acceptodds
Under review as a conference paper at ICLR 2027

Diffusion Models are Secretly One-step Anomaly Detectors

32 views

Abstract

Diffusion-based anomaly detection (AD) commonly relies on iterative denoising to obtain high-quality reconstructions, following the intuition that better reconstruction leads to better AD. In this work, we revisit this intuition and ask whether iterative refinement is truly beneficial for AD. Through a large-scale empirical study across diverse AD datasets and diffusion designs, we uncover a striking phenomenon that we term *one-step optimality*: diffusion models trained solely on normal samples often achieve their best detection performance with just a single denoising step. To understand why, we theoretically analyze the optimal velocity field and show that iterative denoising can increasingly reconstruct anomalous information, leading to greater *anomaly leakage* and reduced separation between normal and anomaly scores. Under our analysis setting, one-step denoising minimizes such leakage and achieves optimal detection performance. This understanding naturally leads to the *One-step Diffusion anomaly Detector (ODD)*, which retains standard diffusion training but performs reconstruction with a single denoising step. Extensive experiments show that **ODD** achieves strong detection performance with substantially improved inference efficiency across diverse datasets and modalities, while also serving as a plug-and-play replacement for iterative denoising in existing diffusion-based AD methods. Code is available at: [https://anonymous.4open.science/r/odd](https://anonymous.4open.science/r/odd).

Primary area:
generative models
Keywords:
Diffusion Models, Anomaly Detection, Flow Matching

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_HQ84OpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

The authors investigate a family of diffusion models for the use as anomaly detectors. The mechanism used is the reconstruction quality of non-normal samples if they are reconstructed with a model that is only trained on a normal distribution. This is a commonly used mechanism for anomaly detection with generative models. In the family of the considered models, the reconstruction is an expensive operation as it usually needs hundreds of network evaluation per sample. The authors challenge this idea with the proposed one-step anomaly detection. They theoretically show that under mild assumptions the proposed one step method is optimal and corroborate the same with experiments across a diverse set of datasets and comparing methods.

Critical strengths

The paper doesn’t not propose a new architecture per se, but challenges the existing use of multiple timesteps of denoising for reconstruction in anomaly detection tasks. The solution of the one-step denoising anomaly detection is fundamentally a novel idea that the authors support with the theoretical as well as empirical results. This method is not only applicable to one specific model, but to basically all diffusion based models in the context of anomaly detection.

Critical weaknesses

Not critical weaknesses but things I’d like to bring to attention:

The git repository is empty/seems to be missing.
I was unable to follow all the proofs in detail, but the assumptions and general ideas seem reasonable. The proofs however are rather technical.

Questions for authors

I’d like to add a few minor things that I’d ask the authors to consider adressing:

  • The softmax notation is a little bit an abuse - I understand that it is more cumbersome to write out the full input/output vectors, but I think it might make it slightly clearer.

  • Many parts of the proofs and chains of equations are quite densely written. I’d like to encourage the authors to spend a little bit more “prose” to explain the intermediate step. Two examples I stumbled across: In proof of lemma A.5 the way z is defined comes a little bit out of the blue. What was the insight in defining it like this? Similarly e.g. on line 1128 the reasoning looks a bit weird, is it maybe an LLM artifact? Finally I’d also prefer the limits of the sums to be made explicit in (at least once per chain of equations), to make it easier to follow the reasoning.

5 comments
AuthorsOpenReview ↗

Response to Reviewer HQ84 (part1)

Thank you for the detailed review and constructive feedback. Here, we address each point below.

1. Code repository accessibility

The git repository is empty/seems to be missing.

Thank you for pointing out that the code repository was inaccessible during review. We have updated the repository’s visibility settings, and the code is now accessible at the following link: https://anonymous.4open.science/r/odd. We will publicly release the code upon acceptance.

2. Softmax notation

The softmax notation is a little bit an abuse - I understand that it is more cumbersome to write out the full input/output vectors, but I think it might make it slightly clearer.

We appreciate the reviewer’s suggestion to clarify the vector-to-vector mapping of the softmax function.

