acceptodds
Under review as a conference paper at ICLR 2027

Divergence-Aware Flow Likelihoods for Industrial Anomaly Localization

29 views

Abstract

Recent work on flow matching for visual anomaly detection has adopted Gaussian endpoint scores and density proxies that avoid evaluating the full flow likelihood. These choices bypass the costly divergence integral, raising the question of when this term can be omitted and what information it contributes to detection and localization. We introduce FlowAD, which learns a flow-matching density over frozen foundation-model patch features. Its likelihood score separates into Gaussian endpoint energy and integrated divergence, the volume-change correction. We estimate divergence at inference using forward-only finite-difference Hutchinson probes. Controlled comparisons of endpoint-only, correction-only and full scores show that the correction improves localization even when image-level detection declines: on VisA, adding it to single-layer C-RADIO endpoint scores raises P-AUPRO by 4.92 percentage points while image AUROC falls by 0.66 points. Localization gains also appear on MVTec AD and with a second backbone, but the correction's additional benefit depends on how layers are combined. In a separate fixed-model-bank control on VisA, multi-layer fusion reduces its mean P-AUPRO gain from 4.86 to 1.15 percentage points and introduces category-level losses. We further prove that, for Gaussian data with positive-definite covariance and an independent standard Gaussian base under linear interpolation, the population-optimal flow-matching field is affine and its divergence is input-independent, so endpoint scoring preserves the full likelihood ranking. These findings establish a sufficient condition for omitting divergence integration and show that the correction's empirical value depends on the task and layer configuration.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Industrial Anomaly Detection, Anomaly Localization, Flow Matching, Density Estimation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_WH87OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper studies the question of "when does the volume-change correction carry anomaly information beyond the Gaussian endpoint, and what is lost by omitting it" in the context of flow-matching anomaly detection for (industrial) images.

By comparing the endpoint-only, correction-only, and full scores of FlowAD, they conclude that dropping the correction costs pixel-level localization accuracy, while image-level detection is largely unaffected or slightly improved.

Critical strengths

The experiments show that the practice of dropping the divergence term in existing methods might be problematic.

Critical weaknesses
  1. The problem statement of when the divergence term can be omitted is meaningful by itself, but the paper studied this problem mostly empirically with one model family (albeit across two backbones and three datasets). This makes the paper a bit shallow, i.e., no novel or interesting insight was revealed about why we cannot omit the divergence term.

  2. While a condition is given in Proposition 1 that describes a case where the divergence term can be omitted theoretically, this result feels isolated from the empirical results which mainly show that the divergence term is needed for localization.

  3. Although the third point in the contribution list highlights that FlowAD exceeds PatchCore by 2.46 P-AUPRO points on VisA, this comparison is selective. In Table 4(c), PatchCore, which requires no training, achieves the best result in half of the twelve columns, including image AUROC on all three datasets, and no single FlowAD configuration is consistently better.

  4. Overall, the contribution is limited. The method combines existing components (flow matching, Hutchinson estimation, frozen pretrained features), the theoretical result covers only a Gaussian case that does not bear on the empirical findings, and the empirical finding itself is not explained or turned into a design principle.

Questions for authors

As the third point of the contribution list states that the authors "introduce FlowAD, a flow-matching method that detects and localizes anomalies ...", would the authors kindly clarify which components of FlowAD, apart from the finite-difference trace estimator, are new / designed by them? If most components are from existing methods, what are the contributions / challenges in making it work?

2 comments
AuthorsOpenReview ↗

Response to Reviewer WH87

Thank you for identifying the need to connect the analysis to the empirical findings. We have strengthened that connection and clarified the contribution and benchmark comparisons.

  1. Contribution. We respectfully disagree that the contribution is adequately characterized as combining existing components. FlowAD addresses a substantive question left open by recent flow-based anomaly detectors: when does replacing a learned density by its endpoint energy preserve the rankings needed for detection and localization? MATCH reports negligible effects from retaining divergence in its Real-IAD setting. Our fixed-field comparisons identify settings with substantial localization effects: retaining the correction improves VisA P-AUPRO by 4.92 points across three seeds, while image AUROC decreases by 0.66 points. The contribution is not a new flow-matching objective, but an analysis of a consequential scoring choice made by current methods.

