acceptodds
Under review as a conference paper at ICLR 2027

OSD-IRF: One-Step Diffusion Anomaly Detection with Inverse Residual Fields

Abstract

Diffusion models have demonstrated strong performance in unsupervised industrial anomaly detection (uIAD) by learning the distribution of normal data. These methods typically assume that off-manifold anomalies are more difficult to reconstruct or generate, leading to larger reconstruction errors in the data space or lower probability densities in a tractable latent space. However, their iterative denoising process incurs substantial inference latency. To overcome this limitation, we propose OSD-IRF, a novel one-step diffusion framework based on inverse residual fields (IRFs). Our key observation is that anomalies become distinguishable in the IRF space, enabling efficient detection in a single diffusion step without iterative denoising or model distillation. Specifically, we first train an unconditional denoising diffusion probabilistic model (DDPM) exclusively on normal data. At inference time, we use the score function learned by the DDPM to estimate the IRF of each test sample. Anomaly detection is then performed by evaluating the IRF likelihood under a Gaussian distribution and comparing it against a predefined threshold. We further leverage the aggregate signal-to-noise ratio (SNR) statistics of the entire training set to automatically determine the probing timestep used during inference, thereby eliminating empirical timestep selection in IRF estimation. Extensive experiments on widely used uIAD benchmarks show that OSD-IRF achieves competitive performance in anomaly detection and localization while substantially reducing the inference cost of diffusion-based methods.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Unsupervised Industrial Anomaly Detection, Diffusion Models, Inverse Residual Fields, One-Step Inference, Signal-to-Noise Ratio

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_stgJOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes OSD-IRF, a diffusion-based model for unsupervised image anomaly detection. It trains an unconditional DDPM on normal image features and uses its predicted noise, with inverse residual field (IRF), to detect and localize anomalies. It has a one-step inference by evaluating the network on a scaled input. The paper also proposes selecting the inference timestep from SNR statistics of the training data. Experiments on MVTec-AD, VisA, and MPDD show competitive detection and localization performance with fast inference.

Critical strengths
  • The method proposes as simple and practical inference process. A model provides detection and localization scores with one network evaluation, without additional distillation. This is useful for applications where inference speed matters.
  • The results are competitive on three industrial benchmarks, with particularly strong improvements on MPDD. The ablation also shows that the method is useful. It achieves comparable or better performance than averaging 16 stochastic passes.
Critical weaknesses
  • The novelty of the method is weak. The paper presents the IRF as the main contribution, and it uses the standard noise as a predictor, proportional to the learned score, and its Gaussian negative log-likelihood reduces to a squared score norm. There are already a few anomaly detection methods that use score norms as anomaly scores, in image anomaly detection [1, 2], in tabular anomaly detection with single-scale DSM [3, 4]. In particular, these tabular anomaly methods are relevant because they could also be applied to pretrained vision embeddings, just like how it is done in this paper’s method. The paper does not discuss or compare against any of these methods. The paper also incorrectly states that DTE requires iterative reverse diffusion, which is false, as it predicts the timestep distribution directly in one inference pass. It is therefore unclear what the method contributes beyond its SNR selection rule and application to industrial features.
  • The empirical comparison omits substantially stronger multi-class methods. As examples, Dinomaly reports image AUROC of 99.6/98.7 and pixel AP of 69.3/53.2 on MVTec-AD/VisA [5], and RAD also reports substantially stronger localization performance [6]. Backbone and resolution can influence comparisons, but these omissions do put in question the state-of-the-art claim. As briefly mentioned in the previous point, since the method works on pretrained image embeddings, comparisons with tabular anomaly detectors on the same features are also relevant.

References

  1. Mahmood, A., Oliva, J., & Styner, M. (2021). Multiscale Score Matching for Out-of-Distribution Detection. ICLR.

  2. Tebbe, J., & Tayyub, J. (2024). Dynamic Addition of Noise in a Diffusion Model for Anomaly Detection. CVPR Workshops, 3940–3949.

  3. Livernoche, V., Zan, J., & Rabbany, R. (2026). Kurtosis-Guided Denoising Score Matching for Tabular Anomaly Detection. arXiv:2605.06955.

  4. Hirth, M., Koberg, L., Jazdi, N., & Kasneci, E. (2026). Denoising without Diffusion: Fixed-Noise Denoiser Anomaly Detection in Tabular Data. ICML.

  5. Guo, J., Lu, S., Zhang, W., Chen, F., Li, H., & Liao, H. (2025). Dinomaly: The Less Is More Philosophy in Multi-Class Unsupervised Anomaly Detection. CVPR, 20405–20415.

  6. Zhang, X., Li, G., Henderson, P., Aragon-Camarasa, G., & Long, Z. (2026). Is Task-Specific Training Necessary for Anomaly Detection? ICML.

