acceptodds
Under review as a conference paper at ICLR 2027

FlowCTRL: Robust Terminal Control of Flow Models for Anomaly Detection

Abstract

Detecting defects by restoring the test image requires keeping every normal pixel and none of the defect, while which pixels are defective is unknown. We cast this as one control problem on a flow matching model trained on normal images, and propose FlowCTRL: the flow generates a normal image from a Gaussian draw, a control steers it towards the observation, and a per-pixel belief sets how far the observation is trusted. Because steering pulls an unflagged defect into the restoration, the method measures the belief first, on probe paths that no control touches, and then steers the flow with a closed-form controller that blends the model's endpoint prediction with the observation pixel by pixel. The restoration keeps the observation where it is trusted and the model's own completion elsewhere, and its residual is the anomaly map. FlowCTRL uses only forward passes of a pixel-space flow model, without pretrained features, at \nRefSeconds s per image. On MVTec-AD and VisA, it shows remarkable pixel-level localization performance matching or surpassing methods built on pretrained feature extractors including DINOv2, and presents strong image-level anomaly detection ability.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
anomaly detection, anomaly localization, flow matching, inference-time control

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_e1jAOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes an anomaly detection method that formulates anomaly localization and image restoration as a control problem using a flow-matching model trained on normal images. The method first estimates pixel-wise anomaly probabilities by comparing the model's predictions along uncontrolled probe trajectories with the observed test image. These estimates are then used to guide the generative flow toward the observation, preserving regions considered normal while allowing the model to reconstruct potentially anomalous regions. The procedure can be repeated over multiple restoration cycles, and the final anomaly map is obtained from reconstruction residuals. The method achieves competitive anomaly localization performance on MVTec-AD and VisA without relying on pretrained feature extractors.

Critical strengths
  • Novelty. While the general concept of guiding image restoration toward the observed input is not new, the specific implementation within a flow-matching framework, including the proposed anomaly belief estimation and control mechanism, appears to be novel.

  • Mathematical treatment. The authors provide a detailed mathematical formulation of the proposed optimization problem and derive several results concerning anomaly estimation and the control mechanism.

  • Experimental evaluation. The method achieves competitive anomaly localization performance on MVTec-AD and VisA, although it falls short of the best results reported in the literature (e.g., [1]). The authors also provide ablation studies investigating the contributions of individual components and the computational budget.

Reference

[1] Qiyu Chen, Huiyuan Luo, Chengkan Lv, and Zhengtao Zhang. A Unified Anomaly Synthesis Strategy with Gradient Ascent for Industrial Anomaly Detection and Localization. In Proceedings of the European Conference on Computer Vision (ECCV), LNCS 15125, pp. 37–54, 2024.

Critical weaknesses

W1. Limited clarity of the conceptual novelty.

The central idea of generating a normal image guided by a potentially anomalous observation is not new. Previous generative anomaly detection methods have already explored this principle, including diffusion-based approaches such as AnoDDPM (2022), MDPS (2024), and DiAD (2024).

While the claim in Section 2 that existing methods assume known guidance may hold for the specific flow-control approaches discussed, it does not generalize to generative anomaly detection methods more broadly. These earlier methods already reconstruct normal images conditioned on observations without requiring known anomaly masks.

The contribution of FlowCTRL therefore appears to lie primarily in its specific formulation of observation-guided restoration as a control problem and its mechanism for estimating pixel-wise anomaly beliefs. The paper would benefit from positioning these contributions more explicitly relative to existing generative reconstruction methods, rather than motivating them primarily through limitations of flow-control approaches.

W2. Insufficient motivation for the reconstruction stage.

An important conceptual question concerns the relationship between anomaly estimation and restoration. FlowCTRL first estimates a pixel-wise anomaly belief using uncontrolled probe trajectories. This belief already provides an anomaly localization signal.

It is therefore not immediately clear why an additional reconstruction stage is necessary for anomaly detection, rather than using the estimated belief directly as the anomaly map. The reconstruction may improve localization by combining anomaly probabilities with reconstruction residuals, but this motivation and its practical benefits should be established more clearly.

W3. Unclear presentation and excessive methodological complexity.

The paper is difficult to follow, primarily because of its presentation. Numerous technical concepts are introduced in rapid succession, including control costs, anomaly support relaxation, entropy regularization, probability factorization, probe trajectories, endpoint predictions, and restoration cycles.

The difficulty is not simply the mathematical complexity. The relationship between these components and their role in addressing the underlying anomaly detection problem is often insufficiently explained. Several passages also contain awkward or ambiguous language that makes the intended meaning unnecessarily difficult to reconstruct.