The revised paper explains this distinction through ranking margins and spatial aggregation rather than score magnitude alone. The Gaussian case identifies an exact ranking-preserving limit; Sections 5.2–5.4 connect it to image-pair reversals, common localization errors and fitted-density comparisons. Forward-only finite-difference evaluation makes the correction accessible without differentiation through the field. We quantify its agreement and cost against shared-probe automatic differentiation. The FM objective, CNF identity and Hutchinson principle are established; our claims concern their implications for anomaly scoring.

  1. Theory and observations. Section 4 gives a sufficient null case: an independent standard-normal base, positive-definite Gaussian data, linear interpolation, an unrestricted population-optimal field and exact integration yield Mahalanobis endpoint energy and input-independent V. The correction therefore cannot change rankings. For a learned field, its variation relative to endpoint margins matters: for anomalous/normal margin mm, m′=m+qa−qnm'=m+q_a-q_n. A pair with m<0m<0 is repaired when qa−qn>−mq_a-q_n>-m; one with m>0m>0 is harmed when qa−qn<−mq_a-q_n<-m. Ties receive half credit.

For image-max scoring, qi=Vi(pE(i))+riq_i=V_i(p_E(i))+r_i, where pEp_E is the endpoint maximum and ri≥0r_i\ge0 is the gain from spatial reselection. Section 5.2 applies this identity to 96,200 within-category pairs per cohort. No ties occur in these cohorts. Category-macro repair minus harm reproduces the AUROC change: -0.6532 points for C-RADIO L31 and +1.2417 for DINOv3. At 5% FPR, Section 5.3 shows that smaller fusion gains can coexist with greater recovery on common endpoint errors. These analyses connect the ranking criterion to measured outcomes; they do not identify a causal mechanism of pretraining. The C-RADIO fusion models are distinct from Table 1's three-seed models.

  1. Gaussian and training controls. The distinction is also empirically important. DINOv3 Gaussian/E/full P-AUPRO is 93.54/85.73/93.45: the correction largely restores the endpoint's gap to Gaussian. In the C-RADIO comparison on identical scoring features, full exceeds Gaussian by 1.07 P-AUPRO, 3.24 pixel-AP and 2.96 image-AUROC points. Thus gains beyond the endpoint are not uniformly gains beyond Gaussian.

In a separate 12-category budget control, normal-training FM MSE falls in every category from 30 to 120 epochs, while E P-AUPRO falls from 85.85 to 83.41 and image AUROC rises from 91.92 to 93.63. Lower measured FM error does not imply better localization rankings. These seed-0 runs use separate cosine schedules and in-sample MSE, not a convergence test. Eight categories use re-extracted features and new models; scores share features within each category. The Gaussian proposition is a population reference, not a finite-network guarantee.

  1. Performance and scope. The evidence extends beyond the highlighted PatchCore result. Table A11 compares five fitted density estimators on identical scoring features and postprocessing. FlowAD has the highest pixel AP and P-AUPRO on both MVTec AD and VisA, and ranks second in image AUROC. This supports a strong localization result within this comparison, not superiority on every metric. Training/model-selection budgets differ, and historical neural training inputs and seeds remain incompletely documented.

Table 4 separately compares recent flow-based methods under their native protocols. These results provide context, not a controlled ranking across papers. Section 5.6 also reports matched max-after-fusion PatchCore scores on all 57 categories. VisA image AUROC ties five-layer FlowAD at two decimals; FlowAD retains a 2.46-point P-AUPRO advantage. PatchCore leads image AUROC on MVTec AD/Real-IAD and P-AUPRO on MVTec AD. Layer selection, normalization and fusion still differ. We now foreground both the fitted-density comparison and these task-specific trade-offs rather than using one metric to imply general dominance.

I thank the authors for the additional analyses and for clarifying the paper’s intended contribution. The revised PatchCore comparison addresses W3 by reporting where each method performs better, but in general I am not fully convinced that my concerns have been addressed.

For W1, my question was why retaining the divergence term improves localization in the settings studied. The additional analyses clarify how the gains vary with layer fusion, though I still find it unclear why the correction helps identify defective regions.

For W2, I appreciate the connection the authors draw between Proposition 1 and changes in score rankings. My remaining concern is that it is still unclear how the Gaussian result helps explain the observed localization gains.