In the revised manuscript, we have modified the notation to explicitly specify both the input and output vectors. Specifically, we have revised the notation as follows:

Before revision (Lines 126–127):

:= \operatorname{softmax}\left( -\frac{\lVert\boldsymbol{x}-(1-t)\boldsymbol{x}^{(i)}\rVert_2^2}{2t^2} \right).$$ **After revision (Lines 126–127):** $$\left[\gamma_i(\boldsymbol{x},t)\right]_{i=1}^{n} := \operatorname{softmax}\left( \left[ -\frac{\lVert\boldsymbol{x}-(1-t)\boldsymbol{x}^{(j)}\rVert_2^2}{2t^2} \right]_{j=1}^{n} \right).$$ We have also added an accompanying definition of the indexed-vector notation, $$[a_i]_{i=1}^{m} := (a_1,\ldots,a_m)^\top,$$ to explicitly clarify this convention. ### 3. Clarity of proofs and derivations > Many parts of the proofs and chains of equations are quite densely written. I’d like to encourage the authors to spend a little bit more “prose” to explain the intermediate step. Two examples I stumbled across: In proof of lemma A.5 the way z is defined comes a little bit out of the blue. What was the insight in defining it like this? Similarly e.g. on line 1128 the reasoning looks a bit weird, is it maybe an LLM artifact? Finally I’d also prefer the limits of the sums to be made explicit in (at least once per chain of equations), to make it easier to follow the reasoning. Thanks for the valuable suggestion! To improve the clarity and readability of the derivations, we have revised the manuscript while preserving the original theoretical claims and their proof strategies. Specifically, we have made the following revisions: - **Added proof sketches:** We have included a proof sketch before each detailed proof to provide an overview of the main arguments and facilitate understanding of the subsequent derivations. - **Clarified summation indices:** We have explicitly specified the ranges of the summation indices to eliminate potential ambiguity. - **Clarified the proof of A.5:** We have expanded the corresponding proof sketch to provide additional context regarding the definition and role of $z$ in the derivation. - **Added intermediate derivation steps:** For the argument originally presented in L1128 (now L1177–1197 in the revised manuscript), we have included additional intermediate steps to improve clarity.
AuthorsOpenReview ↗

Response to Reviewer HQ84 (part2)

In particular, we address the following questions:

Why was zz defined that way in Lemma A.5?

Here, we briefly introduce its motivation (see Proof sketch of A.5 in detail):

To prove ℓΠ1<ℓΠK\ell_{\Pi_1} < \ell_{\Pi_K}, we analyze the anomaly leakage factor ℓΠK=(1−τ)∏k=0K−1(1−hkbtk).\ell_{\Pi_K} = (1-\tau)\prod_{k=0}^{K-1}(1-h_k b_{t_k}).

Define zλ(t):=(λ(1−t),t)⊤\boldsymbol{z}_\lambda(t) := (\sqrt{\lambda}(1-t),t)^\top. Then, each factor can be expressed as

= \frac{\lambda(1-t_k)(1-t_{k+1})+t_k t_{k+1}}{\lambda(1-t_k)^2+t_k^2} = \frac{\langle \boldsymbol{z}_\lambda(t_{k+1}),\boldsymbol{z}_\lambda(t_k)\rangle}{\lVert\boldsymbol{z}_\lambda(t_k)\rVert_2^2}.$$ Thus, each factor has a geometric interpretation as the inner product of consecutive vectors ($\boldsymbol{z}_\lambda(t_k), \boldsymbol{z}_\lambda(t_{k+1})$) normalized by the squared norm at $t_k$. > What justifies the reasoning at the original line 1128? Thank you for pointing this out. We clarify that the reasoning at L1128 follows from the strict monotonicity of the argument of $\boldsymbol{z}_\lambda(t)$ and the ordered denoising schedule. Specifically, for $\lambda > 0$ and $0 = t_K < t_{K-1} < \cdots < t_0 = \tau < 1$, define $$\phi(t):=\arg\boldsymbol{z}_\lambda(t) =\arctan\left(\frac{t}{\sqrt{\lambda}(1-t)}\right).$$ Since $$\phi'(t)=\frac{\sqrt{\lambda}}{\lambda(1-t)^2+t^2} > 0,$$ the argument is strictly increasing on $[0,1)$, implying $0 < \theta = \phi(\tau)-\phi(0) < \pi/2$. Consequently, each consecutive angle satisfies $$\theta_k =\angle\bigl(\boldsymbol{z}_\lambda(t_k),\boldsymbol{z}_\lambda(t_{k+1})\bigr) =\phi(t_k)-\phi(t_{k+1}) > 0.$$ Summing these angular differences yields the telescoping identity $$\sum_{k=0}^{K-1}\theta_k =\sum_{k=0}^{K-1}\bigl[\phi(t_k)-\phi(t_{k+1})\bigr] =\phi(\tau)-\phi(0)=\theta.$$ Thus, the monotonicity of the argument ensures that all consecutive angular differences are positive and additive. We have explicitly clarified this reasoning in the revised proof. We need to clarify that, in all theoretical derivations, **we did not use an LLM to write proofs or claims.** As we specified in the AI use disclosure, we only use an LLM for feedback and to review the correctness of the formulation. Finally, we hope that these revisions have improved the clarity and transparency of the mathematical arguments, making them more accessible and easier for readers to follow.
AuthorsOpenReview ↗