As a result, considerable effort is required to understand the basic methodological idea before the significance of the mathematical contributions can be assessed. A clearer, more intuitive explanation of the complete procedure and its central contribution would substantially improve the paper.

W4. Limited improvements over a substantially simpler flow-based baseline.

Although FlowCTRL achieves competitive results, its improvements over Flow Mismatching (Chen et al., 2026) are generally modest, despite introducing a substantially more elaborate inference procedure involving pixel-wise belief optimization, controlled reconstruction, and iterative restoration cycles. On MVTec-AD, P-AUROC and P-AP improve by only 0.4 percentage points each. On VisA, P-AUROC improves by 1.2 points and P-AP by only 0.1 points. Given the known annotation noise in both datasets, such small differences are difficult to interpret as meaningful improvements.

Furthermore, the reported results do not establish a new state of the art, as higher performance has already been reported in the literature, for example by GLASS [1]. A more explicit comparison of computational costs and the performance–complexity trade-off would help establish the practical benefits of the proposed approach.

[1] Qiyu Chen, Huiyuan Luo, Chengkan Lv, and Zhengtao Zhang. A Unified Anomaly Synthesis Strategy with Gradient Ascent for Industrial Anomaly Detection and Localization. ECCV, 2024 (GLASS).

W5. Limited evaluation of computational efficiency and image resolution.

The authors emphasize the computational efficiency of FlowCTRL, reporting an inference time of approximately 0.28 seconds per image on an NVIDIA A100 GPU. However, all experiments are conducted at a relatively low resolution of 128×128 pixels, whereas industrial anomaly detection methods commonly operate at higher resolutions, such as 256×256 or 512×512. It remains unclear how computational cost and anomaly localization performance scale with image resolution, particularly for small defects. An evaluation at higher resolutions, including runtime measurements, would strengthen the practical relevance of the reported results.

Questions for authors

Q1. What is the specific conceptual novelty relative to existing diffusion-based anomaly detection methods?

Methods such as AnoDDPM, MDPS, and DiAD already generate normal reconstructions conditioned on potentially anomalous observations without requiring known anomaly locations. Beyond replacing the generative model with flow matching, what fundamental capability does the proposed control formulation introduce?

Q2. What is the added benefit of the reconstruction stage for anomaly localization?

The observer already produces a pixel-wise anomaly belief before the controller is applied. While the reconstruction stage provides residual-based anomaly scores and enables further refinement across restoration cycles, its added value for anomaly localization remains unclear. Could the authors report the performance obtained directly from the observer's anomaly estimates and compare it with the complete reconstruction-based method?

For further questions, see the Critical Weaknesses section.

Reviewer_3hEEOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes FlowCTRL for unsupervised industrial anomaly detection, built on a category-conditioned pixel-space flow matching model trained only on normal images. The main contributions are:

  1. Formulating image restoration and anomaly localization as a control problem that considers reconstruction fidelity, sparsity and spatial continuity of anomaly regions, and control cost. The anomaly support is relaxed into a per-pixel anomaly belief.
  2. Measuring endpoint prediction errors along uncontrolled Gaussian-to-target probe paths. These errors are aggregated across paths and time steps to identify potentially unreliable pixels.
  3. Blending the model's endpoint predictions with the observed image based on the estimated belief. Multiple restoration cycles refine the reconstruction, and the final residual is used to generate the anomaly map.

FlowCTRL achieves pixel AP scores of 67.9 and 40.3 on MVTec-AD and VisA, respectively, without pretrained feature extractors. The default configuration requires 8 batched network calls, covering 224 state evaluations, with a controller runtime of approximately 0.28 seconds per image on an A100. The paper also includes component ablations, inference-time noise seed studies, and theoretical analysis.

Critical strengths
  1. Directly pulling generated images toward the observation may reproduce defects and weaken anomaly residuals. FlowCTRL estimates the belief using separate probe paths before guiding controlled restoration. This design is well motivated. Table 2 shows that measuring anomaly evidence along controlled paths instead of probe paths reduces pixel AP from 65.9 / 39.6 to 27.3 / 8.9 on MVTec-AD / VisA under the same ten-step time grid. This strongly supports the need for path separation.

  2. FlowCTRL achieves a pixel AP of 67.9 on MVTec-AD using only a pixel-space Flow Matching model trained on normal images. Table 2 validates key components, including spatial coupling, the control strategy, and evidence aggregation. For example, removing spatial coupling reduces pixel AP from 67.9 / 40.3 to 61.3 / 37.9 on MVTec-AD / VisA. Moreover, the configuration selected on MVTec-AD is directly applied to VisA, providing some evidence of cross-dataset generalizability.

