acceptodds
Desk rejected at ICLR 2027

Meganomaly: Extending Unsupervised Anomaly Detection to Hundreds of Categories

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Recent multi-class unsupervised anomaly detection (MUAD) methods have nearly closed the gap with their single-class counterparts on small-scale benchmarks, yet suffer severe degradation when extended to hundreds of categories. This degradation is commonly attributed to over-generalization, where the reconstruction network learns to restore unseen anomalous patterns. In this work, we show that this explanation is incomplete. Through controlled scaling experiments, we identify __capacity saturation__ as a critical overlooked factor: as the number of categories grows, diverse normal patterns compete and interfere within the decoder's finite weights, causing normal reconstruction to saturate while anomaly reconstruction continues to improve. Guided by this insight, we present __Meganomaly__, a framework for large-scale MUAD based on a simple design principle: expanding decoder's capacity. Specifically, ShareMLP composes image-adaptive MLPs from a dynamic weight pool shared across decoder layers, substantially increasing capacity and reducing inter-category interference at little additional computational cost. To prevent anomalies from exploiting this expanded capacity, the Prior-Preserving Tight Bottleneck (PTB) aggressively compresses content information while preserving spatial and semantic priors that reconstruction relies on. Extensive experiments on seven datasets show that Meganomaly establishes state-of-the-art on large-scale benchmarks, improving I-AUROC by +5.5% on Real-IAD Variety (94.6%, 160 classes) and +7.1% on ADNet (89.1%, 380 classes).

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Anomaly Detection, Unsupervised Learning, Unified Model

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_zfFMOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper studies MUAD on the setting of hundreds of categories. A series of scaling experiments show that anomaly-region reconstruction error keeps falling while normal-region error plateaus or rises. The so-called "capacity saturation" complements the usual over-generalization story. To this end, the authors propose two components: ShareMLP, an image-level (CLS-routed) dynamic FFN whose weights are soft mixtures of a weight pool shared across all decoder layers, and the Prior-Preserving Tight Bottleneck (PTB), a 32-d noisy bottleneck that re-injects positional embeddings and passes the CLS token through a separate noise-free path. The proposed method reports 94.6% I-AUROC on Real-IAD Variety (+5.5 over Dinomaly2) and 89.1% on ADNet (+7.1), with gains on Real-IAD and MANTA and parity on MVTec-AD/VisA.

Critical strengths
  • The diagnosis is the paper's best contribution. Separating anomaly-side and normal-side reconstruction error under a fixed evaluation set is a clean experimental design. And the ablations are thorough: component ablation, static widening up to r=32, sparse MoE, CondConv/DY-Conv/KW-Conv, matched-parameter DynaMLP vs ShareMLP, PTB design, dropout, pool size, encoders, model scale, and plug-in to RD4AD/ViTAD/Dinomaly.
  • Large and consistent gains: 160- and 380-class benchmarks, across image-, object- and pixel-level metrics. The 160-class multi-class result (94.6) nearly matches Dinomaly2's one-model-per-class number (94.9).
  • Compute cost is well controlled. ShareMLP at 8× FFN parameters but ~1× FFN FLOPs beats static MLP at 8× FLOPs (92.00 vs 90.64), and total FLOPs rise only from 26.2G to 27.6G.
Critical weaknesses
  • Technical novelty is modest. Input-conditioned weight mixtures (CondConv, DY-Conv) and cross-layer weight templates (Savarese & Maire; KW) are established; the contribution is choosing image-level routing and sharing for UAD, plus a narrower bottleneck.
  • The "capacity saturation" evidence rests on one baseline, one dataset, one encoder size and normalized errors from single runs. The curves in Fig. 1(b) are fairly flat; it is not shown that the trend is beyond seed noise.
Questions for authors
  1. Were d_t, dropout, N and the stopping iteration selected on the Real-IAD Variety test split? If so, please report results with hyperparameters chosen on a held-out validation split, or on a different dataset and then transferred.
  2. Is it capacity or interference or optimization? Fig. 1(b) is consistent with several explanations: finite capacity, gradient interference between classes, or fewer effective updates per class. Appendix A.3 says models are trained "up to 200,000 iters"; was the per-class number of updates held fixed across C? A run of static MLP r=32 through the same 20→160 protocol would directly test whether pure width removes the normal-side saturation.
2 comments
AuthorsOpenReview ↗

Response to Reviewer zfFM (part1)

Thank you for the thoughtful and detailed review. We sincerely appreciate your recognition of our diagnosis, clean experimental design, thorough ablations, and consistent gains at well-controlled computational cost. Your comments have been very helpful in improving our paper, and we address each of them below.

