Pith. sign in

REVIEW 4 major objections 5 minor 2 cited by

Zero-Shot Image Anomaly Detection Using Generative Foundation Models

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read DiffPathV2 claims that a six-dimensional summary of the error between a diffusion model's predicted Stein score and the ground-truth noise, weighted by (1 − SSIM), separates unseen inlier from outlier images, reaching an average AUROC of…

desk verdict DiffPathV2 reports a plausible 1.8-point AUROC gain over DiffPath, but the paper never defines how a test image is paired with a ground-truth noise epsilon(t), so the central score is not well-defined as written. read the letter →

arxiv 2507.22692 v1 pith:LV3BVZVG submitted 2025-07-30 cs.CV

Lemar Abdi , Amaan Valiuddin , Francisco Caetano , Christiaan Viviers , Fons van der Sommen This is my paper
classification cs.CV
keywords anomalydetectionout-of-distributiondenoisingdiffusionmodelsSteinscoreSSIMweightingzero-shotgenerativefoundationCelebA
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that the denoising trajectory of an unconditional diffusion model can serve as a universal perceptual template for semantic anomaly detection. It claims that a six-dimensional signature made from the error between the predicted Stein score and the ground-truth noise, spatially weighted by (1 − SSIM), separates inlier from outlier images across datasets the model never saw during training. The model is trained once on CelebA and then applied zero-shot to CIFAR-10, CIFAR-100, SVHN, and Textures as inlier or outlier sources, reaching an average AUROC of 94.9. If true, this would mean a single generative model, without retraining, can act as a foundation for open-world vision safety by flagging semantically novel inputs.

What carries the argument

The load-bearing object is the six-dimensional score-error trajectory signature of Eq. (9): the per-timestep mean squared error $\mathrm{mse}(x_0, t) = \|\epsilon_\theta(x_t, t) - \epsilon(t)\|_2^2$ between the denoiser's predicted Stein score (the estimated noise direction) and the ground-truth noise, summed over timesteps in the $\ell^1$, $\ell^2$, and $\ell^3$ norms, together with the same three norms of its time derivative. This signature is then modulated per pixel by $(1 - \mathrm{SSIM}(x_0, \sum_t \epsilon_\theta(x_t, t)))$, where SSIM is the Structural Similarity Index Measure, so that structurally salient regions where noise estimation is poor carry more weight before global pooling. A Gaussian mixture model fitted to the validation set converts the six-dimensional score into a log-likelihood for each test sample. The mechanism's claim is that error statistics, not raw scores, expose the difference between the model's learned manifold and an unseen distribution.

What would settle it

Recompute the DiffPathV2 score after replacing the noise returned by DDIM inversion with any fixed noise tensor of the same shape; if the AUROC separating inliers from outliers stays near 94.9, the claimed dependence on ground-truth noise is not what drives the separation.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is that what discriminates distributions is not the predicted score itself but the way the prediction errs along the diffusion path. For each test image, DiffPathV2 computes the per-timestep mean squared error between the denoiser's output and the ground-truth noise, aggregates the first, second, and third norms of that error and their time derivatives into a six-dimensional score, then multiplies the per-pixel errors by an inverted SSIM map before pooling. The ablation shows that the raw-score baseline already works on far semantic shifts, but the error-plus-SSIM combination sharply separates near-OOD distributions such as CIFAR-10 versus CIFAR-100, lifting that benchmark from 59.0 to 99.4 AUROC. The paper also claims that CelebA is a better base distribution than ImageNet for this trajectory-error scoring, implying a structurally coherent training set sharpens sensitivity to semantic deviations.

Load-bearing premise

The method assumes that every test image can be paired with the actual random noise that produced its noisy version, even though the paper never explains how a real image is matched to that noise.