Critical weaknesses

1-FlowCTRL emphasizes that it does not require external pretrained features, but Flow Mismatching is the only baseline in Table 1 with a similar setting. Compared with this baseline, FlowCTRL improves pixel AP by only 0.4 / 0.1 percentage points on MVTec-AD / VisA, while image AUROC drops by 1.0 / 1.1 points. Meanwhile, inference FLOPs increase from 2.53T to 5.67T. Given the additional belief optimization and controlled restoration mechanisms, these modest gains do not fully justify the added complexity.

2-The no-control ablation in Table 2 still uses the residual between the generated and observed images as the detection score. However, uncontrolled generation does not necessarily preserve correspondence with the input image. Thus, this comparison does not establish whether restoration residuals outperform the Observer's anomaly evidence directly. More importantly, Table 11 shows that merged probe evidence and belief achieve pixel AP scores of 68.7 and 70.2 on MVTec-AD, respectively, both exceeding the final restoration residual score of 67.8. The opposite trend is observed on VisA. Although these results still depend on the full restoration pipeline, they suggest that the advantage of the final detection score is not consistent across datasets. Without a truly independent observer-only baseline, the incremental value of controlled restoration remains uncertain.

Questions for authors
  1. Performance gains versus added complexity: Compared with Flow Mismatching, FlowCTRL introduces a more complex control and restoration pipeline but achieves only modest detection improvements. Could the authors further explain why these additional mechanisms are necessary from a theoretical or methodological perspective, and under what conditions they are expected to offer more substantial benefits?

  2. Observer-only baseline: Could the authors provide an independent observer-only baseline using the same Flow Matching model, evidence aggregation, spatial coupling, and post-processing settings? Specifically, the probe target should remain fixed to the original observation, with no controlled restoration, and detection should rely directly on the evidence or belief. This would help isolate the contribution of the Controller.

  3. Theoretical justification of the control strategy: Could the authors clarify why Eq. (15) uses st2s_t^2 rather than sts_t in the control cost, and how this one-step approximation relates to Eq. (12)? Additionally, given that the practical discrete updates and terminal step use different control rules, it would be helpful to clarify how the full algorithm relates to the original optimization objective.

Reviewer_P9BCOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes FlowCTRL, a restoration-based method for anomaly detection, formulated as a control problem. A class-conditional flow matching model is first trained on normal images only. The flow generates a normal image from Gaussian noise, while a control steers the generation towards the test image. After relaxing the anomaly support into a per-pixel belief, the method first estimates the belief on uncontrolled probe paths, and then uses it in a closed-form controller to restore the test image. Experiments show promising results.

Critical strengths
  1. The method is clearly explained, and the figures are helpful for understanding it. The overall approach makes sense.
  2. The mathematical derivations are clear and rigorous.
Critical weaknesses
  1. The evaluation results are relatively weak. The paper notes that most of the compared baselines rely on pretrained models, so the comparison is not entirely fair. However, from a practical perspective, using a pretrained model is neither particularly difficult nor costly, so the proposed method does not offer a clear practical advantage. As a result, the evaluation is not sufficiently compelling.
    • If the proposed method does not achieve the best performance without a pretrained model but offers other advantages, these advantages should be clearly demonstrated in the evaluation.
    • Alternatively, the authors could incorporate a pretrained model into the proposed method to remove the unfairness of the comparison.
  2. The motivation is not sufficiently developed. The second paragraph of the introduction, which discusses the limitations of existing methods, consists of only a few sentences: it briefly mentions the “identical shortcut” of reconstruction networks, the fact that a generative model’s sample is not the normal version of the observation, and that most current methods compare features of pretrained networks rather than pixels. Each of these points is stated in a sentence or two without further explanation or supporting evidence. In particular, it is unclear why comparing features instead of pixels is problematic. Elaborating on these limitations would provide a clearer and more convincing motivation for this work. In addition, the word “robust” in the title is not reflected in the paper. In what sense, and why, is the method robust?
  3. Minor issues:
    • Inconsistent notation: most of the text uses did_i, while some places use di(L)d_i(L).
    • Eq. (7): “arg” and “min” should be typeset as a single operator, i.e., argmin⁡q∈⋯\operatorname*{argmin}_{q\in\cdots}.
    • Line 37: why is the word “a” in “a normal image” italicized?
Questions for authors

See weaknesses.

1 comment
Reviewer_P9BCOpenReview ↗

No rebuttal?

No rebuttal for this paper?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.