Since OpenReview does not support images in comments, the revised and newly added figures (Figs. A1, A2, and A3) have been included in the updated paper PDF. We invite you to refer to it.


W1.

Technical novelty is modest. Input-conditioned weight mixtures (CondConv, DY-Conv) and cross-layer weight templates ( Savarese & Maire; KW) are established; the contribution is choosing image-level routing and sharing for UAD, plus a narrower bottleneck.

Answer:

Thank you for this comment and for recognizing the value of our diagnosis. As you noted, our primary contribution is identifying capacity saturation as a key failure mode of large-scale MUAD. On the technical side, we do not aim to invent an entirely new mechanism; rather, we identify which design choices matter for alleviating capacity saturation and realize them in the simplest form. We have explored more complex alternatives extensively, including various dynamic weights, finer weight partitioning, and contrasting-driven routing, and found that they bring no clear improvement (Tables 4 and A4). This indicates that the current simple design is sufficient for effectively alleviating capacity saturation. In the revision, we will make this positioning explicit. We hope that our diagnosis and preliminary design can inspire future work on capacity saturation in large-scale UAD.


W2.

The "capacity saturation" evidence rests on one baseline, one dataset, one encoder size and normalized errors from single runs. The curves in Fig. 1(b) are fairly flat; it is not shown that the trend is beyond seed noise.

Answer:

Thank you for this valuable suggestion. It prompted us to extend the category-scaling diagnosis to a larger dataset, different baselines, and different encoders (new Fig. A2).

More datasets, baselines, and encoders. We repeat the scaling protocol on ADNet, scaling the training set from 40 to 380 categories while evaluating on a fixed 40-category test set, with three settings: Dinomaly2-S (ViT-S), Dinomaly2-B ( ViT-B), and RD4AD (CNN-based, WideResNet-50).

On ADNet, a more diverse and challenging benchmark spanning five domains and 380 categories, capacity saturation becomes even more pronounced:

  • Dinomaly2-S and Dinomaly2-B. For both encoder sizes, instead of merely plateauing, the normal-region errors at all levels (E_norm 1%/10%/50%) increase from the very beginning, while the anomaly-region error keeps decreasing, steadily shrinking the reconstruction gap and I-AUROC. For Dinomaly2-S, the hard-normal error eventually exceeds the anomaly-region error. Dinomaly2-B, with a larger decoder, narrows the gap more slowly but follows the same trend, consistent with capacity saturation. We will add Dinomaly2-S-ADNet to Fig. 1b in the revision.
  • RD4AD. For this earlier and relatively weaker method, the hard-normal error already exceeds the anomaly-region error at the very first setting and grows substantially as categories are added, accompanied by a steep decline in I-AUROC.

The trends on ADNet are large, far exceeding typical seed variation in this framework. Moreover, we are running five-seed category-scaling experiments (which is time-consuming) and will post the results during the rebuttal period if time permits.

AuthorsOpenReview ↗

Response to Reviewer zfFM (part2)

Q1.

Were d_t, dropout, N and the stopping iteration selected on the Real-IAD Variety test split? If so, please report results with hyperparameters chosen on a held-out validation split, or on a different dataset and then transferred.

Answer:

Thank you for this careful question. Most hyperparameters are inherited or set by default. Following your suggestion, we transferred the bottleneck d_t to a different dataset and report the results below.

  • Inherited without tuning. The dropout rate (0.4), training schedule (100k iterations), learning rate, betas, and weight decay directly follow Dinomaly2's settings on Real-IAD, without any adjustment. We report the final checkpoint at the end of the fixed schedule.

  • Not selected by performance. Performance increases monotonically with larger pool size N (Table A6), so N is not tuned for optimality. We set N=64 (8× FFN parameters) as a trade-off between accuracy and parameter count; N=96 performs better but requires more parameters.