Regarding W4 and Q1, I asked which components of FlowAD the authors designed beyond the finite-difference trace estimator and what challenges arose in combining the established components. The response clarifies that the authors regard the controlled comparison of scoring rules as their main contribution. It also explains the computational problem addressed by the estimator. Beyond that estimator, the response does not identify an additional component the authors designed or a specific challenge in combining the remaining components.

Reviewer_qcp7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies whether the divergence (volume-change) term of a flow-matching likelihood matters for industrial anomaly detection. Recent flow-based detectors such as MATCH and WT-Flow drop this term or use proxies. FlowAD trains one flow-matching velocity field (a residual MLP) per category and layer on frozen patch tokens from C-RADIOv4-H. It then scores each token by its negative log-likelihood, which splits into a Gaussian endpoint energy E=12∣z0∣2E=\frac12|z_0|^2 and an integrated divergence VV. VV is estimated with Hutchinson probes, where each Jacobian-vector product is replaced by a forward finite difference, so no autodiff is needed.

The main finding comes from paired comparisons on the same trained field. Adding VV to EE improves P-AUPRO by 4.92 points on VisA and 1.08 on MVTec AD with single-layer features, and by 7.72 on VisA with DINOv3. Image AUROC drops slightly with C-RADIO. With five-layer fusion the gain shrinks to 1.15 points, with losses in some categories. The paper also proves that for Gaussian data with a standard Gaussian base and linear interpolation, the optimal FM field is affine, so VV is constant and the endpoint alone preserves the likelihood ranking. Under a common C-RADIO backbone, five-layer FlowAD beats a PatchCore adaptation by 2.46 P-AUPRO on VisA, is roughly tied on MVTec AD and Real-IAD, and trails it on image AUROC everywhere. Finite differences take about 41% of the JVP scoring time.

Critical strengths
  1. The question is well posed and the paired design is careful. Recent flow-based detectors drop the divergence term largely for efficiency, and nobody had really checked what that costs. Here EE, VV and E+VE+V are compared on the same trained field, trajectory, probes and postprocessing, so the effect of the term is isolated from representation and training differences. The localization gain is consistent: 36/36 VisA category–seed pairs, 14/15 MVTec categories, and it reproduces on a second backbone. That makes it a credible and useful empirical observation for this line of work.
  2. The numerical side is checked thoroughly. The paper varies the step count, probe count and difference scale (Table A1). It runs a shared-probe FD-vs-JVP control (P-AUPRO differences below 0.009 points, Table A2) and verifies PRO against the official implementation. This rules out the obvious worry that the gain is an artifact of the finite-difference estimator. The authors also report the negative results openly: fusion shrinks the gain, image AUROC drops, and transistor gets worse.
Critical weaknesses
  1. It is not clear that VV adds "information" rather than repairing a poor endpoint.
    • In a learned (non-optimal) field, EE alone is not a principled score, so an improvement from EE to E+VE+V is partly expected. The more informative reference point is a density that does not need this correction. On the same features, a closed-form global Gaussian already reaches 92.24 P-AUPRO on VisA and 94.61 on MVTec AD (Table A11). By comparison, FlowAD's endpoint alone gets 88.32 / 94.55 (Table 1) and the full likelihood 93.31 / 95.62. So the learned flow's endpoint is worse than a Mahalanobis baseline, and the full likelihood ends up only about 1 point above it.
    • Correction-only scoring matches or beats the full score on VisA in both backbones.
    • Together, these suggest VV may be compensating for how the learned transport distorts the endpoint, rather than carrying anomaly information of its own. The paper offers no mechanism either way. For example, it does not show what VV correlates with: token norm, local feature density, or the training-set Gaussian score.
    • The theory does not help much here. The Gaussian proposition (affine optimal field, constant divergence) is close to textbook. It only covers the regime where the question is trivial, and says nothing about when VV becomes informative for non-Gaussian features, which is what the section title promises.
  2. The practical case for FlowAD is weak and the main comparison has protocol issues.
    • Against the PatchCore adaptation on the same backbone, FlowAD loses image AUROC on all three datasets (e.g. 89.62 vs. 92.02 on Real-IAD) and loses P-AUPRO on MVTec AD. It is tied on Real-IAD P-AUPRO.
    • The one clear win, +2.46 P-AUPRO on VisA, uses five-layer fusion standardized with mean/std over all test pixels of each category. That is transductive. PatchCore is aggregated differently (mean of per-layer maxima vs. max after fusion).
    • The divergence term multiplies scoring cost by about 6× (57 vs. 346 ms/image single-layer). Its marginal gain under the fused configuration is 1.15 points, with losses in 5/12 VisA categories.
    • The decomposition is never run on Real-IAD. That is exactly the dataset where MATCH reported negligible gains from the divergence, so the most direct test of the paper's claim against prior work is missing.
    • Several results rely on "existing" checkpoints whose training inputs, seeds, initialization or even "physical layer identities" are stated as not fully established (A.2, B.3, D.3, Table A11). This makes those numbers hard to reproduce.
Questions for authors
  1. What does VV capture? Can you report the correlation between VV and simple per-token quantities (token norm, closed-form Gaussian log-density, kNN distance to training tokens)? Also, does VV still help when added to a better endpoint, e.g. Gaussian/Mahalanobis + VV, or a field trained longer so that EE alone approaches the Gaussian baseline? If most of VV's gain disappears once the endpoint is reasonable, the main claim needs to be reframed.
  2. Real-IAD decomposition. Please report EE, VV and E+VE+V on Real-IAD. Does the gain hold there, and how do you reconcile the result with MATCH's finding of negligible gains on the same benchmark?
  3. Fusion without test statistics. How does five-layer FlowAD perform when fusion uses train-split (or held-out normal) statistics? And how does PatchCore perform with the same max-after-fusion aggregation? The VisA advantage over PatchCore depends on both choices.
  4. Reproducibility. Which reported numbers come from checkpoints with incompletely established training provenance? Can all main-table results be regenerated end-to-end from the code you plan to release, with fresh training?
3 comments
AuthorsOpenReview ↗