Editorial extensions

If this is right

  • A single diffusion model trained only on CelebA can separate unseen natural-image datasets, with an average AUROC of 94.9 across the reported benchmarks, so zero-shot anomaly detection needs no target-specific retraining.
  • Near-semantic shifts, the hardest case, improve most: CIFAR-10 versus CIFAR-100 goes from 59.0 AUROC with the raw-score baseline to 99.4 with the error-plus-SSIM score.
  • Using the inverted SSIM map as a spatial amplifier means the method can point to which regions drive the anomaly decision, not just whether a sample is anomalous.
  • Heterogeneous base distributions are not automatically better: CelebA outperforms ImageNet for DiffPath and DiffPathV2, so base-dataset choice should be treated as a tunable property of the detector.
  • The success of score-error statistics over raw scores suggests that likelihood-free trajectory signals carry distributional information that exact-likelihood methods miss.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper never shows how a real test image is paired with a ground-truth noise $\epsilon(t)$; if DDIM inversion only returns a deterministic reconstruction, then Eq. (8) measures consistency with an inverted path rather than error against the true noise, and the theoretical framing would need to change even if the empirical scores stand.
  • The same trajectory-error idea could be tested on other generative paradigms, such as flow matching or consistency models, where a 'true noise' analogue is better defined; a positive result would generalize the claim beyond diffusion.
  • The SSIM weighting suggests a cheap way to build pixel-level anomaly maps: thresholding the per-pixel $(1 - \mathrm{SSIM})$-weighted error could localize semantically anomalous regions without any segmentation training.
  • The CelebA result hints that base distributions with low structural variance act like a sharper reference against which deviations are measured; this is testable by sweeping base datasets with controlled semantic breadth while holding the scoring formula fixed.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The manuscript proposes DiffPathV2, a zero-shot semantic anomaly-detection method built on an unconditional denoising diffusion model trained on CelebA. The method computes, for each test image, a six-dimensional score from the per-timestep mean squared error between the predicted denoising score, epsilon_theta(x_t,t), and a quantity called the ground-truth noise epsilon(t); these errors are spatially weighted by (1 - SSIM) before pooling. A GMM is fitted to validation scores and used to assign anomaly scores to test samples. Experiments on CIFAR-10, SVHN, and CelebA as inlier datasets against several OOD sets report an average AUROC of 94.9, outperforming the DiffPath baseline average of 93.1, and ablations show that CelebA is a more effective base distribution than ImageNet for this method.

Significance. If the method can be defined coherently, the paper contributes a useful data point: a single unconditional diffusion model can serve as a zero-shot semantic anomaly detector, and the choice of base training distribution matters in a non-obvious way. The ablation separating raw-score signals, error signals, and SSIM weighting is informative, and the finding that the proposed SSIM-weighted error signal improves near-OOD separation is genuinely interesting. The paper also honestly reports cases with large headroom and does not overclaim universality. However, the central inference procedure, as written, does not specify how a real test image can be paired with a ground-truth noise realization, so the main quantitative claim is not currently established. The availability of the exact inference protocol, and possibly code, is essential before the state-of-the-art claim can be evaluated.

major comments (4)
  1. [Section 4.1.1, Eq. (8); Algorithm 1 line 12] The quantity epsilon(t) called ground-truth noise cannot be obtained for a real test image by DDIM inversion. During training, epsilon(t) is the independent Gaussian draw in Eq. (5) used to construct x_t from x_0. At test time, x_0 is observed without that draw. DDIM inversion is a deterministic map from x_0 to a model-dependent latent trajectory; it does not recover the random epsilon from the forward process. If epsilon(t) is read off the inversion path as (x_t - sqrt(alpha_bar_t) x_0)/sigma_t, it is a deterministic function of the same model predictions that produce epsilon_theta, not ground truth; if epsilon(t) is taken to be the model's own predicted noise, Eq. (8) is identically zero. The paper must specify the exact test-time construction of epsilon(t), for example by explicitly sampling new forward-process noise draws, or the central anomaly score in Eq. (8) is undefined as written.
  2. [Section 3.2, Eq. (6)] Equation (6) omits the square in the denominator: the conditional score of p(x_t|x_0) = N(sqrt(alpha_bar_t) x_0, sigma_t^2 I) is -(x_t - sqrt(alpha_bar_t) x_0)/sigma_t^2. As printed, the equation is dimensionally inconsistent. This matters because Eq. (7), which defines the ground-truth score epsilon, is derived from Eq. (6); the derivation should be corrected and the sign/scale conventions made explicit.
  3. [Section 4.1.2, Eq. (10); Algorithm 1 lines 6-7] Equation (10) multiplies the global six-dimensional vector s_6D(x_0) by a pixelwise SSIM map, which is dimensionally incompatible. Algorithm 1 correctly applies the (1 - SSIM) weighting before spatial pooling. Equation (10) should be rewritten to show that the SSIM modulation is applied to the per-pixel, per-timestep error maps before the p-norm and temporal aggregation, otherwise the notation does not match the implementation.
  4. [Section 5.1, Table 1; Section 5.3, Table 3] All reported AUROC values are single-run point estimates with no standard deviations, confidence intervals, or seeds. Many entries are exactly 100.0, while other cells take extreme low values such as 00.1 in Table 3, and the claimed average improvement over DiffPath is 94.9 versus 93.1. Without repeated runs or a statistical comparison, the reader cannot assess whether the average gain is reliable or driven by particular benchmark combinations. The paper should report variance across runs or at least describe the number of seeds and the stability of the results.
minor comments (5)
  1. [Section 3.2, Eq. (5)] The notation alpha_bar = product over s of alpha_s is ambiguous: the index s over which the product runs should be stated explicitly, and the maximum timestep should be defined.
  2. [Section 4.2] The paper states that optimal GMM parameters are found by grid search but does not report the grid, the number of components, the covariance type, or the regularization. These details are needed for reproducibility.
  3. [Section 5.2, Figure 2] The histograms in Figure 2 have no axis labels or units; adding x-axis labels such as anomaly score and y-axis labels such as count would make the qualitative claim easier to evaluate.
  4. [Section 4.3.1] The text says the non-foundational baselines are categorized into three distinct baselines but then lists four categories: energy-based, flow-based, diffusion-based, and foundation model-based. This should be corrected.
  5. [Section 5.3] The phrase 'in it's diffusion trajectory' contains a typo; it should be 'its diffusion trajectory.'