Gentle Reminder

Dear Reviewer HQ84,

Thank you again for your careful review, constructive feedback, and positive assessment of our work.

In our rebuttal, we addressed your comments on code accessibility and the clarity of the mathematical derivations. In the revised manuscript, we revised the notation for the softmax and summation and added a detailed proof sketch for each theorem.

As the discussion period closes on November 18, we would like to check whether our response has adequately addressed your concerns. If you have any questions, we would be happy to provide further clarification.

Best regards, The Authors of Submission 6795

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

The paper revisits the role of iterative denoising in diffusion-based anomaly detection. It identifies the alleged one-step optimality phenomenon. The paper analyzes this phenomenon through the MMSE property of one-step reconstruction and an analysis of anomaly leakage under stated assumptions. The analysis is further extended to learned velocities and other diffusion frameworks. Based on these analysis, a method called One-step Diffusion Anomaly Detector is proposed. Extensive experiments demonstrate ODD achieves competitive detection performance and substantial inference efficiency gains.

Critical strengths
  • The paper challenges the intuition that better reconstruction necessarily improves anomaly detection. It shows that iterative refinement can reconstruct anomalous information and reduce detection discriminability. The connection between one-step reconstruction, MMSE estimation, and anomaly leakage provides a good explanation under clearly stated conditions. The additional analysis from the viewpoint of learned velocities further supports this explanation. This combination of a useful empirical observation and a principled explanation is the main reason for my positive assessment.

  • ODD is straightforward to implement and does not require distillation or additional training modules. Its competitive accuracy and substantial inference acceleration make the finding practically valuable. In my opinion, the simplicity is a strength.

  • The study performs experiments on multiple datasets and data modalities, which convinces the empirical claim.

Critical weaknesses

I did not find critical weaknesses that would result in rejection. The following issues are minor:

  • Some connections between the claims and their practical scope could be explained further.

    • The supplementary material provides useful sufficient conditions and a learned-velocity extension. However, the connection between the approximate posterior conditions, the empirical velocity discrepancy, and the theoretical guarantee could be clearer.

    • Table 3 would benefit from a more precise description of what “+ODD” changes, particularly whether it only changes denoising steps or also changes the scoring procedure. This matters because Table 10 shows different behavior for scoring.

    • A brief account of run-to-run variability may help contextualize the smaller mAD improvements.

  • The image-based FPS results use batch size 64, whereas other modalities use batch size 1. This distinction, together with the components included in timing, should be stated more clearly.

Questions for authors

Please refer to the minor clarification points listed in Critical Weaknesses.

6 comments
AuthorsOpenReview ↗

Response to Reviewer QAvR (part1)

Thank you for the thoughtful review and constructive suggestions. We address each point below.

1. Connections between theory and practical scope

Some connections between the claims and their practical scope could be explained further.

The supplementary material provides useful sufficient conditions and a learned-velocity extension. However, the connection between the approximate posterior conditions, the empirical velocity discrepancy, and the theoretical guarantee could be clearer.

Thanks for the valuable suggestion. To connect the approximate posterior conditions and the theoretical guarantee, we extend our theoretical analysis of learned velocity to incorporate approximate posterior conditions, providing conceptual insights into how posterior concentration affects the resulting theoretical guarantee (see Corollary A.9).

Regarding the empirical velocity discrepancy (Fig. 9(c) in the original manuscript), Corollary A.9 shows that a smaller discrepancy ee between the learned velocity and the optimal velocity relaxes the sufficient condition for one-step optimality, allowing a larger posterior concentration error η\eta. While the empirical velocity discrepancy alone does not establish one-step optimality, as the sufficient condition also depends on posterior concentration and the ideal AUROC gap, it gains understanding of how the training progress of diffusion models may affect one-step optimality, underscoring the necessity of sufficient training for accurate one-step AD.