  • Roughly set and robust. Since UAD benchmarks provide no validation split due to the unsupervised setting, d_t=32 was roughly set by observing results on Real-IAD Variety, without a fine-grained search. Performance is robust over a wide range of d_t, and we transfer the same setting to a different dataset, ADNet (Meganomaly-S). On ADNet, the variation across d_t from 16 to 48 is within 0.3% I-AUROC, and our default d_t=32 is not even the best setting.

d_tReal-IAD VarietyADNet
I-AUROCP-AUROCAUPROI-AUROCP-AUROCAUPRO
1693.995.383.685.494.082.7
32 (default)94.095.584.685.594.183.2
4893.895.484.685.794.183.4
6493.395.484.885.193.882.9

Q2.

Is it capacity or interference or optimization? Fig. 1(b) is consistent with several explanations: finite capacity, gradient interference between classes, or fewer effective updates per class. Appendix A.3 says models are trained "up to 200,000 iters"; was the per-class number of updates held fixed across C? A run of static MLP r=32 through the same 20→160 protocol would directly test whether pure width removes the normal-side saturation.

Answer:

Thank you for this insightful question and the concrete experimental suggestions. Following them, we ran the static MLP (r=32) through the same 20→160 protocol (revised Fig. A1) and added a control for per-class training budget (new Fig. A3).

Pure width alleviates the saturation. As shown in revised Fig. A1, widening the static MLP to r=32 clearly alleviates the normal-side saturation: compared with the baseline, the normal-region error keeps decreasing over more categories and reaches a lower level, and the I-AUROC drop from 20 to 160 categories shrinks from 7.4% to 4.1%. Since pure width changes nothing but capacity, this directly confirms the existence of capacity saturation.

Capacity and interference are, in our view, two sides of the same coin: interference arises because all categories compete within finite shared weights, as introduced in our Abstract and Sec. 3.1: "diverse normal patterns compete and interfere within the decoder's finite, shared weights." ShareMLP, at the same parameter budget, provides far more weight combinations and lets different categories route to different combinations, thereby both expanding capacity and reducing interference. Accordingly, it sustains the decrease in normal error more effectively.

Full convergence and proportional update budget. In our original setting, for each category configuration, we trained the model for up to 200k iterations (twice the default schedule), evaluated the E_norm every 10k iterations, and selected the checkpoint at which it stopped decreasing. This ensures that every configuration is compared at its fully converged state (no anomaly labels are used, only for diagnostic analysis).

As you suggested, in new Fig. A3, we further compare checkpoints at training budgets proportional to category count ( e.g., 20k iterations for 20 categories and 160k for 160 categories), so that the per-class number of updates is held fixed across C (although we consider the original converged setting more appropriate, as it compares each configuration at its best attainable reconstruction). The same saturation trend persists under this control.

These analysis and figures will be added to the final version.

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

This paper addresses performance degradation in large-scale multi-class unsupervised anomaly detection (MUAD) by proposing Meganomaly, a framework designed to mitigate decoder capacity saturation and over-generalization. Specifically, it introduces ShareMLP to expand effective decoder capacity through dynamic cross-layer weight sharing, and a Prior-Preserving Tight Bottleneck (PTB) to restrict anomaly reconstruction while preserving essential priors. Experiments on seven benchmarks demonstrate competitive performance, with substantial improvements on large-scale datasets containing up to 380 categories.

Critical strengths
  1. Relevant problem formulation and interesting empirical observations. The paper addresses an important practical limitation of unified anomaly detection when scaling to hundreds of categories. The analysis of normal reconstruction saturation provides an interesting complementary perspective to the conventional over-generalization explanation.

  2. Performance is good.

  3. Ablation study is thorough.

Critical weaknesses
  1. Limited methodological novelty. ShareMLP mainly combines existing dynamic weight composition and cross-layer parameter sharing techniques. In Table 3, ShareMLP only marginally improves over DynaMLP (92.00% vs. 91.66% I-AUROC). The authors should better clarify its methodological novelty and demonstrate the necessity of cross-layer sharing.

  2. Insufficient evidence for capacity saturation. The reconstruction-error trends in Section 3.1 do not conclusively establish capacity saturation as the primary cause of performance degradation. Other factors, such as training-data diversity and optimization dynamics, may contribute. More controlled experiments are needed to support this central claim.

