acceptodds
Under review as a conference paper at ICLR 2027

On Uniformly Scaling Flows: A Density-Aligned Approach to Deep One-Class Classification

Abstract

Unsupervised anomaly detection is commonly approached through two major paradigms: deep one-class methods learn compact representations of nominal data, whereas deep density estimators such as normalizing flows model exact likelihoods. We show that uniformly scaling flows (USFs), normalizing flows with a constant Jacobian determinant, provide a principled bridge between these paradigms. Specifically, we prove that maximum-likelihood training of a USF reduces to a one-class objective whose inherent regularization mechanism prevents representational collapse and incentivizes density-aligned latent representations. This connection implies that USFs inherit both the density faithfulness of flows and the distance-based reasoning of one-class methods: their negative log-likelihood is monotonic in the latent norm, whereas in non-USFs the Jacobian term can entangle likelihood with local volume changes. Consequently, we advocate USFs as drop‑in replacements for non‑USFs in modern anomaly detection architectures and identify mechanisms that recover expressivity and enhance numerical stability. Across tabular and image benchmarks, this substitution simplifies optimization, markedly reduces run-to-run variability, and consistently matches or improves detection quality, while mitigating severe failure modes. Thus, our results unify two central perspectives on anomaly detection, indicating that normalizing flow-based methods benefit more from density-faithful geometry than from greater nominal expressivity.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
normalizing flows, one-class classification, anomaly detection, representation learning

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

This work investigates the theoretical connection between uniformly scaling flows (USFs) and deep one-class classification, and claims that maximum-likelihood training of a USF reduces to a one-class objective with implicit volume regularization. The authors further argue that USFs inherently guarantee density-aligned latent representations, and demonstrate improved training stability and competitive detection performance across tabular and image anomaly detection benchmarks.

Critical strengths
  1. The paper provides a clean and step-by-step formal derivation of the equivalence between USF maximum likelihood estimation and the deep one-class objective, which offers a useful reference for understanding the relationship between the two paradigms.
  2. The empirical evaluation covers both tabular and image modalities, includes clean/contaminated training settings and pixel-/image-level evaluation, and reports extensive ablation studies.
Critical weaknesses
  1. Limited theoretical novelty, with worst-case constructions that lack real-world relevance. Uniformly scaling flows are not a new concept, dating back to NICE. Constant-Jacobian flows for anomaly detection were already explored by OneFlow. The core claimed theoretical contribution that MLE training of USFs reduces to a one-class distance objective is a straightforward algebraic consequence of the change-of-variables formula under a constant Jacobian. This connection has been implicitly understood in the field and lacks technical depth; formalizing it does not constitute a substantial theoretical advance.
  2. Unfair baseline comparisons and insufficient SOTA context. The stability advantage of USFs is tested against a weakly optimized non-USF baseline. The paper only applies simple tanh clamping to affine coupling layers, ignoring the large body of established stabilization techniques for normalizing flows. Without thorough hyperparameter tuning for the non-USF side, the conclusion that USFs are more stable is unconvincing. Also, the paper lacks comparison to current state-of-the-art anomaly detection methods.
  3. The paper claims to unify two central perspectives on anomaly detection, but the connection only holds for the narrow subclass of USFs, which have far weaker expressivity than general normalizing flows. This framing exaggerates the scope of the work; it does not unify the two broader paradigms.
  4. The repeated claim that density alignment is more valuable than expressivity is not universally supported by the authors' own data. On multiple tabular datasets, USFs underperform non-USFs by large margins. The paper only reports aggregate averages and does not analyze these failure cases, nor does it define boundary conditions for when USFs help or hurt.
  5. The paper fails to clearly delineate its contribution relative to OneFlow, the most closely related work. OneFlow also uses constant-Jacobian flows for one-class anomaly detection with a minimum-volume objective. The only core difference is that this work uses MLE instead of a custom minimum-volume loss, but the paper never demonstrates that MLE is superior, nor does it provide direct head-to-head comparison with OneFlow.
Questions for authors

Please see Weaknesses

3 comments

W1 (theoretical novelty; NICE; OneFlow; worst-case constructions)

We do not claim that USFs are new; the paper introduces them with reference to NICE (Section 2), and we agree that the derivation is short. Our contribution is its consequence: MLE training of a USF is a one-class objective with built-in collapse prevention, and USFs are exactly the flows for which this holds (G3). This yields a density-alignment guarantee that non-USFs and Deep SVDD provably lack (Proposition 1).