2. What changes in the +ODD evaluation?

Table 3 would benefit from a more precise description of what “+ODD” changes, particularly whether it only changes denoising steps or also changes the scoring procedure. This matters because Table 10 shows different behavior for scoring.

We appreciate the opportunity to clarify the plug-and-play experiments in Table 3. The meaning of “+ODD” depends on the original method:

  • VPOT already uses a reconstruction-based formulation compatible with ours, so its scoring procedure does not require the same adaptation.
  • InvAD originally uses inversion-based, reconstruction-free scoring, and ReFP-AD uses an energy-norm score. For their “+ODD” variants, we use one-step reconstruction and replace the original scoring with our anomaly score (Eq. 6).

Thus, “+ODD” is not simply a change in the number of denoising steps for every baseline; it can also change the scoring procedure. We have clarified this in the revised manuscript.

To investigate the source of performance gains (scoring vs. NFE), we provide additional ablation experiments that measure mAD values across different NFE setups using the original scoring metrics or our scoring (Eq. 6). The results are as below (also in Section D.2 in the revised supplementary material):

ScoringNFE = 1NFE = 3NFE = 4NFE = 8NFE = 20NFE = 40
Inversion72.383.7 (InvAD)—82.580.879.9
Reconstruction83.7 (InvAD + ODD)83.3—83.183.083.0
Energy79.7—78.478.778.778.7 (ReFP-AD)
Reconstruction80.2 (ReFP-AD + ODD)—78.979.279.279.2

Reducing NFE to 1 alone does not achieve the best accuracy across all evaluation settings with the same pre-trained model, and one-step denoising (+ODD) fully exploits the capability of the pre-trained diffusion models with theory-guaranteed one-step optimality, consistent with Table 10. Also, even with non-reconstruction energy-based scoring, one-step denoising still yields the best accuracy compared with multi-step counterparts.

AuthorsOpenReview ↗

Response to Reviewer QAvR (part2)

3. Run-to-run variability

A brief account of run-to-run variability may help contextualize the smaller mAD improvements.

Thank you for the suggestion. We have additionally trained ODD with both DiT-B and DiT-G on all image AD datasets in Table 1 using six random seeds and report the mean ± standard deviation of mAD in revised Tables 1 (compared with diffusion-based AD) and 11 (compared with non-diffusion-based AD).