Questions for authors
  • What is the main difference between the evaluation of IRF and the usual noise-normalized score evaluated at a clean input? What components differentiate the contribution beyond the score principle of [1] and the input construction in [2]?
  • Can you compare with the baselines mentioned in the weaknesses?
  • How does SNR-based timestep selection compare with a common fixed timestep across datasets? Can you also compare against a DSM model, with the same architecture, trained only at the selected noise scale, to establish whether multi-timestep training is useful?
7 comments
AuthorsOpenReview ↗

Response to Reviewer stgJ

W1: Novelty and Comparison with Score-Based Anomaly Detection Methods.

We thank the reviewer for highlighting these relevant studies. We acknowledge that score-magnitude anomaly detection is not new and that our IRF is mathematically related to the learned score function. Yet a shared score function does not imply methodological equivalence: response estimation, noise-scale selection, and anomaly-scoring strategies can substantially affect performance. OSD-IRF's novelty lies in combining deterministic response estimation, SNR-guided timestep selection, and direct spatial anomaly scoring for fast, robust score-based anomaly detection at inference, rather than using score magnitude alone. Our training remains easy to implement without more elaborate design than prior methods.

(1) Comparison with MSMA [1] and K-DSM [3]. MSMA forms multiscale score-norm representations by evaluating the learned score function across noise scales and uses an additional density estimator for OOD detection. OSD-IRF instead detects and localizes anomalies from the learned score-related response at one automatically selected probing timestep, without an additional estimator. K-DSM is particularly relevant but methodologically distinct: it trains a single-scale DSM model with kurtosis-guided, feature-wise noise scales and evaluates the learned response on clean test features. OSD-IRF trains a time-conditioned DDPM and evaluates the noise-prediction response at the conditional mean of an SNR-selected noisy state. K-DSM is an arXiv preprint under ICLR 2027 review, not yet peer-reviewed. ICLR 2027 Reviewer Guidelines do not require comparisons with unpublished arXiv papers. Nevertheless, following the reviewer's suggestion, we compare image-level AUROC on MVTec-AD and VisA (Table R1), using results directly from K-DSM's appendix.

Table R1. Image-level AUROC (%) comparison between K-DSM and OSD-IRF on MVTec-AD and VisA.

MethodMVTec-ADVisA
K-DSM90.0088.10
OSD-IRF (Ours)98.8596.89

(2) Comparison with dynamic noise addition [2]. Tebbe et al. [2] introduce dynamic implicit conditioning: KNN distances estimate anomaly severity and guide per-sample adaptive diffusion timestep selection, adding inference overhead. They also explore deterministic input scaling without stochastic noise, but their reconstruction-based method requires iterative denoising for anomaly-free reconstructions. OSD-IRF selects a shared probing timestep solely from normal-data SNR statistics and directly scores anomalies with one deterministic noise-prediction response, without iterative denoising or explicit reconstruction. Thus, they differ fundamentally in timestep selection and anomaly detection mechanisms. Table R2 compares both methods on MVTec-AD and VisA in image-level detection (I-AUROC) and localization (P-AUROC, PRO), using [2]'s results directly from the original paper. These results further support OSD-IRF's effectiveness and its computational advantage of one network evaluation.

Table R2. Performance (%) comparison between Dynamic Noise Addition [2] and OSD-IRF on MVTec-AD and VisA.

DatasetMethodI-AUROC ↑P-AUROC ↑PRO ↑NFE
MVTec-ADDynamic Noise Addition [2]97.2097.4093.3010
OSD-IRF (Ours)98.8597.7793.681
VisADynamic Noise Addition [2]96.0097.9094.1010
OSD-IRF (Ours)96.8998.9693.281

(3) Comparison with fixed-noise denoising [4]. Hirth et al. [4] estimate expected deviation using a fixed-noise denoiser and repeated stochastic perturbations. OSD-IRF instead uses deterministic mean-input evaluation without test-time sampling, requiring one network evaluation. On MVTec-AD, Fig. 5 reports 86.0 mAD for input-aggregated IRF (one evaluation), versus 85.3 with 16 stochastic evaluations and 57.4 with one stochastic evaluation. Although not a direct comparison with [4], these results demonstrate the advantage of deterministic response estimation over stochastic aggregation within our framework. We reproduce their method under the same experimental setup as OSD-IRF for a fair comparison on MPDD (Table R3).

Table R3. Performance (%) comparison between fixed-noise denoising [4] and OSD-IRF on MPDD.

MethodInferenceI-AUROC ↑P-AUROC ↑PRO ↑mAD ↑
Hirth et al. [4]Stochastic, K=195.6793.1979.3876.59
Hirth et al. [4]Stochastic, K=497.0898.6293.7982.34
Hirth et al. [4]Stochastic, K=1696.9598.7994.8782.76
OSD-IRF (Ours)Deterministic97.6598.8395.0583.52

(4) Correction and clarification of DTE's contributions. We thank the reviewer for noting this distinction. DTE directly estimates the diffusion-time distribution without iterative reverse diffusion. We will revise this description in the manuscript.