Response to Reviewer qcp7 (1/2)

Thank you. Section 5.4 now separates gains over E from gains over a fitted Gaussian.

  1. Gaussian scoring, endpoint weakness and V.

On all 12 DINOv3 VisA categories, Gaussian reaches 93.54 P-AUPRO and 50.07 pixel AP, versus 85.73/43.48 for E, leading P-AUPRO in every category. V/full reach 93.46/93.45 P-AUPRO, below Gaussian, while gaining 1.14/0.85 AP and 0.45/0.57 image-AUROC points. Large gains over E therefore do not imply P-AUPRO gains over Gaussian. This comparison fixes the original flow models and uses identical training and test features within each category.

On 12 C-RADIO VisA categories, methods share scoring features and fixed fields. Training inputs/seeds are incompletely documented; these models differ from the three-seed models in Table 1.

ScoreI-AUROCPixel APP-AUPRO
Gaussian92.6143.2792.24
E96.3142.5288.72
V94.6246.5493.36
Full95.5746.5193.31

Gaussian leads E in P-AUPRO in 9/12 categories; E leads image AUROC in ten, with one tie. Full gains 1.07 P-AUPRO, 3.24 AP and 2.96 image-AUROC points over Gaussian. Adding E to V gains 0.95 image-AUROC points but loses 0.05 P-AUPRO; adding V to E gains 4.59 P-AUPRO but loses 0.74 image AUROC. The three-seed MVTec control gives +0.35 image AUROC over V and +1.08 P-AUPRO/-0.17 image AUROC over E. DINOv3 full gains 0.13 image-AUROC points over V but slightly loses AP/P-AUPRO; V itself exceeds E in detection. Thus C-RADIO gains extend beyond Gaussian; component contributions remain task dependent (Section 5.1).

On the same DINOv3 features, mean Spearman correlations of V with Gaussian negative log-density, token norm, full-bank Euclidean 1-NN and cosine 1-NN are 0.974/0.177/0.824/0.756 on 4,096 native patches per category; Gaussian log-density has the opposite sign. For these 12 fixed models, Gaussian+V reaches 92.26 image AUROC, 50.33 AP and 93.67 P-AUPRO versus Gaussian's 92.10/50.07/93.54. P-AUPRO improves in 7/12 categories. Equal-weight z-score fusion uses 100 normal training images per category, in sample, without test statistics or label-selected weights. The high correlation and small, nonuniform fusion gains show substantial overlap in fitted-density rankings (Appendices H.2–H.3).