Regarding "implicitly understood": the literature reports the empirical advantage of constant-Jacobian flows without explaining it. OneFlow finds its constant-Jacobian variant stronger than the general one; recent tabular evidence highlights likelihood-based detection with USFs (Kim et al., 2026); Behrmann et al. (2021) give stability bounds favouring additive coupling. None of these attribute the advantage to the density alignment of the latent geometry. Moreover, established flow-based detectors (FastFlow, CFlow, U-Flow) all default to affine coupling, suggesting that this connection may not yet be widely reflected and its implication considered in detector design. We would welcome references to prior work that states it explicitly and would be happy to cite them.

Worst-case constructions. We agree that they are not realistic data-generating processes, as the paper says (Remark 1, Limitations). They show that the setups of Deep SVDD and non-USFs provide no guarantee of alignment, not that misalignment is typical. That it arises in benign settings is shown by Fig. 3a and the Gaussian-mixture experiment in G2. OneFlow is addressed under W5.

W2 (non-USF baseline strength; stabilization techniques; SOTA comparison)

We respectfully note that the non-USF is not examined under tanh clamping alone. The paper evaluates several established stabilization techniques:

  • Normalization layers. ActNorm and spectral normalization, for both models on all 20 datasets (App. C.4, Table 9; Section 5.1). Neither closes the gap to the base USF; both widen it. With ActNorm, the non-USF remains 2.302.30 (clamp 1.0) and 3.143.14 (clamp 2.0) points below; with spectral normalization, 7.107.10 and 16.9416.94, and at clamp 2.0, 10 of 60 runs fall below 50% AUC-ROC, none for any USF configuration.
  • Clamp strength. Fig. 3b sweeps the log-scale clamp. The non-USF improves almost monotonically as the clamp is tightened, and the zero-clamp endpoint is the uniformly scaling limit. Even the best non-zero clamp (1.0, Table 9) stays below the USF (87.47±3.3087.47\pm3.30 vs. 88.28±1.4988.28\pm1.49), with more than twice the run-to-run variability. Clamp 2.0 in the main results is the default of common implementations (Akcay et al., 2022).
  • Feature mixing. Both models use Householder reflections, because LU-parameterized mixing costs the non-USF 26.6 points vs. 0.57 for the USF (App. E).

Tuning. USF, non-USF and OneFlow share one configuration of identical capacity (Section 5.1); the study tests a mechanism under controlled conditions and is not a leaderboard comparison (Limitations). Where tuning matters most for affine flows (clamp, mixing), it was done in favour of the non-USF. In the Gaussian-mixture experiments, both models are selected by the same hyperparameter search; on images, we use the hyperparameters of the original authors for the affine models (G4).

SOTA context. On tabular data, the comparison includes recent detectors, DOHSC (ICLR 2024), DASVDD (TKDE 2024) and URD (NeurIPS 2025), alongside Deep SVDD and OneFlow; the USF attains the best mean AUC-ROC and mean rank in both regimes (Table 1). On images, the experiments test a drop-in substitution; we make no state-of-the-art claim there and will add references to recent diffusion-based and vision–language-based detectors (G4).

W3 (scope of the "unification" claim; USF expressivity)

We agree that "unify" overstates the scope and have replaced it by "connect". As detailed in G3, the MLE objective of a flow is a Deep-SVDD-type objective, i.e., a latent-distance term plus a term independent of xx, if and only if the flow is a USF. The connection is therefore specific to USFs by necessity, not by choice. We agree that USFs with a Gaussian base are less expressive; learnable radial bases keep the constant Jacobian and lower the held-out NLL below that of the non-USF on 17/20 and 19/20 datasets (now Section 5.1, Table 2b).

W4 (density alignment vs. expressivity; tabular failure cases; boundary conditions)

The paper does not report aggregate averages only. App. C gives per-dataset AUC-ROC for all eight methods in both regimes (Tables 5 and 6); Table 8 reports paired per-dataset effects of the learnable radial base against the Gaussian-base USF and the non-USF, with the number of seeds with a positive effect; App. D gives per-class image results.