AuthorsOpenReview ↗

Response to Reviewer stgJ

W2: Comparison with Stronger Multi-Class Anomaly Detection Methods.

We thank the reviewer for highlighting these strong multi-class anomaly detection methods. We would like to clarify that, as stated in our Abstract, our primary objective is to achieve competitive anomaly detection and localization performance while substantially reducing the inference cost of diffusion-based methods. We do not claim state-of-the-art performance over all existing industrial anomaly detection methods, including discriminative and retrieval-based approaches.

(1) Comparison with Dinomaly[5]. We acknowledge the strong performance of Dinomaly and agree that including it provides a more comprehensive evaluation. Accordingly, we conduct additional experiments comparing Dinomaly and OSD-IRF on the VisA and MPDD datasets, using the same input image resolution of 256 × 256. As reported in Table R4, OSD-IRF achieves competitive performance compared with Dinomaly, further demonstrating the effectiveness of our one-step, reconstruction-free diffusion framework.

Table R4. Performance (%) comparison between Dinomaly and OSD-IRF on VisA and MPDD.

DatasetMethodI-AUROC ↑P-AUROC ↑P-AP ↑
VisADinomaly96.7098.8242.90
OSD-IRF (Ours)96.8998.9643.73
MPDDDinomaly94.6898.5352.76
OSD-IRF (Ours)97.6598.8348.53

(2) Comparison with RAD [6]. We acknowledge that RAD achieves stronger anomaly localization performance. However, RAD follows a substantially different retrieval-based paradigm, leveraging multi-layer visual features, global semantic context, and test-sample-adaptive reference retrieval to construct spatially conditioned anomaly scores. In contrast, OSD-IRF directly derives anomaly scores from a learned diffusion response without test-time memory-bank retrieval. We recognize that RAD's multi-level and context-aware representations may contribute to its strong fine-grained anomaly localization, while such information is not fully exploited by our current framework. Incorporating complementary multi-scale or context-aware information into OSD-IRF represents a promising direction for future research.

Overall, these comparisons help clarify the scope of our contribution: OSD-IRF aims to provide an efficient diffusion-based anomaly detection framework with competitive performance, rather than outperforming all discriminative or retrieval-based methods. We will revise the experimental discussion to make this distinction explicit and include the additional comparison with Dinomaly.

Q1. Difference between IRF and conventional noise-normalized score evaluation

See W1.

Q2. Comparison with the baselines mentioned in the weaknesses

See W1 and W2.

Q3. SNR-based timestep selection and single-noise-scale DSM training

(1) SNR-based timestep selection vs. fixed timesteps. As shown in Fig. 3 of the original manuscript, detection performance is sensitive to the probing timestep, with high performance observed near logSNR = 0. Although a common fixed timestep can be used, selecting it based on detection performance would require anomalous validation data. In contrast, our SNR-guided strategy automatically identifies effective probing timesteps using only normal-data statistics.

(2) Single-noise-scale DSM vs. multi-timestep training. Following the reviewer's suggestion, we trained a single-noise-scale DSM model at t∗=588t^*=588, using the same architecture and noise-prediction objective as OSD-IRF. Both models use identical deterministic inference and IRF-based anomaly scoring at t∗=588t^*=588, differing only in their training noise scales. On MPDD, single-timestep training achieves 82.59 mAD, compared with 83.52 mAD for multi-timestep training, demonstrating its empirical benefit while retaining inference with only one network evaluation.

References

  1. Mahmood, A., Oliva, J., & Styner, M. (2021). Multiscale Score Matching for Out-of-Distribution Detection. ICLR.
  2. Tebbe, J., & Tayyub, J. (2024). Dynamic Addition of Noise in a Diffusion Model for Anomaly Detection. CVPR Workshops, 3940–3949.
  3. Livernoche, V., Zan, J., & Rabbany, R. (2026). Kurtosis-Guided Denoising Score Matching for Tabular Anomaly Detection. arXiv:2605.06955.
  4. Hirth, M., Koberg, L., Jazdi, N., & Kasneci, E. (2026). Denoising without Diffusion: Fixed-Noise Denoiser Anomaly Detection in Tabular Data. ICML.
  5. Guo, J., Lu, S., Zhang, W., Chen, F., Li, H., & Liao, H. (2025). Dinomaly: The Less Is More Philosophy in Multi-Class Unsupervised Anomaly Detection. CVPR, 20405–20415.
  6. Zhang, X., Li, G., Henderson, P., Aragon-Camarasa, G., & Long, Z. (2026). Is Task-Specific Training Necessary for Anomaly Detection? ICML.

Thank you for the rebuttal and the extra experiments. I appreciate the DTE correction, the reimplemented baselines, and the single-noise-scale ablation, and I'm retracting the request for K-DSM.