  3. Parameter efficiency. Although the computational overhead is relatively small, Table A6 shows that the parameter count increases substantially from 36.3M to 102.7M. The authors should further justify this trade-off and discuss its implications for practical scalability and deployment.

  4. Missing important related work. The paper overlooks UniMMAD: Unified Multi-Modal and Multi-Class Anomaly Detection via MoE-Driven Feature Decompression (CVPR 2026), which also addresses cross-category interference through dynamic expert-based feature reconstruction. The authors should discuss this relevant work and better distinguish their methodological contributions from existing approaches.

Questions for authors

Please see the weaknesses part.

5 comments
AuthorsOpenReview ↗

Response to Reviewer xbtv (part1)

Thank you for recognizing the practical relevance of scaling unified anomaly detection to large categories, the interesting perspective offered by our analysis, the strong performance, and the thoroughness of our ablations. Your comments have been very helpful in improving our paper, and we address each of them below.

Since OpenReview does not support images in comments, the revised and newly added figures (Figs. A1, A2, and A3) have been included in the updated paper PDF. We invite you to refer to it.


W1.

Limited methodological novelty. ShareMLP mainly combines existing dynamic weight composition and cross-layer parameter sharing techniques. In Table 3, ShareMLP only marginally improves over DynaMLP (92.00% vs. 91.66% I-AUROC). The authors should better clarify its methodological novelty and demonstrate the necessity of cross-layer sharing.

Answer:

Thank you for this question, which prompted us to evaluate cross-layer sharing more thoroughly, including new experiments on ADNet. The modest gain in Table 3 occurs at a relatively generous 8× FFN budget for 160 categories. Cross-layer sharing provides larger improvements in more demanding settings.

Novelty. Our primary contribution is identifying capacity saturation as a key failure mode of large-scale MUAD. On the technical side, we do not aim to invent an entirely new mechanism; rather, we identify which design choices matter for large-scale UAD and realize them in the simplest form. Notably, more complex alternatives (DY-Conv, CondConv, KW-Conv, finer weight partitioning, contrasting-driven routing) either underperform or bring no consistent gains (Tables 4 and A4).

Necessity of cross-layer sharing. The gap in Table 3 appears small because, for the 160 categories of Real-IAD Variety, an 8× FFN budget is already ample for both designs. The advantage of sharing becomes clear in two more demanding regimes.

(1) Smaller parameter budgets. When the budget is limited, sharing gives each layer access to many more bases, and ShareMLP outperforms DynaMLP by a clear margin (Table 6):

Real-IAD Variety:

ParamsI-AUROCAUPRO
DynaMLPShareMLPΔDynaMLPShareMLPΔ
1×86.34*87.47+1.1379.21*80.02+0.81
2×88.2789.67+1.4080.8382.07+1.24
4×90.1890.81+0.6382.4883.10+0.62

*At 1×, DynaMLP reduces to the static MLP (K=1).

(2) More categories. On ADNet with 380 categories, where greater diversity demands more capacity, the advantage of ShareMLP grows even at the same 8× budget (DINOv2R-S):

ADNet:

BottleneckI-AUROCAUPRO
DynaMLPShareMLPΔDynaMLPShareMLPΔ
NB83.8084.62+0.8281.9182.69+0.78
PTB84.1785.49+1.3282.1483.17+1.03

The I-AUROC gains are 2–3× larger than those on Real-IAD Variety. These results show that cross-layer sharing becomes increasingly important as the number of categories grows relative to the parameter budget, which is precisely the regime that large-scale MUAD targets. We will add the ADNet comparison and this budget analysis to the revision.

AuthorsOpenReview ↗

Response to Reviewer xbtv (part2)

W2.

Insufficient evidence for capacity saturation. The reconstruction-error trends in Section 3.1 do not conclusively establish capacity saturation as the primary cause of performance degradation. Other factors, such as training-data diversity and optimization dynamics, may contribute. More controlled experiments are needed to support this central claim.

Answer:

Thank you for this important comment. It prompted us to strengthen the evidence for capacity saturation by organizing the existing results more systematically and by adding controlled experiments. We address your concern at four parts.