DatasetMethodNFE (↓)Image AUROCImage APImage F1_maxPixel AUROCPixel APPixel F1_maxAU-PROmADFPS
MPDDDiAD (AAAI'24)1074.682.182.593.015.921.278.464.00.1
VPOT (ICML'24)100077.480.083.695.916.422.585.365.93.1
OmiAD (ICML'25)193.192.990.698.536.140.793.978.049.8
InvAD (CVPR'26)396.596.594.498.340.143.991.280.1120.0
ReFP-AD (ECCV'26)4074.676.379.397.528.731.290.268.3159.3
ODD (DiT-B)195.0 ± 1.1095.8 ± 1.4593.9 ± 1.7498.4 ± 0.1140.9 ± 1.5544.4 ± 1.1992.6 ± 0.1180.1 ± 0.94325.5
ODD (DiT-G)197.1 ± 0.5798.0 ± 0.3496.1 ± 0.5198.7 ± 0.1044.8 ± 0.8647.5 ± 0.6093.1 ± 0.2582.2 ± 0.36168.2
RealIADDiAD (AAAI'24)1075.666.469.988.02.97.158.152.60.1
VPOT (ICML'24)100064.361.163.193.711.117.975.155.22.6
OmiAD (ICML'25)190.389.383.399.034.440.493.175.752.1
InvAD (CVPR'26)383.877.175.697.122.631.687.767.990.5
ReFP-AD (ECCV'26)4072.668.968.195.112.020.078.359.3159.5
ODD (DiT-B)187.9 ± 0.2883.6 ± 0.4578.7 ± 0.2699.0 ± 0.0233.8 ± 0.3340.6 ± 0.2693.8 ± 0.1473.9 ± 0.23406.0
ODD (DiT-G)188.3 ± 0.1284.0 ± 0.3179.0 ± 0.1299.1 ± 0.0034.2 ± 0.1640.8 ± 0.1394.0 ± 0.0874.2 ± 0.13163.6
VisADiAD (AAAI'24)1086.888.385.196.026.133.075.270.10.1
VPOT (ICML'24)100068.275.275.192.713.819.966.158.73.2
DiffusionAD (TPAMI'25)192.594.488.897.637.642.792.178.132.3
OmiAD (ICML'25)195.496.692.098.839.343.689.479.335.3
InvAD (CVPR'26)396.997.293.799.139.243.192.980.374.1
ReFP-AD (ECCV'26)4088.990.786.598.436.341.188.775.8159.3
ODD (DiT-B)196.6 ± 0.1397.1 ± 0.1293.4 ± 0.3098.8 ± 0.0041.3 ± 0.2345.6 ± 0.1291.5 ± 0.1180.6 ± 0.12322.7
ODD (DiT-G)197.0 ± 0.1097.5 ± 0.0693.9 ± 0.2198.1 ± 0.0541.1 ± 0.1845.6 ± 0.1190.5 ± 0.0980.5 ± 0.00167.5
MVTecADDiAD (AAAI'24)1097.299.096.596.852.655.590.784.00.1
VPOT (ICML'24)100082.892.389.189.630.635.676.671.03.2
DiffusionAD (TPAMI'25)192.497.093.794.457.456.086.582.532.3
OmiAD (ICML'25)198.299.398.197.551.756.293.484.939.4
InvAD (CVPR'26)399.099.698.597.546.552.392.783.788.1
ReFP-AD (ECCV'26)4090.895.792.995.442.046.288.078.7160.7
ODD (DiT-B)198.7 ± 0.1799.5 ± 0.0798.5 ± 0.3997.7 ± 0.1251.8 ± 1.3155.6 ± 0.9993.5 ± 0.2385.1 ± 0.47324.0
ODD (DiT-G)198.6 ± 0.1899.4 ± 0.0798.6 ± 0.1297.8 ± 0.0053.7 ± 1.1556.7 ± 0.7494.0 ± 0.0685.5 ± 0.24164.7

As shown, ODD (DiT-G) achieves a higher mean mAD than previous state-of-the-art methods on MVTec AD, MPDD, and VisA, with performance improvements that exceed the observed standard deviations across different training seeds (e.g., ODD: (82.2 ± 0.36) vs. InvAD: (80.1) on MPDD). These results again demonstrate that ODD consistently achieves strong yet stable detection performance despite the stochastic nature of diffusion model training, while also providing substantially faster inference.

AuthorsOpenReview ↗

Response to Reviewer QAvR (part3)

4. FPS batch sizes and timing scope

The image-based FPS results use batch size 64, whereas other modalities use batch size 1. This distinction, together with the components included in timing, should be stated more clearly.

Thank you for highlighting the need to clarify our FPS measurement protocol. For image-based AD, we strictly follow the evaluation settings adopted in [1, 2], using a batch size of 64 to ensure consistency and comparability with prior studies. In contrast, standardized FPS evaluation protocols are not yet widely established for other modalities. We therefore adopt a batch size of 1 to better reflect practical real-time streaming inference scenarios. Across all modalities, we measure FPS based on the time required for anomaly scoring, including feature extraction, denoising, anomaly score computation, and score post-processing. In the revised manuscript, we have clarified these measurement protocols.

References

[1] Y. Feng, W. Chen, Y. Li, B. Chen, Y. Wang, Z. Zhao, H. Liu, and M. Zhou. “OmiAD: One-Step Adaptive Masked Diffusion Model for Multi-class Anomaly Detection via Adversarial Distillation.” Proceedings of the International Conference on Machine Learning (ICML), PMLR 267, pp. 16604–16633, 2025.

[2] S. Sakai, X. He, C. Gu, L. Sigal, and T. Hasegawa. “InvAD: Inversion-based Reconstruction-Free Anomaly Detection with Diffusion Models.” Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2026.

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

This paper investigates the role of iterative denoising in diffusion-based anomaly detection and identifies a phenomenon termed one-step optimality, where single-step reconstruction often achieves better detection performance than multi-step denoising. The authors provide a theoretical analysis based on optimal velocity fields, showing that iterative denoising can increase anomaly leakage and reduce the separation between normal and anomalous samples under certain assumptions. Based on these findings, the paper introduces the One-step Diffusion anomaly Detector (ODD), which retains standard diffusion training while performing reconstruction in a single step. Experiments across different datasets, diffusion architectures, and data modalities demonstrate comparable detection performance and improved inference efficiency.