However, my main concerns still remain:

  1. Novelty (Q1): The rebuttal doesn't address this. With σ2=(1−αˉt)/αˉt\sigma^2 = (1-\bar\alpha_t)/\bar\alpha_t, we have ϵθ(αˉt x0,t)=−σ∇xlog⁡pσ(x0)\epsilon_\theta(\sqrt{\bar\alpha_t}\,x_0, t) = -\sigma \nabla_x \log p_\sigma(x_0). This means the input-aggregated IRF is the noise-normalized score evaluated at the clean input, as in prior score-norm methods. "Mean-input evaluation" is a reparametrization, not a new estimator. Your single-noise-scale DSM ablation is only marginally below the full method and is only reported on one dataset. It would be better to show it on all three datasets and report the variance over seeds.

  2. Timestep selection (Q3): The SNR rule picks nearly the same timestep across all three datasets, so the current experiments can't distinguish it from a fixed common timestep. To show its benefit, you would need a setting where the selected timestep actually changes, such as features with a different scale from another backbone or layer set.

  3. Dinomaly: The new comparison leaves out MVTec-AD and runs at a lower resolution than Dinomaly's default, where its performance falls well below published results. Since Section 4.5 and Table 1 frame this as a state-of-the-art comparison, these methods need to be included there.

Based on this, I'm maintaining my score for now.

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

Given the inference latency caused by iterative denoising or multi-step inversion in existing diffusion model-based anomaly detection methods, this work proposes OSD-IRF, which employs only the one-step score function of DDPM during inference to perform anomaly detection based on an adaptively determined time step derived from SNR statistics.

Critical strengths
  1. The paper is well written, with a clear and reasonable research motivation, methodology, and verification.

  2. The results demonstrate that the proposed method achieves comparable or even superior detection performance while incurring lower inference cost.

  3. The ablation study provided demonstrates the rationale behind the empirical settings.

Critical weaknesses
  1. This work does not clarify the basic assumptions on which it relies. Such clarification is important for establishing the fundamental effectiveness of any method; while a method cannot cover all conditions, it should specify the conditions under which it works and those under which it does not.

  2. Regarding the anomaly score introduced in Eq. (12), the numerical scale between snulls_\text{null} and sdiffs_\text{diff} appears unreasonable.

Questions for authors
  1. Does the proposed method require training different models for different data classes?

  2. Can the analysis of the numerical scale between snulls_\text{null} and sdiffs_\text{diff} be provided?

  3. Why is only one t∗t^* used for each dataset, rather than being determined by the number of classes?

  4. From the perspective of the overall denoising process, is the selected t∗t^* based on SNR a good choice? A comprehensive analysis across all datasets is necessary.

7 comments
AuthorsOpenReview ↗

Response to Reviewer to vVPf

W1: Basic assumptions.

We thank the reviewer for this valuable comment. We agree that explicitly clarifying the underlying assumptions and applicability of OSD-IRF is essential. Our method relies on three main assumptions under the multi-class (one-for-all) anomaly detection setting, which we will clarify, along with their associated applicability conditions and limitations, in the revised manuscript.

(1) Training strategy and scope. OSD-IRF trains a single diffusion model on normal samples from all categories within each dataset, rather than a separate model for each category. Specifically, one model is shared across all 15 categories of MVTec-AD, while separate models are trained for VisA (12 categories) and MPDD (6 categories). Thus, our one-for-all setting refers to a shared model within each dataset, rather than across different datasets.

(2) Underlying assumptions and applicability. Given this multi-class setting, OSD-IRF relies on three main assumptions:

(i) Normal-distribution learnability. The normal training samples sufficiently represent the target categories, enabling the DDPM to learn their joint normal-data distribution. Insufficient coverage or substantial distribution shifts may cause false positives.

(ii) IRF-response discriminability. At an appropriate timestep, normal and anomalous inputs induce distinguishable responses in the learned score field. We assume that anomalies produce distinguishable response magnitudes or spatial patterns, as empirically supported by our toy-data analysis and experimental results. However, anomalies with responses similar to normal patterns may remain undetected.

(iii) Effectiveness of SNR-guided timestep selection. We assume that the transition between signal-dominated and noise-dominated diffusion states provides an informative operating region for anomaly discrimination. Accordingly, the probing timestep is selected using the aggregated SNR statistics of all normal training features within each dataset, without requiring anomalous validation samples or category-specific timestep tuning.

(3) Limitations and scope clarification. Our experiments validate OSD-IRF on MVTec-AD, VisA, and MPDD under the known-category, multi-class setting. Generalization to unseen categories or substantially different datasets without retraining, which relates to zero-shot or cross-domain anomaly detection, remains beyond our current scope. Furthermore, a shared model and dataset-level timestep may be suboptimal when category distributions or anomaly characteristics vary considerably.