1. Training-data diversity is the source of capacity saturation, not an alternative explanation. Increased training-data diversity is precisely the premise of capacity saturation: the decoder must fit a broader set of normal patterns, where “diverse normal patterns compete and interfere within the decoder’s finite, shared weights” (Abstract and Sec. 3.1). Merely adding images of the same categories would create no extra capacity demand. Meanwhile, the evaluation distribution is kept fixed (the same 20-category set).

A key probe is thus whether the effect of increased diversity is mediated by capacity. Revised Fig. A1 answers this directly: under the identical 20→160 diversity increase, widening only the static MLP (r = 4 → 32), which changes nothing but capacity, substantially alleviates the normal-side saturation. This indicates that the saturation is largely mediated by capacity rather than by the diversity increase itself.

2. Controlling for optimization dynamics. In scaling experiments, we keep the optimization as consistent as possible across category settings: all scaling configurations were trained to convergence (Appendix A.3). Furthermore, in new Fig. A3, we compare checkpoints at training budgets proportional to category count (e.g., 20k iterations for 20 categories and 160k for 160 categories), which keeps the per-category training budget constant. The same saturation trend persists under this control.

3. Beyond Fig. 1(b) and Sec. 3.1, multiple lines of evidence throughout the paper points to capacity saturation.

  • (a) Fixed capacity, increasing diversity → normal-side saturation. In Fig. 1(b) and revised Fig. A1(a), we expand the training set from 20 to 160 categories while keeping the model capacity and the 20-category evaluation set fixed. Normal-region reconstruction errors plateau or increase, accompanied by declining detection performance.

  • (b) Fixed diversity, increased capacity → better reconstruction and detection. At C = 160, widening only the static decoder MLP from r = 4 to r = 32 consistently improves detection (Table 4) and reduces normal reconstruction error (revised Fig. A1(b)).

  • (c) Same diversity increase, more capacity → less saturation. Under the identical 20-to-160 protocol (revised Fig. A1), increasing decoder capacity, either by statically widening the MLP (r = 32, panel b) or through ShareMLP (panel c), alleviates normal-side saturation, and the larger the effective capacity, the more sustained the decrease in normal error. Since all models face the same distribution change, and larger MLP width changes nothing but capacity, we believe it is safe to deduce that the saturation is mostly mediated by capacity rather than by distribution shift or optimization.

  • (d) Across datasets and baselines. On ADNet (380 categories), the saturation is even more pronounced: for Dinomaly2-S/B, normal-region errors increase from the very beginning, and for RD4AD, the hard-normal error exceeds the anomaly-region error from the first setting (new Fig. A2).

4. Scope of our claim. We identify capacity saturation as an important, previously overlooked contributor to degradation alongside over-generalization, not as its sole cause. Other factors may also contribute, although exhaustively examining them is beyond the scope of this paper and current understanding. We will make this scope explicit in the revised manuscript as a further direction of exploration.

AuthorsOpenReview ↗

Response to Reviewer xbtv (part3)

W3.

Parameter efficiency. Although the computational overhead is relatively small, Table A6 shows that the parameter count increases substantially from 36.3M to 102.7M. The authors should further justify this trade-off and discuss its implications for practical scalability and deployment.

Answer:

Thank you for raising this practical concern. While Meganomaly is actually parameter-efficient overall, your concern prompted us to compare against an additional eight-model baseline and to discuss deployment trade-offs more explicitly. We address it from three aspects.

Meganomaly is actually parameter-efficient overall. Combining Tables A6 and A7, Meganomaly-S uses fewer parameters than Dinomaly2-B yet improves I-AUROC by +5.0%, with 4× fewer FLOPs and 2.4× higher throughput (581 vs. 240 FPS). We further trained 8 Dinomaly2-S models, each responsible for 20 categories. Despite using 2.9× more parameters than Meganomaly-S, this ensemble yields lower image- and object-level detection with comparable localization. Moreover, it requires knowing each test sample's category group at inference.

ModelTotal ParamsStorageI-AUROCO-AUROCP-AUROCAUPRO
Dinomaly2-B146.6M0.55 GiB89.094.294.081.8
8 × Dinomaly2-S298.7M1.11 GiB93.397.095.584.8
Meganomaly-S (N=64)102.7M0.38 GiB94.097.395.584.6
Meganomaly-S (N=16)45.8M0.17 GiB92.796.895.083.9

N=64 is a performance-oriented operating point, which can be lower. The pool size N provides a flexible trade-off between parameters and accuracy (Table A6). With N=16, Meganomaly-S uses 3.2× fewer parameters than Dinomaly2-B (45.8M vs. 146.6M) while still improving I-AUROC by +3.7%.