Critical strengths
  1. The paper provides a theoretical analysis of one-step optimality of anomaly detection, offering useful insights into the relationship between reconstruction quality and anomaly detection performance.

  2. The experiments covers multiple datasets and diffusion architectures.

Critical weaknesses

1.Limited methodological novelty: The proposed ODD primarily analyzes why single-step reconstruction can be superior to multi-step reconstruction for anomaly detection; however, one-step diffusion-based anomaly detection has already been explored in prior research. In this sense, the paper functions more like an analytical report.

2.The theoretical guarantee for the optimality of a single-step approach relies on assumptions such as Gaussian mixture distributions, additive anomalies, posterior anomaly invariance, and prototype fidelity. The applicability of these assumptions to complex, real-world anomaly distributions remains unclear—a point further underscored by the counterexamples presented in the supplementary material.

3.While ODD achieves competitive results and significant acceleration in inference, it does not consistently outperform existing methods across all benchmarks.

Questions for authors

N/A

4 comments
AuthorsOpenReview ↗

Response to Reviewer DAgG (part1)

Thank you for your careful assessment of our work. We appreciate the opportunity to clarify our contributions, theoretical scope, and empirical results.

1. Methodological novelty and prior one-step work

Limited methodological novelty: The proposed ODD primarily analyzes why single-step reconstruction can be superior to multi-step reconstruction for anomaly detection; however, one-step diffusion-based anomaly detection has already been explored in prior research. In this sense, the paper functions more like an analytical report.

We would like to clarify that our one-step paradigm is conceptually different from prior attempts that explore one-step denoising in diffusion-based AD. To the best of our knowledge, [1, 2] represent the most relevant prior studies in this direction. Specifically, the central focus of both [1] and [2] lies primarily in devising dedicated refinement procedures to compensate for the performance degradation associated with one-step denoising, rather than investigating the intrinsic properties and potential of the one-step denoising itself.

In stark contrast, our work is the first to systematically investigate native one-step denoising, uncovering one-step optimality in diffusion-based AD through comprehensive empirical and theoretical analyses. Accordingly, our proposed ODD establishes a generalizable one-step denoising paradigm that leverages the intrinsic capabilities of diffusion models without relying on specialized refinement modules, fundamentally distinguishing it from [1, 2]. (See Related Works for a more detailed discussion.) As also recognized by the other two reviewers (HQ84 and QAvR), methodological simplicity constitutes a key strength of our work, which therefore allows ODD to be readily incorporated into other diffusion-based AD techniques (see Table 3 in the revised manuscript).

2. Assumptions, applicability, and counterexamples

The theoretical guarantee for the optimality of a single-step approach relies on assumptions such as Gaussian mixture distributions, additive anomalies, posterior anomaly invariance, and prototype fidelity. The applicability of these assumptions to complex, real-world anomaly distributions remains unclear—a point further underscored by the counterexamples presented in the supplementary material.

Thanks for pointing this out.

We need to clarify that these counterexamples (Fig. 9 in the original paper) barely narrow the applicability of our one-step optimality. First, the anomalous background bias is a benchmark issue, which we resolve in Fig. 9(a) by simply focusing on meaningful foreground objects, where we observe one-step optimality. Also, since the remaining two examples are tied to the geometry of the pre-trained feature space and an insufficient training budget, they can again be easily bypassed by adopting a more discriminative feature extractor or increasing training budgets.

Second, our theoretical derivation follows the widely adopted assumptions in the understanding of diffusion modeling and unsupervised AD fields. Precisely, additive anomaly assumptions are extensively used in the AD literature [3,4,5,6] because of their flexibility, and GMM data assumptions are broadly adopted to characterize complex distributions [7,8,9,10,11]. Moreover, prototype fidelity is often introduced to reduce the nonlinearity in the responsibility terms [12,13,14], and we extend this assumption with normal-anomalous invariance. Consequently, as acknowledged by reviewer (HQ84), our theoretical assumptions are reasonable and do not strictly limit their applicability to complex scenarios.

