern.
  3. Statistical stability: How much do the principal ablation results vary across independent training runs and inference-noise realizations?
7 comments

We sincerely appreciate the time and care you devoted to reviewing our paper. Your thoughtful comments gave us a clearer sense of where the manuscript needed further explanation. We have revised those passages and strengthened the supporting analysis.

Summary of Revisions:

  • We clarify the backbone configurations and comparability of existing baselines on Real-IAD and Real-IAD Variety, distinguishing literature-reported results from strictly controlled comparisons. The corresponding clarifications have been incorporated into the discussions accompanying Table 1 in the main text and Table 10 in the appendix.

  • We conduct controlled comparisons with flat prototype conditioning, demonstrating that hierarchical prototype modeling improves sample-/image-level AUROC by 5.2/6.6 percentage points. These additional experiments and analyses have been incorporated into Section 4.6 (Table 5) of the revised manuscript.

  • We compare hierarchical prototype interaction (HDPM) with dense cross-view attention under matched settings, with dual masking disabled in both configurations. HDPM improves sample-, image-, and pixel-level AUROC by 10.5, 12.1, and 1.9 percentage points, respectively, and reduces inference latency by 43.5%. We report the results and analysis in Appendix K.2 (Table 13).

  • We evaluate fusion reliability on a subset comprising normal samples and anomalous samples with defects visible in only one view. PPAFS achieves higher sample-level AUROC, AP, and F1-max than both mean and max fusion. Among the anomalous samples, PPAFS assigns an average weight of 30.4% to the defective view, compared with 20.0% under mean fusion. We report the results in Appendix D (Table 7).

  • We evaluate statistical stability across independent training runs and inference-noise realizations, observing maximum standard deviations of only 0.25 and 0.09 percentage points, respectively. The detailed statistical results and analyses have been added to Appendix L (Table 14).

All newly added and revised content is highlighted in blue in the revised PDF for ease of reference.

We provide detailed point-by-point responses and supporting experimental results below.

Q1(Part 1/2). Clarification and Controlled Comparison

To address this concern, we clarify the comparability of existing baselines in terms of backbone architectures, input preprocessing, and evaluation protocols. We further conduct controlled comparisons with flat prototype conditioning and dense cross-view attention under matched experimental settings to isolate the contribution of hierarchical prototype modeling. These analyses provide empirical evidence for the effectiveness of hierarchical conditioning in terms of both detection accuracy and computational efficiency.

(1) Baseline comparability.

We clarify the backbone architectures of representative single-view and multi-view baselines under the multi-class setting, as shown in the following table.

DatasetMethodBackboneSample AUROC (%)Image AUROC (%)Pixel AUROC (%)
Real-IADUniADEfficientNet-B487.383.097.4
Real-IADViTADDINOv1-S88.882.897.1
Real-IADMVADResNet-3490.286.697.9
Real-IADDinomalyDINOv2-B–89.398.8
Real-IADMVAG-FLDINOv2-B–89.798.9
Real-IADMVEADDINOv2-B94.687.8–
Real-IADHPMD-ADDINOv2-B94.689.999.1
Real-IAD VarietyViTADDINOv1-S83.678.390.9
Real-IAD VarietyMVADResNet-3471.369.286.5
Real-IAD VarietyMambaADResNet-3485.079.090.2
Real-IAD VarietyEILADINOv2-B91.387.193.8
Real-IAD VarietyDinomalyDINOv2-B92.587.093.5
Real-IAD VarietyINP-FormerDINOv2-B91.887.593.6
Real-IAD VarietyHPMD-ADDINOv2-B96.592.194.3

Several representative baselines employ the same DINOv2-B backbone architecture as HPMD-AD, including Dinomaly, MVAG-FL, and MVEAD on Real-IAD, as well as EILA, Dinomaly, and INP-Former on Real-IAD Variety. All compared methods follow the same official training and testing splits defined in the JSON annotation files provided by Real-IAD and Real-IAD Variety, ensuring consistent data partitioning across methods. On Real-IAD Variety, HPMD-AD achieves 96.5% sample-level AUROC and 92.1% image-level AUROC, outperforming the strongest baselines on the respective metrics by 4.0 and 4.6 percentage points, respectively.

(2) Hierarchical vs. flat prototype conditioning.

To isolate the contribution of hierarchical prototype conditioning, we conduct controlled comparisons under identical backbone features, input preprocessing, data splits, and diffusion reconstruction settings, including the denoiser architecture, noise schedule, training protocol, and DDIM sampling steps. We vary only the prototype conditioning mechanism, comparing view-only prototypes with different numbers, flat prototypes extracted directly from the patch tokens of all five views, and our HDPM.

The results are as follows.

Prototype configurationSample AUROC (%)Sample AP (%)Sample F1-max (%)Image AUROC (%)Image AP (%)Image F1-max (%)Pixel AUROC (%)Pixel AP (%)Pixel F1-max (%)AUPRO (%)
No prototype90.895.590.585.983.276.498.335.640.389.6
View-only (4)92.096.091.487.585.178.198.635.140.489.3
View-only (8)92.496.291.688.085.678.898.835.740.689.8
View-only (12)92.596.291.888.185.879.098.835.640.690.0
Flat cross-view (12)88.594.289.182.580.073.697.932.237.985.5
HDPM93.796.992.689.186.280.299.035.640.690.9

We investigate the effects of prototype quantity and organization. Increasing the number of view-level prototypes from 0 to 4, 8, and 12 per view improves sample-level AUROC from 90.8% to 92.0%, 92.4%, and 92.5%, respectively, with similar trends at the image level. The diminishing gains beyond eight prototypes suggest that simply increasing the prototype count provides limited additional benefits. In contrast, directly extracting 12 flat cross-view prototypes from all five views yields only 88.5%/82.5% sample-/image-level AUROC. HDPM, employing 8 view-level prototypes per view and 4 sample-level prototypes, achieves 93.7%/89.1%, outperforming the best view-only configuration by 1.2/1.0 percentage points and the flat cross-view configuration by 5.2/6.6 points, respectively.

Q1(Part 2/2). Clarification and Controlled Comparison

(3) Hierarchical vs. dense cross-view attention.

We further compare hierarchical prototype interaction (HDPM) with dense cross-view attention under the same frozen backbone features, input preprocessing, data splits, and diffusion reconstruction settings. Both configurations are evaluated without our proposed dual-masking strategy to better isolate the contribution of hierarchical prototype interaction.

The results are as follows.

MethodSample AUROC (%)Image AUROC (%)Pixel AUROC (%)Latency (ms/batch, mean ± SD)Objects/sPeak memory (GiB)Parameters (M)
Dense cross-view attention83.277.097.16048.32 ± 126.414.9619.51120.25
HDPM93.789.199.03416.18 ± 13.558.7819.53119.99

HDPM improves sample-/image-/pixel-level AUROC over dense cross-view attention by 10.5/12.1/1.9 percentage points, respectively. With comparable parameter counts (119.99M vs. 120.25M), the HDPM-based configuration reduces end-to-end inference latency by 43.5% (3416.18 vs. 6048.32 ms/batch). Inference latency is averaged over 10 repeated measurements, with the detailed evaluation protocol provided in Appendix J. These results demonstrate that hierarchical prototype interaction achieves a more favorable accuracy–efficiency trade-off than dense cross-view attention.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.