Industrial deployment focuses more on throughput than storage Since ShareMLP adds parameters with nearly constant FLOPs, a model size below 0.4 GiB is readily acceptable on typical deployment hardware, while inference speed remains high. We will add this discussion and the multi-model comparison to the revision.

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

This paper proposes Meganomaly to investigate the performance degradation problem when scaling unified anomaly detection to hundreds of categories. The authors argue that, besides the over-reconstruction of abnormal patterns, the insufficient reconstruction capacity for normal patterns is also a crucial cause. To this end, ShareMLP expands the effective capacity of the decoder by combining cross-layer shared weights through image-level routing; meanwhile, PTB compresses local content information while preserving positional and semantic priors. The core perspective is that large-scale unified detection requires simultaneously enhancing the reconstruction capability for normal patterns and restricting the transmission of abnormal information. The paper achieves significant improvements across multiple datasets.

Critical strengths

s1: The paper addresses an important and highly practical problem: unified anomaly detection across hundreds of categories. Compared to maintaining separate models for each category, a unified model better aligns with the needs of large-scale applications. The proposed method achieves significant improvements on large-scale datasets, demonstrating the practical value of this research direction.

s2: The method design has clear motivations, with capacity expansion and information constraints complementing each other. The authors validate the method through various capacity configurations, component ablations, and cross-framework experiments, making the main results highly persuasive. The overall design is also relatively concise, easy to understand, and applicable.

Critical weaknesses

w1: The mechanistic explanation for capacity saturation still requires more adequate support. Increasing training categories also changes data distributions and optimization relationships; therefore, the current results mainly demonstrate the effectiveness of capacity expansion, but cannot definitively attribute the degradation to capacity shortages.

w2: The independent contribution of the method relative to existing studies on dynamic weights and parameter sharing remains to be clarified. It is recommended to explain the key benefits brought independently by image-level routing, cross-layer sharing, and prior preservation, as well as their specific value for large-scale anomaly detection.

w3: The paper lacks discussion and comparison with related retrieval methods such as RAD [1]. Both utilize pre-trained features and normal samples for multi-category anomaly detection. It is recommended to supplement comparable experiments to clarify the differences between the two routes in terms of accuracy, cost, and applicable conditions, in order to better position the contribution of this paper.

[1] Xingwu Zhang, Guanxuan Li, Paul Henderson, Gerardo Aragon-Camarasa, Zijun Long. “Is Task-Specific Training Necessary for Anomaly Detection?” ICML 2026.

Questions for authors

See the weaknesses.

4 comments
AuthorsOpenReview ↗

Response to Reviewer oRMN (part 1)

Thank you for recognizing the practical importance of scaling unified anomaly detection to hundreds of categories, the clear motivation of our complementary designs, and the extensive experimental validation. Your comments have been very helpful in improving our paper, and we address each of them below.

Since OpenReview does not support images in comments, the revised and newly added figures (Figs. A1, A2, and A3) have been included in the updated paper PDF. We invite you to refer to it.


W1.

The mechanistic explanation for capacity saturation still requires more adequate support. Increasing training categories also changes data distributions and optimization relationships; therefore, the current results mainly demonstrate the effectiveness of capacity expansion, but cannot definitively attribute the degradation to capacity shortages.

Answer:

Thank you for this insightful question. It prompted us to strengthen the evidence for capacity saturation by organizing the existing results more systematically and by adding controlled experiments. We address your concern at four levels.

1. The distribution change is the source of capacity saturation, not an alternative explanation. Adding categories necessarily broadens the training distribution, and this is precisely the premise of capacity saturation: the decoder must fit a broader set of normal patterns, where “diverse normal patterns compete and interfere within the decoder’s finite, shared weights” (Abstract and Sec. 3.1). Merely adding images of the same categories would create no extra capacity demand. Meanwhile, the evaluation distribution is kept fixed (the same 20-category set).