Gaussian fits normal-feature moments; E is energy after an integrated learned inverse. Their ideal equivalence requires independent standard-normal base sampling, linear interpolation, a positive-definite Gaussian population, the unrestricted population-optimal field and exact integration. This is an ideal reference, not equality for the finite model or a guarantee of correct anomaly ranking. The controls establish marginal task gains, without proving independence, excluding simple feature geometry or identifying the endpoint gap's cause.

  1. Does more training close the Gaussian–E gap?

The budget study covers all 12 categories (Section 5.4; Appendix H.4). The added eight use re-extracted features and new fields; Gaussian and flow scores share each category's train/test tensors. At 30/60/120 epochs, E P-AUPRO is 85.85/85.08/83.41 versus Gaussian's fixed 93.54; full reaches 93.37/93.59/93.79. With in-sample calibration, Gaussian+V gains 0.11/0.20/0.31 P-AUPRO over Gaussian, but its mean image AUROC and AP stay below full. Mean fixed-probe FM training MSE is 0.0894/0.0823/0.0780, falling in every category. E image AUROC reaches 91.92/92.42/93.63 versus Gaussian's 92.10. From 30 to 120 epochs, detection improves in all 12 categories but P-AUPRO falls in ten. Each budget uses seed 0 and its own cosine schedule: separate runs, not one trajectory or multiple seeds. The 4,096-patch normal-training MSE is in sample. Lower measured FM training error need not improve rankings; these budgets do not establish convergence.

  1. Real-IAD, V-only and MATCH.

Section 5.1 and Appendix I.7 compare all 30 categories with one fixed single-layer C-RADIO model per category and scoring seed 20261110. Training seeds are unknown; this evaluation is separate from the main benchmark.

ScoreI-AUROCP-AUROCPixel APP-AUPRO
E89.853297.503530.029686.4077
V87.469398.895541.205194.6344
Full88.914398.955939.331694.7343

Full minus E is +8.33 P-AUPRO (29/30 categories improve) but -0.94 image AUROC. Full minus V is +1.45 image AUROC, -1.87 AP and +0.10 P-AUPRO; P-AUPRO improves in 12/30 categories. A same-feature Gaussian baseline is unavailable here.

MATCH reports negligible within-model effects on both tasks. MATCH uses WRN50 at 256 pixels, view conditioning, five Euler steps and task-specific layer fusion; our control uses one C-RADIO layer for both. Our result establishes a different within-model tradeoff under this protocol; Section 5.2 explains such effects through beneficial and harmful ranking changes. The protocol differences are not demonstrated causes, and neither study establishes a universal omission rule.

AuthorsOpenReview ↗

Response to Reviewer qcp7 (2/2)

  1. Does fusion depend on test statistics?

Yes. Anomaly prevalence can change layer variances and fusion weights without using labels. A shared positive affine transform of the final score preserves rankings; scaling components or layers separately before fusion can reorder them. Section 5.3 evaluates two calibration controls.

First, all 12 VisA categories and three scoring runs use 100 normal training images per layer to fix full-score calibration. Test maps, layer averaging and spatial-max scoring stay fixed. Test-pool to normal-training calibration changes I-AUROC/P-AUPRO/AP from 96.7234/95.1872/44.4310 to 96.7323/95.1593/44.7771. Small macro changes do not make the weights equivalent: chewinggum gains 2.56 AP points but loses 0.20 P-AUPRO points. These are in-sample normal training images. This evaluation has no paired E/V results, and its models cannot be established as identical to those in the component comparison. It therefore does not establish that comparison's training-calibrated gains or PatchCore margin.

Second, for the models used in the component comparison, 20 normal test images per category provide calibration; both policies score the same remaining 1,922 images. E and V share each layer's full-score scale. The full-minus-E P-AUPRO increment is 1.3509 points with evaluation-pool statistics and 1.4212 with disjoint normal calibration; 7/12 categories improve under either. The image-AUROC cost persists (-0.8322/-0.8630 points). The localization/detection tradeoff thus persists with disjoint normal calibration; it is not solely an effect of evaluation-pool standardization. This post-hoc repartition of previously evaluated maps is distinct from untouched validation and original training-split calibration.