To further support these theoretical assumptions with experimental evidence, we investigate the applicability of the GMM data assumption and posterior anomaly invariance (see D.4 in the revised manuscript). Through systematic analysis, we find that: (i) the normal feature distributions in a realistic AD dataset (MVTec AD) can be modeled using a small number of GMM components, with the lowest held-out NLL attained using only five components on average; and (ii) the posterior responsibilities of paired normal and pseudo-anomalous features remain approximately invariant during reconstruction, exhibiting consistent stability across the evaluated texture categories and anomaly intensities. These findings provide empirical support for the applicability of our theoretical assumptions in realistic, complex AD scenarios.

AuthorsOpenReview ↗

Response to Reviewer DAgG (part2)

3. Performance across benchmarks

While ODD achieves competitive results and significant acceleration in inference, it does not consistently outperform existing methods across all benchmarks.

Thanks for your comment. We need to clarify that the central contribution of this paper is to establish one-step optimality through systematic and theoretical analysis, not simply chasing the state-of-the-art performance.

Yet, across the datasets shared with each baseline, ODD (DiT-G) scores higher on 22 of 28 dataset–metric comparisons against OmiAD (one-step diffusion with distillation) and 12 of 14 against DiffusionAD (one-step diffusion using pseudo-anomalies). Furthermore, ODD achieves the best speed–accuracy trade-off across all evaluation settings in our paper, delivering competitive AD accuracy while maintaining remarkable efficiency of one-step denoising. This highlights that our simple yet generalizable paradigm can achieve strong detection performance compared with methods specifically developed for one-step denoising.

To further investigate the generality of ODD, we additionally compare our methods against three SoTA diffusion-based AD methods (OmiAD, InvAD, ReFP-AD) on three medical AD benchmarks (RESC, BraTS, RSNA) and two industrial AD benchmarks (BTAD, VIADUCT).

Quantitative results on different AD datasets under the multi-class setting. We compare against the best-performing diffusion-based methods: OmiAD, InvAD, and ReFP-AD. Since RSNA only provides image-level annotations, we omit its pixel-level metrics and report mAD as the average of all image-level metrics. Best results for each dataset and metric are highlighted in bold.

DatasetMethodNFE (↓)Image AUROCImage APImage F1_maxPixel AUROCPixel APPixel F1_maxAU-PROmADFPS
RSNAOmiAD (ICML'25)164.297.397.7————86.437.1
InvAD (CVPR'26)373.198.097.7————89.693.3
ReFP-AD (ECCV'26)4071.497.897.7————89.0162.4
ODD (DiT-G)185.099.097.7————93.9168.3
RESCOmiAD (ICML'25)177.666.269.994.138.947.277.867.438.4
InvAD (CVPR'26)388.986.378.396.259.660.484.179.193.7
ReFP-AD (ECCV'26)4090.888.480.197.067.563.686.982.1164.0
ODD (DiT-G)195.694.887.097.473.767.788.886.4168.5
BraTSOmiAD (ICML'25)179.293.491.597.151.453.581.678.238.5
InvAD (CVPR'26)389.697.093.095.850.051.373.678.693.4
ReFP-AD (ECCV'26)4072.489.991.190.921.330.864.865.9162.2
ODD (DiT-G)194.698.794.798.570.467.188.887.6168.4
BTADOmiAD (ICML'25)193.492.288.097.043.449.878.277.439.0
InvAD (CVPR'26)394.298.897.096.744.049.575.279.393.2
ReFP-AD (ECCV'26)4094.498.896.297.550.353.578.981.4160.8
ODD (DiT-G)194.798.995.897.652.055.679.682.0169.0
VIADUCTOmiAD (ICML'25)152.968.178.692.77.912.578.755.936.8
InvAD (CVPR'26)376.383.984.590.723.130.172.265.893.4
ReFP-AD (ECCV'26)4065.275.781.094.614.419.582.761.8163.3
ODD (DiT-G)182.989.486.596.626.533.390.472.2168.9

As shown, ODD mostly outperforms these strong competitors across datasets and metrics (27/31) while retaining efficient inference, again confirming the generality. We have added these results to the revised manuscript (see Table 12).

AuthorsOpenReview ↗

Response to Reviewer DAgG (part3)

References

[1] Y. Feng, W. Chen, Y. Li, B. Chen, Y. Wang, Z. Zhao, H. Liu, and M. Zhou. “OmiAD: One-Step Adaptive Masked Diffusion Model for Multi-class Anomaly Detection via Adversarial Distillation.” Proceedings of the International Conference on Machine Learning (ICML), PMLR 267, pp. 16604–16633, 2025.