A key probe is thus whether the saturation effect of the distribution change is mediated by capacity. Revised Fig. A1 answers this directly: under the identical 20→160 distribution change, widening only the static MLP (r = 4 → 32), which changes nothing but capacity, substantially alleviates the normal-side saturation. This indicates that the saturation is largely mediated by capacity rather than by the distribution change itself.

2. Controlling for optimization. Regarding the change in optimization relationships, we consider two aspects: inter-category interference and insufficient optimization. Inter-category interference is, in our view, a manifestation of capacity saturation, as it arises from categories competing within finite shared weights; ShareMLP reduces it by routing different categories to different weight combinations. For insufficient optimization, we keep the optimization as consistent as possible across category settings: all scaling configurations were trained to convergence (Appendix A.3). Furthermore, in new Fig. A3, we compare checkpoints at training budgets proportional to category count (e.g., 20k iterations for 20 categories and 160k for 160 categories), which keeps the per-category training budget constant. The same saturation trend persists under this control.

3. Multiple lines of evidence point to capacity saturation, not just UAD performance gain. Our conclusion rests on how reconstruction errors respond to controlled changes in data diversity and capacity, not only on the improved AD results:

  • (a) Fixed capacity, increasing diversity → normal-side saturation. In Fig. 1(b) and Fig. A1, we expand the training set from 20 to 160 categories while keeping the model capacity and the 20-category evaluation set fixed. Normal-region reconstruction errors plateau or increase, accompanied by declining detection performance.

  • (b) Fixed diversity, increased capacity → less saturation. At C = 160, widening only the static decoder MLP from r = 4 to r = 32 consistently improves detection (Table 4) and reduces normal reconstruction error (revised Fig. A1(b)).

  • (c) Same diversity increase, increased capacity → less saturation. Under the identical 20-160 protocol (revised Fig.A1), increasing decoder capacity, either by statically widening the MLP (r = 32, panel b) or through ShareMLP ( panel c), alleviates normal-side saturation.

  • (d) Across datasets and baselines. On ADNet (380 categories), the saturation is even more pronounced: for Dinomaly2-S, normal-region errors increase from the very beginning, and for RD4AD, the hard-normal error exceeds the anomaly-region error from the first setting (new Fig. A2).

4. Scope of our claim. We identify capacity saturation as an important, previously overlooked contributor to degradation alongside over-generalization, not as its sole cause. Other factors may also contribute, although exhaustively examining them is beyond the scope of this paper and current understanding. We will make this scope explicit in the revised manuscript as a further direction of exploration.

AuthorsOpenReview ↗

Response to Reviewer oRMN (part 2)

W2.

The independent contribution of the method relative to existing studies on dynamic weights and parameter sharing remains to be clarified. It is recommended to explain the key benefits brought independently by image-level routing, cross-layer sharing, and prior preservation, as well as their specific value for large-scale anomaly detection.

Answer:

Thank you for this valuable suggestion. We will present the role of each design choice more clearly in the revision, as discussed below.

Our primary contribution is identifying capacity saturation as a key failure mode of large-scale MUAD. On the technical side, we do not aim to invent an entirely new mechanism; rather, we identify which design choices matter for large-scale UAD and realize them in the simplest form. Below we consolidate the ablations scattered across the paper to isolate each design choice (Meganomaly-S on Real-IAD Variety, 160 classes).

(1) Image-level routing. Routing on the CLS token makes all patches of an image share the same composed weights, so anomalous patches cannot be dispatched to the weights that reconstruct them best. Under the same 8× parameters and 1× FLOPs, it clearly outperforms per-token MoE (Table 4).

FFNRoutingParamsFLOPsI-AUROC
Static MLP–1×1×86.34
MoE-top1per-token8×1×87.38
DynaMLPimage-level8×1×91.66

(2) Cross-layer sharing. Sharing aims to enlarge the space of weight combinations and thus the effective capacity. ShareMLP lets every layer route over a common pool, expanding each layer's accessible bases from K to N without extra parameters. It consistently outperforms DynaMLP at every matched parameter budget (Table 6).