W2. Clarification of the Numerical Scales of the Anomaly Scores in Eq. (12)

The two anomaly scores, s_diff and s_nll, are independently min–max normalized using parameters estimated exclusively from normal training data before being combined, thereby mitigating potential numerical scale imbalances in Eq. (12). We will explicitly clarify this normalization. Furthermore, we examine the numerical scales of these two scores computed from normal and anomalous test samples and report the corresponding statistics in Table R1.

Table R1. Numerical scales of s_diff and s_nll.

DatasetNormal Test (Diff / NLL, mean ± std)Anomalous Test (Diff / NLL, mean ± std)
MVTec-AD0.345 ± 0.162 / 0.366 ± 0.1633.910 ± 1.043 / 3.565 ± 1.381
VisA0.288 ± 0.139 / 0.276 ± 0.1250.904 ± 0.343 / 0.614 ± 0.301
MPDD0.398 ± 0.161 / 0.375 ± 0.1923.334 ± 0.641 / 4.446 ± 1.020

As shown in Table R1, s_diff and s_nll exhibit comparable numerical scales within each of the three datasets, providing empirical support for their equal-weight combination in anomaly detection. The effectiveness of this combination is further validated by the ablation results in Table 2.

AuthorsOpenReview ↗

Response to Reviewer vVPf

Q1. Does the proposed method require training separate models for different categories?

See "Training strategy and scope" in W1.

Q2. Analysis of the Numerical Scale.

See W2.

Q3. Why is a single timestep selected for each dataset rather than separate timesteps for individual categories?

OSD-IRF follows the established One-for-All multi-class anomaly detection setting, in which a single model is jointly trained using normal samples from all categories within each dataset. Consistent with this setting, our timestep selection does not rely on category information or category-specific tuning.

To address the reviewer's concern, we further evaluate category-specific SNR-guided timestep selection using the same trained model for each of the three datasets, with results reported in Table R2. The comparable performance suggests that dataset-level timestep selection provides an effective alternative to category-specific tuning, without requiring separate timestep selection for each category.

Table R2. Performance (mAD ↑, %) comparison between dataset-level and category-specific timestep selection

DatasetDataset-level selectionCategory-specific selection
MVTec-AD85.9885.96
VisA81.4581.51
MPDD83.5283.59

Q4. Justification and cross-dataset analysis of SNR-guided timestep selection.

We thank the reviewer for this insightful question. Our SNR-guided timestep selection is motivated by the trade-off between preserving anomaly-relevant information and smoothing normal-data variations during diffusion.

(1) Theoretical motivation. According to our input-aggregated IRF response, at small timesteps (high SNR), the perturbed features remain close to their original values, preserving anomaly information but potentially retaining fine-grained normal variations that affect the learned score response. Conversely, at large timesteps (low SNR), the perturbed normal-data distribution approaches a Gaussian distribution, while the mean-perturbed inputs of normal and anomalous samples become increasingly indistinguishable, reducing the discriminative information available to our input-aggregated IRF.

Therefore, an intermediate timestep may provide a favorable balance between feature preservation and distributional smoothing. Motivated by this observation, we select

t⋆=arg⁡min⁡t∣log⁡SNRN(t)∣,t^\star=\arg\min_t|\log\mathrm{SNR}_N(t)|,

where the calibrated SNR is estimated exclusively from normal training features and the predefined diffusion schedule. This criterion identifies the transition region where signal and noise have approximately equal average power, without requiring anomalous validation samples. We emphasize that this is a theoretically motivated, normal-data-driven selection criterion rather than a guarantee of globally optimal anomaly detection performance.

(2) Cross-dataset empirical validation. To further address the reviewer's concern, we extend the timestep sensitivity analysis in Fig. 3 to VisA and MPDD, in addition to MVTec-AD. The supplementary results in Table R3 illustrate the relationship among probing timestep, calibrated SNR, and anomaly detection performance across all three datasets, providing a quantitative assessment of the effectiveness of our selection criterion.

The results reveal a region of relatively stable, high anomaly detection performance around the SNR transition. Our simple, normal-data-driven SNR criterion identifies a suitable timestep within this region, without requiring anomalous validation samples or performance-based timestep tuning.

These analyses further clarify the relationship between diffusion noise levels and anomaly discriminability and evaluate the general applicability of the proposed SNR-guided timestep selection across different industrial anomaly detection benchmarks.

Table R3. Calibrated log SNR and mAD (%, ↑) at different probing timesteps across three datasets.

ttMVTec log SNRMVTec-AD mADVisA log SNRVisA mADMPDD log SNRMPDD mAD
100+5.5682.86+5.6242.46+5.6480.18
200+4.0684.04+4.1265.41+4.1481.73
300+2.9984.56+3.0475.63+3.0682.19
400+1.9985.11+2.0578.17+2.0782.68
500+0.9585.68+1.0080.38+1.0283.25
600−0.2286.03−0.1781.43−0.1583.86
700−1.5685.72−1.5075.58−1.4882.70
800−3.0881.54−3.0265.40−3.0072.30
900−4.8069.37−4.7456.94−4.7257.30

Thank you for the authors' response. The clarified assumptions, acknowledged limitations, and additional empirical evaluations partically address my concerns.