  1. PatchCore aggregation and fairness.

Section 5.6 matches image aggregation on all 57 categories by taking the maximum after fusing PatchCore's layer maps. I-AUROC changes from 99.2002/97.0113/92.0202 to 98.9788/96.7239/91.9560 on MVTec AD/VisA/Real-IAD. Benchmark tables use this aggregation order; the control table retains original values. VisA equals five-layer FlowAD at two decimals. PatchCore still leads image AUROC on MVTec AD/Real-IAD and P-AUPRO on MVTec AD. The localization comparison survives this aggregation change because pixel maps and metrics are unchanged. Selected layers, normalization and fusion still differ between three-layer PatchCore and five-layer FlowAD; this is not a fully matched-input comparison.

  1. Scope of reproduction.

The comparison in point 1 holds the original DINOv3 models and features fixed. Fresh 30-epoch training reproduces four models and native scores on fixed features. The separate budget study extends this four-category control to eight additional categories with re-extracted features and newly trained models. Gaussian+V calibration covers all 12 categories at each of 30/60/120 epochs. These controls do not regenerate all benchmark rows end to end. The C-RADIO main-table and five-layer models lack complete records of initialization, feature extraction and backbone-layer identity; the neural density-estimator comparison also lacks established training inputs and seeds. Published results and the U-Flow reproduction retain their separate sources.

Reproducible initialization requires setting the seed before constructing the network. The accompanying training code and dedicated DINOv3 control follow this order; the earlier generic code did not ensure it. Implementation tests accompany the code, but the five-layer calibration evidence remains limited to the two controls in point 4.

  1. Why pay for the correction, or use full rather than V?

Sections 5.1 and 5.3 give empirical reasons to retain E alongside V: it adds detection gains in several settings, and after five-layer fusion full exceeds V by 3.20 AP and 0.45 P-AUPRO points. The combination also incurs losses: full P-AUPRO is below E in 5/12 fused VisA categories. The practical choice therefore depends on the task and baseline, not a fixed role for either component.

V-only still requires the reverse trajectory and divergence probes, so removing E does not avoid the main cost. In the matched three-category native-patch comparison (Section 5.5), FD full scoring takes about 346–348 ms/image versus 57 ms for E. The 41% ratio compares FD with JVP full scoring. Their agreement assesses finite-difference accuracy on learned fields; both retain Euler and Hutchinson errors.

Single-layer E+V estimates the flow density score; standardized fusion is not a joint likelihood. A density identity alone does not select the best anomaly score. The evidence supports comparing both tasks against Gaussian before accepting the roughly sixfold cost relative to endpoint scoring, without providing an unsupervised selection rule. How representation, aggregation, task objectives and field/derivative errors shape density-based discrimination remains open.

Thank you for the extensive response and the additional controls, and for being upfront about their limitations. Having read the response and the revised Sec. 5.4, I will keep my rating at 2.

The Gaussian comparisons mostly confirm my original concern. On the DINOv3 models, the same-feature Gaussian reaches 93.54 P-AUPRO, slightly above both V and the full likelihood. Gaussian+V adds 0.13, and V correlates with the Gaussian score at 0.974. In the budget study, E gets worse with longer training, while full ends at most 0.25 above the Gaussian, and those are single-seed runs. The 1.07-point gain on C-RADIO is the largest margin over the Gaussian, but it comes from checkpoints whose training inputs and seeds are not documented. Against a simple density on the same features, the localization gain of the full likelihood is therefore about one point at most. This compares with the 4.92-point gain over E in the headline and a roughly sixfold scoring cost. To me this still reads mainly as V repairing a weak learned endpoint. The detection and AP gains over the Gaussian in some settings are interesting, but they are not what the paper claims.

The Real-IAD decomposition is a welcome addition. Without a Gaussian on that benchmark, however, it is hard to say how much of the 8.33-point gain goes beyond what a Gaussian gives. The disjoint-calibration control settles my question about test-pool statistics for the E-vs-full comparison. As you note yourselves, the PatchCore margin on VisA is not established without those statistics. With matched aggregation, PatchCore still leads on image AUROC on MVTec AD and Real-IAD, and on P-AUPRO on MVTec AD.