We agree that the USF does not win everywhere; on some datasets (e.g., NSL-KDD, clean) it trails the non-USF clearly. We do not claim that alignment always outweighs expressivity. The claim is suite-level: the USF wins on 17/20 (clean) and 14/20 (contaminated) datasets against the non-USF; the clean difference is significant (Wilcoxon p=0.0056p=0.0056), the contaminated one is not (p=0.11p=0.11). Section 5.1 now states this, including the win/loss counts.

On boundary conditions, the theory makes a clear prediction. A USF must realize all density variation by transport under constant volume change. It is therefore at a disadvantage when the nominal data require strongly varying local volume change, e.g., components of very different scale. A controlled experiment in the revised paper confirms this: with anomalies just outside the narrow component of the Gaussian mixture (G2), the USF falls clearly behind the non-USF in high dimension (89.589.5 vs. 98.298.2 AUC-ROC at d=128d=128), and a learnable radial base does not close the gap (App. F; stated as a limitation in App. H). We will add a per-dataset failure analysis of the tabular results.

W5 (delineation from OneFlow; MLE vs. minimum-volume loss; head-to-head comparison)

A head-to-head comparison is part of the paper. OneFlow and its non-constant-Jacobian variant OneFlow-Gen are baselines in all tabular experiments (Table 1; per dataset in Tables 5 and 6), trained with the same coupling architecture and budget as USF and non-USF (Section 5.1, App. C). The comparison thus isolates the objective: maximum likelihood vs. minimal volume. The USF achieves a higher mean AUC-ROC (88.388.3 vs. 87.987.9 clean; 75.775.7 vs. 74.974.9 contaminated) and a better mean rank (3.153.15 vs. 3.203.20; 3.053.05 vs. 4.254.25). The gap between OneFlow and OneFlow-Gen mirrors that between USF and non-USF (Section 5.1), consistent with the constant Jacobian being the main driver in both families.

Conceptual delineation. The difference is not only the loss. OneFlow learns a minimum-volume hypersphere through a quantile-based criterion: only the points inside the target quantile shape the objective, and the model is not trained as a density estimator. Nothing ties the ordering by latent distance to density outside the sphere, which is where anomalies lie. MLE training of a USF yields p(x)∝pB(ϕ(x))p(x) \propto p_B(\phi(x)) on the whole space (Proposition 1(c)), so the alignment holds everywhere, including the tails. Both ingredients are needed: the constant Jacobian (without it, alignment fails, Proposition 1(b)) and the likelihood objective (without it, there is no density to align with). In practice, USFs need no custom objective and can replace affine flows inside existing likelihood-based detectors, as the image experiments show. We do not claim that MLE outperforms OneFlow's criterion; the margins are small. Our claim is equal-or-better performance together with a guarantee and drop-in compatibility.

Thanks for your response. However, I think the main flaw has not been fully addressed in this rebuttal. Therefore, I want to keep the original score.

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

Authors argue that constraining a normalizing flow to expand and contract space at the same rate everywhere yields anomaly detectors that are more stable and more interpretable. To this end, they unify two major lines of deep anomaly detection research. One-class methods such as Deep SVDD map normal data close to a single center in latent space and use the distance from that center as the anomaly score, whereas normalizing flows estimate the data density directly and flag samples with low likelihood as anomalous. However, the former are prone to representational collapse, in which all embeddings converge to a single point, while in the latter the input-dependent Jacobian term can decouple likelihood from latent distance and make training numerically unstable. To address these limitations, the paper introduces uniformly scaling flows (USFs), whose Jacobian determinant is constant across inputs, as a theoretical bridge between the two paradigms. Specifically, the authors show that maximum-likelihood training of a USF decomposes into a center-distance term and a regularization term that prevents volume collapse, mirroring the Deep SVDD objective, so that the model likelihood decreases monotonically with distance from the latent center. Building on this result, they propose a drop-in modification that replaces affine coupling with additive coupling in existing flow-based detectors such as FastFlow, CFlow, and U-Flow, and demonstrate on tabular and image benchmarks that it maintains or improves detection performance while substantially reducing run-to-run variability.