Regarding the rationale for the IRF-response discriminability assumption, attributing its validity to the empirical results establishes an incorrect causal relationship. Instead, the authors can claim that the empirical results provide support for its effectiveness. The underlying rationale for why this assumption may hold is that only normal samples are used during training.

I still have two main reservations.

First, although the results in Table R1 suggest that the two terms have comparable numerical scales, their definitions in Eq. (12) suggest a significant difference in their numerical scales. Why do the empirical results appear to contradict this theoretical expectation? A clear explanation and further analysis of this discrepancy should be provided.

Second, Table R2 indicates that separate t* selection yields almost identical results to using a unified t*. I am curious about the underlying reason for this observation. Do all classes tend to select approximately the same t*, or is the anomaly detection performance relatively insensitive to the selection of t*?

The exact code implementation should also be provided to facilitate verification of the reported empirical results.

Reviewer_qN6LOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes a method for unsupervised industrial anomaly detection and localization. It trains a DDPM on normal image features and uses an input-aggregated inverse residual field (IRF) for inference. The probing timestep is selected using SNR statistics of normal training features. Anomaly scores and localization maps are computed from the IRF response with one diffusion-network evaluation. Experiments on MVTec-AD, VisA, and MPDD show competitive performance and improved inference speed over the compared diffusion-based methods.

Critical strengths
  • The proposed method avoids iterative denoising, multi-step inversion, and model distillation.
  • The method achieves the highest mAD among the compared methods on all three datasets. The SNR-guided timestep selection also provides a practical rule that uses only normal training features.
Critical weaknesses
  • The central IRF formulation and its probabilistic interpretation need further justification. Substituting the forward perturbation equation into Eq. (5) gives Δxt=ϵ\Delta x_t=\epsilon. This residual is the sampled Gaussian noise for both normal and anomalous inputs. However, the noise prediction network generally estimates a conditional expectation, ϵθ⋆(xt,t)=E[ϵ∣xt]\epsilon_\theta^\star(x_t,t)=\mathbb E[\epsilon\mid x_t], under the normal training distribution. It does not recover each sampled noise exactly. Thus, the equality introduced in Eq. (6) conflates two different quantities. The claim that they approximately follow the same distribution does not follow from the DDPM training objective. Moreover, the deterministic input-aggregated response is not shown to follow a standard Gaussian distribution, although this distribution motivates the anomaly score in Eq. (12). These gaps affect the interpretation of the central contribution. They do not invalidate the empirical results, but they leave the proposed formulation without adequate theoretical support.

  • One diffusion-network evaluation does not guarantee low computational cost, particularly with the reported DiT-giant configuration. The paper reports FPS but omits parameter counts, FLOPs, training time, and inference memory. The timing protocol is also unclear, including batch size, numerical precision, and whether feature extraction is included.

  • Important implementation details are missing, such as the learning rate, feature-layer selection and fusion, feature normalization, and patch/token configuration.

Questions for authors
  • The authors are suggested to distinguish the sampled residual in Eq. (5) from the predicted response in Eq. (6), and revise the derivation accordingly. The reviewer wonders what assumptions justify the claimed distributional relationship and the standard Gaussian reference for the input-aggregated IRF.

  • Can the authors provide a direct methodological comparison with [Shin et al., 2023] and the method in [a]? In particular, how does the proposed IRF response differ from existing score-based or predicted-noise anomaly statistics? A controlled comparison would help establish the contribution beyond the single-step, reconstruction-free formulation.

  • It would be better to provide the missing model-cost measurements, timing protocol, and implementation settings listed above. Please also specify the linear noise schedule endpoints needed to reproduce the selected probing timesteps.

  • Section 4.1 reports 2,629 normal training images for MVTec-AD, whereas the standard dataset contains 3,629 training images. Please clarify the actual training split and how the checkpoints and deployment thresholds were selected without anomalous validation data.

[Shin et al., 2023] Anomaly Detection using Score-based Perturbation Resilience [a] Moradi M, Grasso M, Colosimo BM, Paynabar K. Single-step reconstruction-free anomaly detection and segmentation via diffusion models. In2025 International Conference on Machine Learning and Applications (ICMLA) 2025 Dec 3 (pp. 663-670). IEEE.

4 comments
AuthorsOpenReview ↗

Response to Reviewer qN6L

W1: Theoretical Justification of IRF (Eqs. 5–7 and 12).

We acknowledge that the original formulation does not sufficiently distinguish sampled forward noise from the learned noise-prediction response, and that the Gaussian interpretation requires clarification. We will revise these descriptions accordingly, while emphasizing that these issues do not affect the validity of the proposed method or the correctness of the experimental results.