FFN ParamsDynaMLPShareMLP
1×86.34 (static MLP)87.47 (N=8)
2×88.27 (K=2)89.67 (N=16)
4×90.18 (K=4)90.81 (N=32)
8×91.66 (K=8)92.00 (N=64)

(3) Prior preservation. Expanded capacity can also facilitate anomaly reconstruction (Fig. A1 c), so we tighten the bottleneck. However, naive compression and noise destroy the spatial and category information that ViT decoders rely on. Explicitly preserving both recovers and further improves performance, confirming their importance (Table 5).

Bottleneck (with ShareMLP)I-AUROCP-AUROC
Wide NB (d→4d→d)92.0095.12
Tight (d_t=32)88.4393.46
Tight + positional info93.6595.44
Tight + positional & category info (PTB)93.9995.46

In summary, image-level routing makes capacity expansion safe for UAD, cross-layer sharing enlarges the space of weight combinations, and prior preservation enables tight information control. More complex alternatives either underperform or bring no consistent gains (Tables 4 and A4), indicating that the improvements stem from these simple design choices rather than mechanism complexity.

AuthorsOpenReview ↗

Response to Reviewer oRMN (part 3)

W3.

The paper lacks discussion and comparison with related retrieval methods such as RAD [1]. Both utilize pre-trained features and normal samples for multi-category anomaly detection. It is recommended to supplement comparable experiments to clarify the differences between the two routes in terms of accuracy, cost, and applicable conditions, in order to better position the contribution of this paper.

Answer:

Thank you for pointing out this relevant work. Following your suggestion, we reproduced RAD on Real-IAD Variety and ADNet, will add it to Table 1, and will include a discussion of retrieval-based methods in the related work.

Results. We evaluate RAD following its official configuration with a DINOv3-B encoder. Storage refers to the total model size for parametric methods and the memory bank size for RAD; FPS is measured on the same RTX 4090 GPU.

Real-IAD Variety (160 categories):

MethodI-AUROCO-AUROCP-AUROCAUPROStorageFPS
Dinomaly2-B89.094.294.081.80.55 GiB240
RAD91.194.594.279.4179.3 GiB12.9
Meganomaly-B94.697.895.885.41.52 GiB215

ADNet (380 categories):

MethodI-AUROCP-AUROCAUPROStorageFPS
Dinomaly2-B82.093.181.40.55 GiB240
RAD85.394.983.51043.9 GiB3.2
Meganomaly-B89.194.985.71.52 GiB215

On Real-IAD Variety, Meganomaly-B outperforms RAD by +3.5% I-AUROC and +6.0% AUPRO. On ADNet, Meganomaly-B leads by +3.8% I-AUROC and +2.2% AUPRO with comparable P-AUROC. Meanwhile, RAD's memory bank grows from 179.3 GiB on Real-IAD Variety to 1043.9 GiB on ADNet (100× and 700× the storage of Meganomaly-B), and its speed drops from 12.9 to 3.2 FPS (17× and 67× slower), whereas the cost of Meganomaly stays constant.

RAD’s performance is, interestingly, consistent with our diagnosis: its non-parametric memory grows with the training data, so it avoids capacity saturation and outperforms fixed-capacity baselines such as Dinomaly2, but at a storage and inference cost that scales with the number of categories. Meganomaly expands capacity parametrically at a nearly fixed cost. RAD is thus attractive when training is undesirable or normal samples are scarce, whereas Meganomaly suits large-scale deployment where throughput and storage matter.

I thank the authors for the additional experiments. The controlled scaling results address my concern about capacity saturation, the ablations clarify the role of each design choice, and the RAD comparison helps put the accuracy and deployment costs in context.

Overall, the response addresses my concerns, and I will keep my original rating of 3. Good luck!

Comments on the paper

2 comments
AuthorsOpenReview ↗

Rebuttal begin!

We sincerely thank all reviewers for their thoughtful and constructive feedback. We are encouraged that the reviewers recognized the practical importance of scaling unified AD model to hundreds of categories, the value of our capacity-saturation diagnosis. Your comments have substantially improved our paper.

As OpenReview does not support images, all revised and new figures are included in the updated PDF.

We hope our responses address your concerns, and we would be happy to discuss any further questions.

Dear reviewers, Thank you for your reviews. The author-reviewer discussion closes on November 18, so please respond to the authors' rebuttal before then. Please say whether it addressed your concerns, and update your score if your view has changed.

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

Related papers

Discussion (0)

Sign in to comment.