Critical strengths

S1. The proposed modification is simple and practical, as it only replaces affine coupling with additive coupling. It can therefore be applied as a drop-in change to existing flow-based detectors such as FastFlow, CFlow, and U-Flow without altering other components.

S2. By showing that maximum-likelihood training of a USF reduces to a Deep SVDD–like objective, the paper provides a novel and intuitive theoretical link between one-class and density-based anomaly detection.

S3. The presented experimental results show consistent improvements in training stability. USFs substantially reduce run-to-run variance across tabular and image benchmarks and eliminate severe failure cases of affine baselines, such as CFlow on MVTec metal_nut and toothbrush. This is further supported by the clamp sensitivity analysis in Fig. 3b and the normalization experiments.

Critical weaknesses

W1. The connection between the theoretical analysis and the empirical gains is weak. In practice, the non-USF is already nearly aligned (Fig. 3a, r = 0.996), and USF achieves higher AUC-ROC despite a worse NLL (Table 7). This suggests that the performance gains stem primarily from improved training stability rather than from density alignment. Moreover, the restriction that USFs can only represent unimodal densities is not discussed.

W2. The claim of improved training stability is supported only by the variance of the final scores over three seeds. Direct evidence, such as training loss curves, gradient norms, or the evolution of the log-determinant during training, is lacking.

W3. The image experiments compare only untuned flow-based models, whose performance falls well below the originally reported results (e.g., 72.5 for CFlow on MVTec AD). Even if improvements are observed, comparisons with and citations of recent state-of-the-art methods, including diffusion-based and multimodal (e.g., vision–language) approaches, are insufficient.

W4. Failure cases (e.g., FastFlow on cable and CFlow on grid) are not analyzed, and it would be worth discussing region-wise approaches as a potential remedy. In addition, neither qualitative anomaly maps nor the PRO metric is provided.

Questions for authors

Q1. Given Fig. 3a and Table 7, it is unclear how much of the performance gain comes from density alignment rather than training stability. It would also be helpful to know how USFs perform on multimodal normal data.

Q2. Showing the training dynamics, such as loss curves and the evolution of the log-determinant, together with results over more seeds, would make the stability claim more convincing.

Q3. The paper would benefit from comparisons with properly tuned baselines and recent SOTA methods. An explanation of the performance drops on classes such as cable and grid is also needed.

3 comments

W1 / Q1 (theory vs. empirical gains; Fig. 3a, Table 7; unimodality)

We address this in G1 and G2. The non-USF is less aligned than r=0.996r=0.996 suggests (over individual fits, 0.9210.921; G1), and the performance effects are moderate. Alignment and stability are not competing explanations. Both follow from the input-dependent log-determinant, and the alignment view explains geometrically why removing it helps detection. Table 7 supports this view: better nominal NLL with worse AUC (at least 11 of 20 datasets at clamp 1.0) is the decoupling of likelihood and latent geometry that Proposition 1 predicts. With known density (G2), the latent geometry of the non-USF is not density-faithful and degrades under contamination, while that of the USF is exact. We have revised the presentation accordingly (Section 4.4).

Unimodality: we agree and have added it to the limitations (App. H). Learnable radial bases (now Section 5.1, Table 2b) lower the held-out NLL of the USF below that of the non-USF on 17/20 (clamp 1.0) and 19/20 (clamp 2.0) datasets while keeping the constant Jacobian (see MHV2, Q1).

W2 / Q2 (direct evidence for stability; training dynamics; more seeds)

Beyond the three-seed variance, the paper contains more direct evidence:

  • Collapsed runs. With LU-parameterized mixing, the non-USF loses 26.6 AUC-ROC points on average vs. 0.57 for the USF (App. E, Table 14); hence Householder mixing for the non-USF in the main results. With spectral normalization at clamp 2.0, 10 of 60 non-USF runs fall below 50% AUC-ROC, none for any USF configuration (App. C.4). On images, affine U-Flow seeds collapse on individual classes (e.g., metal_nut, 82.7±29.982.7\pm29.9 vs. 100.0±0.0100.0\pm0.0 for the USF variant; Table 13); the USF variants never do. In the configuration search of the Gaussian-mixture experiment (G2), 14 of 30 non-USF trials for d≥8d\geq8 diverge in the first epoch, none of the USF trials with identical configurations (App. F).
  • Mechanism. App. E traces the failures to locally exploding log-determinants, which compound across layers (Section 4.2).
  • Statistics. In the clean regime, the gap to the non-USF is significant (Wilcoxon signed-rank p=0.0056p=0.0056, 17/20 wins) at less than half the seed variability; in the contaminated regime, it is not (p=0.11p=0.11, 14/20). Both are now reported in Section 5.1.

New: training dynamics. We retrained USF and non-USF on five tabular datasets (Arrhythmia, Optdigits, Pendigits, Campaign, NSL-KDD) with the configuration of the paper over ten seeds, logging per epoch the training loss, gradient norms, test AUC-ROC, and the spread of log⁡∣det⁡J∣\log|\det J| over 1000 fixed training points (new Fig. 4 in Section 5.1; full results in App. E). In the clean regime:

  • On Arrhythmia, Optdigits and Pendigits, the final AUC-ROC of the non-USF varies 3–9$\times$ more across seeds (std 6.2/6.2/7.36.2/6.2/7.3 vs. 2.1/1.1/0.82.1/1.1/0.8), its test AUC-ROC changes 2–6$\times$ more between epochs (mean absolute change 3.8/4.4/2.53.8/4.4/2.5 vs. 0.6/1.1/1.40.6/1.1/1.4 points), and gradient spikes (maximum >10×>10\times median within an epoch) occur in 5–10% of epochs vs. at most 1%. Single non-USF seeds drop permanently to 35–60% AUC-ROC; no USF seed does.
  • The spread of log⁡∣det⁡J∣\log|\det J| stays at 3–55 nats for the non-USF and is exactly zero for the USF. This is the input-dependent term that G1 identifies as the common source of instability and misalignment.
  • The effect is not universal: on Campaign, both models behave alike; on NSL-KDD, the non-USF is stable and better (92.792.7 vs. 86.786.7), consistent with the moderate average effect.

In the contaminated regime, both models overfit the anomalies similarly, so the curves do not discriminate (shown in the appendix). The runs reproduce the three-seed means of the paper (e.g., Arrhythmia, seeds 42–44: 70.1/79.370.1/79.3 vs. 69.0/78.969.0/78.9 reported).

W3 / Q3a (untuned baselines; SOTA comparison and citations)

On tabular data, the comparison focuses on the most recent methodologically related approaches, namely, recent one-class and flow-based (hybrid) detectors: DOHSC (ICLR 2024), DASVDD (TKDE 2024) and URD (NeurIPS 2025), alongside Deep SVDD and OneFlow. Against these, the USF attains the best mean AUC-ROC and the best mean rank in both regimes (Table 1). On images (G4), the experiments test a drop-in substitution with the hyperparameters of the original authors for the affine models; the CFlow numbers are affected by a per-batch normalization in the Anomalib evaluation, which we found while preparing this response, and are being re-evaluated. We make no state-of-the-art claim on images, but rather argue for the pair-wise advantages of using USF-based backbones in flow-based anomaly detection methods. We will add references to recent diffusion-based and vision–language-based detectors.

W4 / Q3b (failure cases cable/grid; region-wise approaches; anomaly maps; PRO)

Please see G4. The drops on cable (FastFlow) and grid (CFlow) occur at image level only. At pixel level, the USF variant is better on cable (94.9→97.294.9 \to 97.2, the best of all six models) and nearly unchanged on grid (97.4→96.797.4 \to 96.7). For cable (FastFlow), the anomalies are thus localized correctly, and the loss arises in aggregating the anomaly map into an image score; grid (CFlow) will be re-assessed after the CFlow re-evaluation. This makes region-wise approaches, as you suggest, a natural remedy, and we will discuss this in the paper. We report pixel-level AUC-ROC for all classes (App. D) and agree that PRO and qualitative maps would complement it.

