acceptodds
Withdrawn from ICLR 2027

Modulation-Only Feature Reconstruction for Unified Anomaly Detection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Feature reconstruction for anomaly detection uses the difference between the input features and a reconstruction learned from normal data. We propose Modulation-Only Feature Reconstruction (MOFR), which pairs an image-independent initial state with per-token feature conditioning. The decoder of MOFR starts from learnable positional embeddings shared by all images, and the input features generate the modulation coefficients of the residual updates and of the output head instead of forming the initial state. Using the per-token conditions and a global condition obtained from their mean, the decoder reconstructs features without token-to-token attention. On MVTec-AD, VisA, and Real-IAD, MOFR reaches image-level AUROCs of 99.8%, 99.0%, and 91.1% and AUPROs of 96.1%, 95.8%, and 96.3%. Crossing the initialization pathway with per-token conditioning shows that the contribution of per-token conditioning depends strongly on whether the initial state contains the input features. The differences from feature-stream reconstruction, which uses an input-derived state, are small and vary with the dataset and the metric. These results show that pairing a shared initial state with per-token conditioning is a reconstruction scheme that replaces input-derived initialization with conditional updates while retaining strong anomaly detection performance.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
anomaly detection, multi-class anomaly detection, unsupervised anomaly detection

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_ZHEXOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes MOFR, a feature reconstruction method for multi-class unsupervised anomaly detection built on the Dinomaly/Dinomaly2 framework. Instead of initializing the decoder state with input-derived features, the decoder starts from a learnable positional.

Critical strengths

The paper is clearly written, and the authors honestly acknowledge that the differences from the baseline are small.

Critical weaknesses
  1. Unaligned training settings. The method is built on the Dinomaly framework, but the training schedule differs. Since the baseline numbers are copied from the literature, the comparison is not controlled. The authors should compare with Dinomaly-based baselines under identical training settings.

  2. Missing comparison with Dinomaly2. The method adopts the CLS re-centering of Dinomaly2, and the ablation (Table 3) shows that it is one of the main sources of the improvement over Dinomaly, yet Dinomaly2 is not compared. The Dinomaly2 baseline reports 99.8%, 99.2%, and 91.6% I-AUROC on MVTec-AD, VisA, and Real-IAD, which appear similar or even higher than MOFR's results (99.8%, 99.0%, 91.1%).

  3. Limited novelty. Using a learnable, image-independent embedding as the decoder's initial state (rather than input-derived features) is not new; similar designs already exist, e.g., the learnable queries in UniAD and INP-former.

  4. No meaningful improvement from the proposed design. Table 2 shows that the proposed modulation-only design performs almost identically to the feature-stream baseline (I-AUROC differences of only 0.02–0.07 pp, and AUPRO is even lower on VisA and Real-IAD). With such small differences, results from multiple random seeds (mean and standard deviation) are necessary to establish whether there is any stable improvement.

Questions for authors

See weaknesses

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

This paper proposes MOFR for unified unsupervised anomaly detection. The method allows the decoder to start from shared learnable positional embeddings, and completes modulated reconstruction through token-wise and global conditions generated from input features, replacing input-dependent initialization. The authors validate the method on three industrial anomaly detection datasets and analyze the relationship between initialization methods and conditional inputs through cross-experiments.

Critical strengths

s1: The core design of the paper is clear. The cross-comparison between the initialization method and the token-wise condition helps to understand the relationship between the two information injection paths.

s2: MOFR maintains strong performance across three datasets; its throughput at the same resolution is higher than that of Dinomaly and INP-Former, and inference does not require normal reference memory.

Critical weaknesses

w1: The benefits of the core design are marginal. In Table 2, I-AUROC only improves by 0.02–0.07%, AUPRO slightly decreases on two datasets, and there is a lack of repeated training validation; thus, the independent value of the new information path is not yet fully demonstrated.

w2: The related comparisons are incomplete. The main table lacks the directly related reconstruction method RLR [2]. RAD [1] similarly utilizes frozen pre-trained features to handle multi-class anomaly detection, but replaces reconstruction training with normal sample retrieval; it is worth comparing their trade-offs in accuracy, training cost, and memory.

w3: The source of the gains remains unclear. Table 2 simultaneously changes the input initialization and the positional prior, making it difficult to distinguish the respective roles of the two; the increase in the average error margin in Figure 4 does not necessarily imply an improvement in detection capability.

w4: Appendix A, cited in the main text, could not be found in the current PDF, making it impossible to verify the complete trade-off between speed and localization accuracy under different resolutions.

[1] Xingwu Zhang et al. “Is Task-Specific Training Necessary for Anomaly Detection?” ICML 2026.

[2] Liren He et al. “Learning Unified Reference Representation for Unsupervised Multi-class Anomaly Detection.” ECCV 2024.

Questions for authors

See the weaknesses.

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

This paper proposes Modulation-Only Feature Reconstruction (MOFR) for unified unsupervised anomaly detection. Instead of initializing the decoder with input-derived features, MOFR employs shared learnable positional embeddings and reconstructs features through per-token and global modulation. The attention-free decoder aims to reduce reconstruction shortcuts while preserving anomaly discrimination. Experiments on MVTec-AD, VisA, and Real-IAD demonstrate competitive anomaly detection and localization performance.

Critical strengths
  1. The motivation is good. The paper explores an alternative feature reconstruction mechanism by separating decoder initialization from input-dependent conditioning, providing an interesting perspective on shortcut learning in unified anomaly detection.

  2. The organization is good.

  3. The ablation studies are thorough.

Critical weaknesses
  1. Limited methodological novelty and marginal improvements. The proposed MOFR mainly builds upon existing conditional modulation techniques, including FiLM and adaLN-Zero. More importantly, Table 2 shows that MOFR only marginally improves I-AUROC over Feature-stream Reconstruction (+0.04, +0.02, and +0.07 pp across three datasets), with some localization metrics performing slightly worse. The authors should better justify the methodological novelty and practical significance of this alternative reconstruction scheme.

  2. Insufficient evidence for mitigating reconstruction shortcuts. Although MOFR avoids input-derived decoder initialization, input information can still flow through the per-token modulation coefficients. Moreover, the larger reconstruction-error gap in Figure 4 does not consistently translate into better detection performance. Stronger evidence is needed to demonstrate that modulation-only reconstruction effectively mitigates shortcut learning.

  3. Incomplete comparisons and limited evaluation scope. The experimental comparisons omit the important Dinomaly2 baseline, despite discussing and adopting techniques from this work. Additionally, evaluation is limited to three relatively small-scale benchmarks (15, 12, and 30 categories). Comparisons with stronger recent methods and experiments on larger-scale benchmarks would better establish the effectiveness and scalability of MOFR.

  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 investigates unified anomaly detection through adaptive feature reconstruction and addresses interference across categories and domains. Although the technical approaches differ, this work is relevant to the motivation and methodological positioning of MOFR. The authors should discuss this connection and better distinguish their contributions from existing unified reconstruction frameworks.

Questions for authors

Please see the weaknesses part.

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

Related papers

Discussion (0)

Sign in to comment.