[2] H. Zhang, Z. Wang, D. Zeng, Z. Wu, and Y.-G. Jiang. “DiffusionAD: Norm-guided One-step Denoising Diffusion for Anomaly Detection.” IEEE Transactions on Pattern Analysis and Machine Intelligence, 47(8), pp. 7140–7152, 2025. https://doi.org/10.1109/TPAMI.2025.3570494.

[3] D. Datta, F. Chen, and N. Ramakrishnan. "Framing Algorithmic Recourse for Anomaly Detection." Proceedings of the ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD), pp. 283–293, 2022. https://doi.org/10.1145/3534678.3539344.

[4] Y. Cai, H. Chen, and K.-T. Cheng. "Rethinking Autoencoders for Medical Anomaly Detection from A Theoretical Perspective." Medical Image Computing and Computer-Assisted Intervention (MICCAI), pp. 544–554, 2024.

[5] H. Yao, M. Liu, Z. Yin, Z. Yan, X. Hong, and W. Zuo. "GLAD: Towards Better Reconstruction with Global and Local Adaptive Diffusion Models for Unsupervised Anomaly Detection." Computer Vision – ECCV 2024, pp. 1–17, 2025. https://doi.org/10.1007/978-3-031-73209-6_1.

[6] W. Dai, K. Hwang, and J. Fan. "Unsupervised Anomaly Detection for Tabular Data Using Deep Noise Evaluation." Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), vol. 39, no. 11, pp. 11553–11562, 2025. https://doi.org/10.1609/aaai.v39i11.33257.

[7] T. Bonnaire, R. Urfin, G. Biroli, and M. Mezard. "Why Diffusion Models Don't Memorize: The Role of Implicit Dynamical Regularization in Training." Advances in Neural Information Processing Systems (NeurIPS), vol. 38, 2025.

[8] Z. Ye, Q. Zhu, M. Tao, and M. Chen. "Provable Separations between Memorization and Generalization in Diffusion Models." International Conference on Learning Representations (ICLR), 2026.

[9] S. Buchanan, D. Pai, Y. Ma, and V. De Bortoli. "On the Edge of Memorization in Diffusion Models." Advances in Neural Information Processing Systems (NeurIPS), vol. 38, pp. 96113–96157, 2025. https://doi.org/10.52202/085713-3213.

[10] Z. Zhang, X. Li, X. Li, L. Shi, M. Wu, M. Tao, and Q. Qu. "Generalization of Diffusion Models Arises with a Balanced Representation Space." International Conference on Learning Representations (ICLR), pp. 101444–101475, 2026.

[11] H. Cui, C. Pehlevan, and Y. Lu. "A Solvable Model of Learning Generative Diffusion: Theory and Insights." Advances in Neural Information Processing Systems (NeurIPS), vol. 38, pp. 5253–5296, 2025. https://doi.org/10.52202/085713-0187.

[12] R. Levi and M. Elad. "Mixture-of-Gaussians-Guided Schedule Design for Brownian Bridge Diffusion Models." Transactions on Machine Learning Research (TMLR), 2026. https://arxiv.org/abs/2607.03517.

[13] K. Shah, S. Chen, and A. Klivans. "Learning Mixtures of Gaussians Using the DDPM Objective." Advances in Neural Information Processing Systems (NeurIPS), 2023.

[14] J. Ma, L. Xu, and M. I. Jordan. "Asymptotic Convergence Rate of the EM Algorithm for Gaussian Mixtures." Neural Computation, vol. 12, no. 12, pp. 2881–2907, 2000.

AuthorsOpenReview ↗

Gentle Reminder

Dear Reviewer DAgG,

Thank you for your careful review and for the critical points you raised, which helped us sharpen the paper.

In our rebuttal, we addressed your comments on the methodological novelty, the applicability of the theoretical assumptions to real-world anomalies, and the performance of ODD. In response, we have clarified our novelty relative to the prior one-step diffusion AD literature, included additional experiments verifying the theoretical assumptions in real-world AD scenarios, and verified the performance advantages of ODD with additional results across six AD benchmarks.

Since the discussion period ends on November 18, we would appreciate hearing whether our rebuttal and revisions address your comments. We would be pleased to discuss any remaining issues in the time left.

Best regards, The Authors of Submission 6795

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.