(1) IRF definition and its connection to the score function. We agree that Eq. (5) yields the sampled Gaussian noise, which is independent of whether the input is normal or anomalous and thus is not inherently anomaly-discriminative. Under the DDPM objective, the optimal noise predictor estimates the conditional expectation of the sampled noise rather than recovering each realization exactly. Accordingly, we clarify that IRF refers to the learned noise-prediction response, which is proportional to the estimated score of the perturbed normal-data distribution, rather than the sampled forward residual. This corrects the unjustified distributional equivalence in Eq. (6).

(2) Discriminability of the learned IRF response. Since the DDPM is trained exclusively on normal data, normal and anomalous inputs may induce distinguishable responses in the learned score field. Our input-aggregated IRF evaluates the predictor at the conditional mean of the perturbed input, eliminating test-time sampling variability. Although this deterministic evaluation is not mathematically equivalent to averaging stochastic predictions, it provides an efficient and effective response estimator. The toy-data analysis (Fig. 2) and estimator comparison (Fig. 5) empirically support its improved anomaly discriminability and detection performance.

(3) Gaussian reference for anomaly scoring. We agree that the input-aggregated IRF is not guaranteed to follow a standard Gaussian distribution. Therefore, Eq. (12) should be interpreted as an energy-based anomaly statistic under a fixed Gaussian reference, rather than the exact negative log-likelihood of the IRF under a standard Gaussian distribution. Table 2 supports its effectiveness, showing 94.7% image-level AUROC on MVTec-AD using this statistic alone, with further improvements when combined with the response-induced Gaussian reference.

W2. Computational Complexity and Inference Efficiency

We evaluate the anomaly detection performance (mAD) and inference speed (FPS) of OSD-IRF using both DiT-Giant and DiT-Base architectures, with results reported in Table R1. We further report the parameter count, FLOPs, and inference memory usage in Table R2. The results demonstrate that OSD-IRF achieves a favorable trade-off between computational cost and anomaly detection performance. Notably, the substantially smaller DiT-Base significantly improves inference efficiency while maintaining competitive detection performance, highlighting the scalability of OSD-IRF across different model sizes.

Table R1. mAD and FPS across three benchmarks.

DatasetGiant mAD (%)Base mAD (%)Difference (pp)Giant FPSBase FPS
MVTec-AD85.9885.36−0.62133.0832.46
VisA81.4581.37−0.08105.9819.76
MPDD83.5282.44−1.08212.4862.72

Table R2. Complexity and GPU memory consumption.

ModelParameters (M)FLOPs (G/image)Memory (MiB)
DiT-Giant1,223.83422.318,526
DiT-Base155.8255.011,642
Reduction87.27%86.97%80.74%

Parameters and FLOPs refer to the denoising network, excluding the EfficientNet-B4 feature extractor. FLOPs represent a single denoiser forward pass per image. FPS includes feature extraction. Both FPS and GPU memory are measured with batch size 64 under the same hardware and evaluation settings.

W3. Implementation details.

Implementation details are provided for reproducibility.

(1) Architecture and features. A pretrained EfficientNet-B4 fuses blocks [1, 5, 9, 21] into a 272 × 16 × 16 feature map from each normalized 256 × 256 image. No augmentation is used. DiT-Giant is trained from scratch with 16 Transformer blocks, hidden dimension 2048, and 768 latent channels.

(2) Training. We use the DDPM noise-prediction MSE objective with 1,000 timesteps and a linear beta schedule from 0.0001 to 0.02. Training runs for 300 epochs with AdamW, global batch size 32, and seed 42. The learning rate warms up for 40 epochs to 5 × 10⁻⁵, then cosine-decays to 5 × 10⁻⁶. Weight decay is 0, gradient clipping is 1.0, and EMA decay is 0.999.

(3) Inference. We use EMA weights and select probing timesteps using only normal-training-feature SNR statistics: 582 for MVTec-AD, 587 for VisA, and 588 for MPDD. Deterministic input-aggregated IRF inference requires one denoiser evaluation, with no iterative denoising or stochastic sampling.

AuthorsOpenReview ↗

Response to Reviewer qN6L

Q1.

See W1.

We thank the reviewer for highlighting these studies. We acknowledge prior use of learned scores or predicted noise for reconstruction-free anomaly detection. OSD-IRF's novelty instead lies in response evaluation, anomaly scoring, and probing timestep selection.