Thank you for the detailed rebuttal. The ten-seed training-dynamics results resolve W2, and the known-density experiment (G2) partially resolves W1, though the relative contributions of alignment and stability remain unclear. But, the corrected CFlow results are still pending, the gain under contamination is not significant, and the unimodality limitation, PRO metric, and qualitative maps remain unaddressed. Therefore, I will therefore keep my original score.

Reviewer_MHV2OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes uniformly scaling flows (USFs) for anomaly detection. The main idea is to use a constant Jacobian determinant to make latent distance consistent with the learned density. The authors provide theoretical analysis and experiments on tabular and image datasets.

Critical strengths
  1. The motivation is clear, and the theoretical connection between latent distance and likelihood is interesting.
  2. The method can be applied to existing flow-based methods such as FastFlow, U-Flow and CFlow without changing the overall framework.
  3. Some experiments show clear improvements in training stability. The appendix also provides useful ablation results.
Critical weaknesses
  1. Limited theoretical guarantee. The theory mainly guarantees alignment with the model's learned density, not the true data distribution. I am not sure how well this works for more complex or multimodal normal data.
  2. The source of improvement is unclear. It is difficult to tell whether the gains come from density alignment or simply better training stability. The current experiments do not clearly separate these effects.
  3. Baseline comparison is questionable. The reported CFlow baseline achieves only 72.5% AUROC on MVTec AD, while the original paper reports 98.26%. Although the settings may differ, this gap is quite large and needs explanation.
  4. Improvements are not always consistent. The average gains on tabular datasets are quite small (around 0.4–0.5%), while some image classes show large performance drops, such as cable and grid. These results make the overall effectiveness less convincing.
Questions for authors
  1. Can the authors explain how the method handles multimodal normal data? Does the theoretical guarantee still hold with a more flexible base distribution?
  2. How can we know that the improvement comes from density alignment rather than training stability? A more direct ablation would be helpful.
  3. Why is the CFlow baseline much lower than the original result? Also, could the authors explain the large performance drops on some image classes?
2 comments

W1 / Q1 (theoretical guarantee; multimodal normal data)

The guarantee indeed concerns the learned density, as the paper states (Section 4.3, after Proposition 1: agreement with the true density additionally depends on the quality of the fit). This is the strongest guarantee a likelihood-based detector can offer, and non-USFs and Deep SVDD lack even this one: a non-USF can fit pXp_X well and still place high-density points far from the latent centre (Item (b)).

More flexible bases. Item (c) does not depend on the base: for any USF that maps pXp_X to the base pBp_B, we have p(x)∝pB(ϕ(x))p(x) \propto p_B(\phi(x)). The model's density ordering is thus the base-density ordering of the latent code, for every base. For radial monotone bases, this is the ordering by latent norm. For a learned radial base with a non-monotone norm density, the ordering is still exact, but by base density rather than by norm alone.

Multimodal normal data. With a Gaussian base, a USF inherits the unimodality of the base; modes can only be connected by (arbitrarily thin) bridges (Draxler et al., 2024). The revised limitations state this explicitly (App. H), and the Gaussian mixture of G2 shows that a USF fits bimodal nominal data well (Section 4.3). The paper also evaluates the remedy, learnable radial bases (Section 4.4; Zaid et al., 2024), now in the main text (Section 5.1, Table 2b; details in App. C.3). A radial base whose norm distribution is a mixture of 20 Gamma distributions lowers held-out nominal NLL relative to the Gaussian-base USF (median 1.25 per dimension, interquartile range [0.47,1.81][0.47, 1.81]) and below the non-USF on 17/20 (clamp 1.0) and 19/20 (clamp 2.0) datasets, while keeping the constant Jacobian. Detection improves on average (+0.84 AUC-ROC points), though not uniformly.

W2 / Q2 (density alignment vs. training stability)

Please see G1 and G2. In short: both effects stem from the same input-dependent log-determinant, so no ablation can remove one while keeping the other; the alignment view is the geometric account of why removing it helps detection. Its predictions are borne out by Fig. 3a (12/20 non-USF datasets and 38/60 individual fits off the diagonal), the clamp sweep in Fig. 3b, and Tables 7/9 (at least 11 datasets where the non-USF fits the nominal density better yet detects worse). In a controlled experiment with known density (G2), the latent geometry is density-faithful only for the USF, and the latent radius of the non-USF loses discriminative power under contamination.