I am also more concerned about reproducibility than before. The response confirms that the main C-RADIO and five-layer models lack records of initialization, feature extraction and even backbone-layer identity, and that the new controls have not regenerated the main benchmark rows. For a paper whose contribution is a controlled comparison, this matters. In addition, Sec. 5.4 and the conclusion now discuss the Gaussian, but the abstract and introduction still lead with the 4.92-point gain over E. They do not mention that a same-feature Gaussian nearly matches the full likelihood on localization.

I think the question is a good one and the paired design is careful. Three things would change my view: regenerating the main results from documented training runs, including the Gaussian in the main tables on all three datasets, and showing on those models that V adds localization information beyond the Gaussian.

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

This paper investigates the role of the divergence integral in flow-matching-based industrial anomaly detection. The authors propose FlowAD, which models the density of frozen visual foundation-model patch features using Flow Matching. The anomaly score is decomposed into Gaussian endpoint energy and an integrated divergence correction, with the latter estimated using forward-only finite-difference Hutchinson probes.

Through controlled comparisons of endpoint-only, divergence-only, and full-likelihood scores, the paper demonstrates that incorporating the divergence correction can substantially improve pixel-level anomaly localization, although image-level detection performance may decrease in certain settings. The authors further investigate multi-layer feature fusion and observe that the additional localization benefit of the divergence correction becomes smaller after fusion.

The paper also provides a theoretical result showing that, under specific Gaussian assumptions and a population-optimal Flow Matching field, the divergence integral is input-independent. Consequently, endpoint scoring preserves the same sample ranking as the full likelihood.

Critical strengths

The theoretical analysis is a notable strength of the paper. Proposition 1 provides a mathematically sound derivation showing that, under the stated Gaussian assumptions, the population-optimal Flow Matching velocity field is affine and its integrated divergence is independent of the input. This establishes a clear sufficient condition under which the divergence correction can be omitted without affecting sample rankings.

Critical weaknesses

1. Insufficient explanation of the impact of divergence correction on image-level anomaly detection.

Although incorporating the divergence correction improves pixel-level anomaly localization, it leads to decreased image-level detection performance in some settings. As shown in Table 1, image-level AUROC decreases from 96.23% to 95.57% on VisA and from 98.64% to 98.48% on MVTec AD, despite improvements in P-AUPRO. In contrast, Table 2 reports an improvement in image-level AUROC when using DINOv3 features.

The paper does not sufficiently explain why the divergence correction has different effects on image-level detection and pixel-level localization, or why its effect varies across feature representations. Since this task-dependent behavior is one of the paper's main findings, the lack of a deeper explanation limits the interpretation of the experimental results.

2. Insufficient analysis of the reduced marginal benefit of divergence correction after multi-layer fusion.

Table 3 shows that the P-AUPRO improvement from incorporating the divergence correction decreases from 4.87 percentage points in the single-layer setting to only 1.15 percentage points after five-layer fusion.

However, the endpoint-only baseline already improves substantially after multi-layer fusion, from 88.37% to 94.04% P-AUPRO. Therefore, the reduced marginal gain may be partly explained by the higher baseline performance and the smaller remaining room for improvement, rather than necessarily indicating that the divergence correction provides less useful information in the multi-layer setting.

The current analysis does not sufficiently distinguish this possibility from other explanations, such as information redundancy across feature layers or the influence of the fusion mechanism itself. This limits the strength of the paper's conclusions regarding the dependence of divergence correction benefits on layer configuration.

Questions for authors

1. Could the authors further explain why incorporating the divergence correction improves pixel-level anomaly localization but decreases image-level detection performance in some settings? In particular, why does its effect on image-level AUROC differ between C-RADIO and DINOv3 features?

2. Could the authors clarify why the marginal benefit of divergence correction decreases substantially after multi-layer fusion? To what extent is this reduction attributable to the higher endpoint-only baseline performance and the limited remaining room for improvement, rather than information redundancy across layers or the fusion mechanism itself?

1 comment
AuthorsOpenReview ↗

Response to Reviewer isuw

Thank you for highlighting the need to explain these two effects. We have expanded Sections 5.2–5.3 with analyses of image-max aggregation and shared endpoint errors.

  1. Why localization and image detection can move in opposite directions.

To address your first question, let pE(i)p_E(i) be the endpoint's maximum-scoring location in image i, and define ri=max⁡p(Ei+Vi)−[Ei(pE(i))+Vi(pE(i))]≥0r_i=\max_p(E_i+V_i)-[E_i(p_E(i))+V_i(p_E(i))]\ge0. The full image score is exactly

sF(i)=sE(i)+Vi(pE(i))+ri+constant.s_F(i)=s_E(i)+V_i(p_E(i))+r_i+\text{constant}.