Circularity Check

1 steps flagged · score 8.0 of 10

The central anomaly score in Eq. (8) is defined against a 'true noise' that Algorithm 1 obtains from DDIM inversion, a deterministic function of the same model; the score therefore reduces by construction to a self-consistency residual (or zero), not a ground-truth error.

  1. self definitional [Section 4.1.1, Eq. (8); Algorithm 1, lines 11-12]
    "for x0 in Xval do {ϵθ(xt, t), ϵ(t)} ←DDIMInversion(x0, gθ)"

    Eq. (8) defines the anomaly signal as the error between the model's predicted noise and the 'true noise' ϵ(t). In Eqs. (4)-(7), ϵ(t) is an independent Gaussian draw used to construct xt from x0. For a real test image, no such independent draw exists. Algorithm 1 instead supplies ϵ(t) as a second output of DDIMInversion. DDIM inversion is a deterministic map from x0 to a latent trajectory computed with the same network gθ; it does not sample the forward-process noise of Eq. (5). The only noise-like quantity available on that trajectory is the model's own prediction, so the 'error' in Eq. (8) is either identically zero or a self-consistency residual, not an error against ground truth.

full rationale

The paper's derivation chain is short: it hypothesizes that OOD samples induce larger denoising score errors, defines the error in Eq. (8) using an alleged ground-truth noise, and summarizes it in a 6D score. The fatal circular step is the provenance of ϵ(t). The training definition (Eq. 5) makes ϵ(t) an independent random draw, but Algorithm 1 line 12 says DDIMInversion returns it; DDIM inversion is deterministic and model-dependent, so the MSE is not an error against an independent ground truth. This makes the central anomaly score self-referential (and, in the common reading, identically zero), so the headline 94.9 AUROC is not supported by the written method. Other potential concerns: references [1] and [33] are self-citations, but they appear only in related-work context and are not load-bearing; the main comparison is against the external DiffPath baseline [9]. The GMM fitting on the validation set is a standard thresholding step, not a circular prediction. Given that the core quantity in Eq. (8) reduces by definition to the model comparing itself with itself, the circularity score is high, though the flaw is also a correctness/definitional gap that could be repaired by explicitly sampling ϵ(t) from the forward noising process.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The contribution is empirical and heuristic. It assumes transferability of denoising-error statistics across datasets, that OOD samples produce larger score errors, and that an SSIM map computed against accumulated noise predictions localizes semantic deviations. The only fitted components are GMM hyperparameters and manually chosen score norms and weighting. No new entities are introduced.

free parameters (3)
  • GMM hyperparameters = not reported
    Section 4.2: a lightweight GMM is fit to the ID validation set's six-dimensional scores; components and covariance are selected by grid search, with no reported values or sensitivity analysis.
  • Trajectory norm orders = p = 1, 2, 3
    Eq. (9): the six-dimensional score sums the first, second, and third norms of the per-timestep MSE plus their temporal derivatives; these orders are chosen without a dedicated ablation.
  • SSIM modulation = 1 - SSIM (no exponent or temperature)
    Eq. (10): the inverse SSIM map is applied directly; the choice is justified by the ablation table, not by a separate tuning process.
assumptions (5)
  • standard math The denoising score-matching objective equips epsilon_theta with an estimate of the Stein score.
    Section 3.2, Eq. (4), based on Vincent [32]; standard background, not introduced by the paper.
  • domain assumption The denoising trajectory of a diffusion model trained on a single base dataset is informative enough to separate any unseen inlier/outlier dataset pair.
    Invoked in Section 4.1 as the foundation of DiffPathV2; Section 5.3 only ablates the base dataset between ImageNet and CelebA, leaving the general claim outside evidence.
  • domain assumption Test samples from pOOD induce larger Stein score estimation errors than samples from pID.
    Stated in Section 4.1.1: 'If a test sample x0 is drawn from pOOD, its trajectory under the noising process will induce greater errors'. This premise is what makes the score discriminative.
  • domain assumption The SSIM map between the original image and the accumulated predicted noise highlights semantically important error regions.
    Section 4.1.2 postulates that fine-grained structures and high-frequency textures are more indicative of semantic content and that SSIM localizes them; no separate evidence is given.
  • domain assumption The ID validation set is drawn from the same distribution as the ID test set, so a GMM fit on validation scores calibrates test-time likelihoods.
    Section 4.2 fits the GMM on the validation split; any train/validation drift would break the thresholding.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Zero-Shot Image Anomaly Detection Using Generative Foundation Models." pith.science (2026). https://pith.science/paper/LV3BVZVG