W3 / Q3a (CFlow baseline 72.5% vs. 98.26%)

Please see G4. While preparing this response, we found that the Anomalib v2.0.0 implementation of CFlow normalizes the scores per test batch, whereas the original implementation normalizes over the whole test set. Scores from different batches are therefore not comparable, although AUC-ROC is computed over the test set. This affects both CFlow variants and plausibly explains much of the gap to 98.2698.26, including the collapsed seeds (e.g., metal_nut 32.6±42.632.6\pm42.6). FastFlow and U-Flow are not affected. Otherwise, all image models run in the identical pipeline with the hyperparameters of the original papers and no tuning. We are re-evaluating the trained CFlow models (inference only) and will report corrected numbers; we will also make the protocol explicit in the paper.

W4 / Q3b (small tabular gains; drops on cable and grid)

Tabular gains. The 0.4–0.5 points are the margin over the best competing method (OneFlow clean, Deep SVDD contaminated). Against the directly comparable non-USF, the gain is +2.2+2.2 (clean) and +1.0+1.0 (contaminated) points (Section 5.1). Over the 20 datasets, the USF wins on 17/20 in the clean regime (Wilcoxon signed-rank p=0.0056p=0.0056), with less than half the run-to-run variability (mean std 1.51.5 vs. 3.93.9). In the contaminated regime, it wins on 14/20, but the difference is not significant (p=0.11p=0.11). Section 5.1 now reports these statistics.

Cable and grid. Please see G4. The drops occur at image level only; at pixel level, the USF variant is better on cable (94.9→97.294.9 \to 97.2, the best of all six models) and nearly unchanged on grid (97.4→96.797.4 \to 96.7). For cable, localization works and the loss arises when the anomaly map is aggregated into an image score; grid (CFlow) will be re-assessed after the re-evaluation. We will discuss these classes explicitly in the paper.

Thanks for the detailed response. The rebuttal addresses most of my concerns and provides some useful additional evidence. I find the paper more convincing after these revisions.

Regarding W1, the clarification of the theoretical guarantee is helpful. I agree that the guarantee is about the learned density, not the true data distribution. The additional experiments with learned radial bases also help address my concern about multimodal data. However, I still think the limitation on modeling general multimodal distributions should be made clear. A more flexible radial base helps, but it does not necessarily solve the problem for arbitrary multimodal distributions.

For W2, the new G1/G2 results and the clamp experiments make the density-alignment argument more convincing. The controlled experiment with known density is particularly useful. I understand that density alignment and training stability are related to the same input-dependent Jacobian term, so it is difficult to fully separate their effects. The new results provide better support for the authors' explanation, although the contribution of each effect is still not completely clear.

For W3, finding the score normalization issue in the Anomalib implementation is important and could explain much of the CFlow performance gap. I appreciate that the authors checked this issue and are re-evaluating the trained models. However, the corrected results are not available yet, so this concern is not fully resolved. I would like to see the updated results and the corresponding evaluation protocol in the revised paper.

For W4, the additional statistical results are helpful. The improvement over the directly comparable non-USF baseline is larger than the margin over the best competing method. The clean setting is also supported by the Wilcoxon test, while the improvement under contamination is not statistically significant. The explanation for the cable results is reasonable, since the pixel-level performance actually improves and the main drop comes from image-level score aggregation. The grid result still needs to be checked after the CFlow re-evaluation.

One minor point is that the distinction between preventing exact collapse and ensuring good conditioning is still worth clarifying in the paper. A constant Jacobian determinant alone does not guarantee well-conditioned transformations.

Overall,thanks for the detailed response. The additional explanations and experiments address several of my concerns, especially regarding the theoretical guarantee and density alignment. However, some issues remain. The corrected CFlow results are still pending, and the improvement under contaminated data is not statistically significant. I also think some limitations regarding multimodal distributions and the theoretical guarantees still need further clarification. Overall, the rebuttal helps improve the paper, but I do not think the remaining concerns have been sufficiently resolved to change my initial assessment. Therefore, I will maintain my original rating of 2 (Weak Reject).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.