(1) Comparison with SPR [1]. Both methods exploit score fields learned by models trained exclusively on normal data but differ in their anomaly statistics. SPR evaluates the discrepancy between injected and predicted perturbations: SSPR(x0)=Eε[∥ε−εθ(xt,t)∥22]S_{SPR}(x_0)=E_{ε}[∥ε-ε_{θ}(x_t,t)∥_2^2] where xt=αˉtx0+1−αˉtε,ε∼N(0,I)x_t=\sqrt{\bar{\alpha}_t}x_0+\sqrt{1-\bar{\alpha}_t}ε,\quad ε\sim N(0,I)

Estimating this statistic typically requires averaging over multiple stochastic perturbations, whereas OSD-IRF directly evaluates the learned response at the perturbed input's conditional mean: Rt(x0)=ϵθ(αˉtx0,t)R_t(x_0)=\epsilon_{\theta}(\sqrt{\bar{\alpha}_t}x_0,t) It derives anomaly statistics from this deterministic response's spatial magnitude and global energy, eliminating test-time noise sampling and requiring only one network evaluation. Whereas SPR empirically selects a perturbation scale, OSD-IRF introduces SNR-guided timestep selection based solely on normal training statistics and the predefined diffusion schedule, without anomalous validation samples. Thus, OSD-IRF differs from SPR in its deterministic response estimation, direct anomaly scoring, and normal-data-guided timestep selection.

We compare OSD-IRF with SPR [1] on MVTec-AD, VisA, and MPDD (Table R3). For a fair comparison, we reimplement SPR under experimental settings consistent with OSD-IRF. OSD-IRF consistently outperforms SPR across all three benchmarks while requiring only one network evaluation per test sample.

Table R3. Performance (mAD, %) comparison between OSD-IRF with SPR.

MethodMVTec-ADVisAMPDD
SPR (K=1)84.8779.1079.90
SPR (K=3)84.9579.5979.95
OSD-IRF (Ours)85.9881.4583.52

(2) Comparison with RADAR [2]. RADAR also uses predicted noise for single-step, reconstruction-free anomaly detection. It smooths the predicted-noise map, extracts Sobel edges, constructs global and maximum local L2L_2-norm features, and trains an Isolation Forest for image-level detection.

In contrast, OSD-IRF directly derives anomaly statistics from its deterministic input-aggregated response: channel-wise response magnitude for pixel-level localization, and spatial response contrast combined with a Gaussian-reference energy statistic for image-level detection.

Unlike RADAR, OSD-IRF requires neither complex post-processing nor an additional one-class classifier. Its deterministic mean-input evaluation and SNR-guided timestep selection explicitly address stochastic response variability and timestep sensitivity, respectively.

Thus, the distinction lies in response estimation and scoring mechanisms, rather than single-step inference or predicted-noise use alone.

Q3. Experimental Protocol and Diffusion Schedule Details

See W3.

Q4. Training Data, Checkpoint Selection, and Detection Thresholds

(1) Training data. We use all 3,629 normal MVTec-AD training images, without anomalous samples, and will clarify this in Section 4.1 of the revised manuscript.

(2) Checkpoint selection. All models use the final EMA weights (decay = 0.999) after a predefined 300-epoch training schedule, without checkpoint selection based on anomalous validation data or test-set performance.

(3) Detection thresholds for deployment without anomalous validation data.

OSD-IRF produces continuous anomaly scores, with no fixed deployment threshold selected in our experiments. AUROC, AP, F1-max, and AU-PRO are evaluated across thresholds. Specifically, F1-max is the maximum F1 score obtained by sweeping thresholds using test labels solely for evaluation, not deployment.

For deployment without anomalous validation data, a threshold can be calibrated using the constant false alarm rate (CFAR) principle [3], by selecting a quantile of the anomaly-score distribution estimated from normal calibration samples according to a predefined target false-positive rate.

However, CFAR-based calibration alone cannot guarantee a desired detection probability. If an application imposes specific detection-rate requirements, representative anomalous validation samples are needed to assess whether the selected operating point meets them.

Thus, deployment threshold calibration is application-dependent and separate from our threshold-sweeping evaluation protocol. We will clarify this distinction in the revised manuscript.

[1] Shin et al., Anomaly Detection using Score-based Perturbation Resilience, ICCV 2023.

[2] Moradi et al., Single-Step Reconstruction-Free Anomaly Detection and Segmentation via Diffusion Models, ICMLA 2025.

[3] Data Adaptive Constant False Alarm Rate Normalizer Design for Active Sonar and Radar. ICASSP 1999.

Thank you for providing the rebuttal, which basically addresses my previous concerns. I am now changing my rating from weak rejection to borderline Accept.

Since revision uploads are still permitted, the authors are encouraged to incorporate the necessary corrections and supporting evidence into the revised manuscript, rather than leaving them only in the rebuttal.

We sincerely thank the reviewer for their constructive feedback and positive reassessment of our work. We greatly appreciate the acknowledgment of our rebuttal and additional experimental evidence. We have submitted a revised manuscript incorporating the necessary corrections, methodological clarifications, and supporting experimental results. All revisions have been highlighted in blue in the updated manuscript. Due to time constraints, the current revision may still contain minor issues, such as formatting inconsistencies. We will continue to refine the manuscript to further improve its clarity, rigor, and overall presentation. Thank you again for your valuable comments and suggestions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.