An anomalous/normal pair's detection-margin change equals the difference of V at the old maxima plus ranomalous−rnormalr_{\mathrm{anomalous}}-r_{\mathrm{normal}}. Each rir_i is nonnegative, but its difference need not be. For endpoint margin mm and added margin dd, a wrong pair is repaired exactly when m<0m<0 and d>−md>-m; a correct pair is harmed when m>0m>0 and d<−md<-m. A correction can improve the ordering of many defect pixels while leaving a high normal-image peak, or increasing that peak relative to anomalous-image maxima. Pixel/region localization and image detection consequently impose different ranking requirements.

We tested this decomposition on all 12 VisA categories by fixing the endpoint maximum, adding V there, and then allowing the maximum to move. Image AUROC (%) follows:

Backbonemax⁡E\max E(E+V)(E+V) at pEp_Emax⁡(E+V)\max(E+V)
C-RADIO L3196.26695.26495.613
DINOv391.43391.05292.674

At the old maxima, V hurts detection ordering in both cohorts. Reselecting the maximum recovers 0.349 points for C-RADIO and 1.622 for DINOv3; only the latter exceeds the preceding loss. This explains the opposite net changes at the scoring level: detection depends on the balance between fixed-location corrections and spatial reselection. Moreover, the maximum's frequency of falling inside a defect mask increases in both cohorts (59.25% to 60.42%; 55.25% to 58.92%), so improved spatial placement alone does not ensure better normal/anomalous image ordering. Appendix I.1 analyzes all 96,200 within-category pairs per cohort (the same 2,162 images, not new independent samples), with no ties. For E to full, category-macro repaired/harmed rates are 1.1973%/1.8505% for C-RADIO and 4.1801%/2.9384% for DINOv3. Their differences exactly recover the AUROC changes (-0.6532/+1.2417 points): the opposite net signs reflect which crossings are repaired versus harmed, not merely how often scores change. The C-RADIO diagnostic uses layer 31 of the fusion study, distinct from the separately trained three-seed models of Table 1.

Together, these analyses explain the contrast through image-max aggregation (Section 5.2), while the origins of the map differences in pretraining or feature geometry remain open.

  1. Why the correction's marginal gain shrinks after fusion.

As you point out, fusion strengthens the endpoint and leaves fewer errors to repair. Section 5.3 therefore separates this effect from losses of previously correct detections. At matched 5% pixel FPR, define repaired coverage as endpoint-missed foreground recovered by the full score, and lost coverage as endpoint-detected foreground missed by the full score. With the same region weights, the net PRO change equals repaired minus lost coverage exactly. Across all 12 categories:

LayersEndpoint PRORepairedLostFull PRO
182.177.571.0888.66
590.883.471.6092.76

PRO here is region coverage at this operating point; repair/loss are percentage points, not integrated AUPRO. Fusion reduces the endpoint's missed region mass from 17.83% to 9.12%. The smaller net increment contains both 4.10 points less repaired mass and 0.52 points more lost mass. For example, five-layer macaroni1 repairs 0.17 points but loses 0.92, yielding a negative net effect. Thus a smaller gain does not by itself show that V has become uninformative or redundant: there are fewer errors available to repair, and rescoring/rethresholding can also lose endpoint successes.

Holding residual errors fixed, on foreground missed by both endpoints, 1L and 5L recover 20.14% and 41.62%, respectively, when recovery is first computed within each region containing common misses and then averaged over regions and categories. Pooling pixels within each category instead gives 28.29% and 42.14%; Table A30 reports both and per-category support. The correction thus recovers more of the same missed foreground after fusion: the smaller aggregate gain does not imply uniformly weaker correction on these shared errors. This conditional comparison excludes losses of endpoint successes and is not the overall PRO increment.

Appendix I.2 gives the accounting, thresholds and achieved FPRs for each category. These diagnostics use test masks and the original fusion scales. They describe region coverage at 5% FPR, not integrated P-AUPRO, and do not identify inter-layer redundancy as a cause or rule out every ceiling effect.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.