@misc{pith2026250722692,
  author       = {Pith},
  title        = {Pith review of: Zero-Shot Image Anomaly Detection Using Generative Foundation Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LV3BVZVG}},
  note         = {Machine review of arXiv:2507.22692}
}
read the original abstract

Detecting out-of-distribution (OOD) inputs is pivotal for deploying safe vision systems in open-world environments. We revisit diffusion models, not as generators, but as universal perceptual templates for OOD detection. This research explores the use of score-based generative models as foundational tools for semantic anomaly detection across unseen datasets. Specifically, we leverage the denoising trajectories of Denoising Diffusion Models (DDMs) as a rich source of texture and semantic information. By analyzing Stein score errors, amplified through the Structural Similarity Index Metric (SSIM), we introduce a novel method for identifying anomalous samples without requiring re-training on each target dataset. Our approach improves over state-of-the-art and relies on training a single model on one dataset -- CelebA -- which we find to be an effective base distribution, even outperforming more commonly used datasets like ImageNet in several settings. Experimental results show near-perfect performance on some benchmarks, with notable headroom on others, highlighting both the strength and future potential of generative foundation models in anomaly detection.

Figures

Figures reproduced from arXiv: 2507.22692 by the authors.

Figure 1
Figure 1. High-level overview of anomaly detection using generative foundation models [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Qualitative comparison of Stein score errors and SSIM [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Uncertainty-Aware Distribution-to-Distribution Flow Matching for Scientific Imaging

    cs.LG 2026-03 unverdicted novelty 6.0 of 10

    Bayesian Stochastic Flow Matching augments flow models with stochastic diffusion for better generalization and uses Monte Carlo Dropout with antithetic sampling to disentangle uncertainties and detect out-of-distribut...

  2. Uncertainty-Aware Distribution-to-Distribution Flow Matching for Scientific Imaging

    cs.LG 2026-03 unverdicted novelty 5.0 of 10

    SFM improves generalization under distribution shift for scientific imaging tasks while AVUQ supplies sample-efficient epistemic and aleatoric uncertainty estimates plus anomaly scores.

Reference graph

Works this paper leans on

37 extracted references · 30 canonical work pages · cited by 1 Pith paper

  1. [1]

    Typicality excels likelihood for unsupervised out-of-distribution detec- tion in medical imaging

    Lemar Abdi, MM Amaan Valiuddin, Christiaan GA Viviers, Peter HN de With, and Fons van der Sommen. Typicality excels likelihood for unsupervised out-of-distribution detec- tion in medical imaging. In Uncertainty for Safe Utiliza- tion of Machine Learning in Medical Imaging: 6th Interna- tional Workshop, Held in Conjunction with MICCAI 2024, Proceedings, pa...

    work page 2024

  2. [2]

    Improving normalizing flows with the approx- imate mass for out-of-distribution detection

    Samy Chali, Inna Kucher, Marc Duranton, and Jacques- Olivier Klein. Improving normalizing flows with the approx- imate mass for out-of-distribution detection. In Proceed- ings of the IEEE/CVF Conference on CVPR, pages 750–758,

    work page

  3. [3]

    Hyunsun Choi, Eric Jang, and Alexander A. Alemi. W AIC, but Why? Generative Ensembles for Robust Anomaly De- tection, 2019. arXiv:1810.01392 [cs, stat]. 2, 5

    arXiv 2019

  4. [4]

    Cimpoi, S

    M. Cimpoi, S. Maji, I. Kokkinos, S. Mohamed, , and A. Vedaldi. Describing textures in the wild. In Proceedings of the IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2014. 5

    work page 2014

  5. [5]

    Nice: Non-linear independent components estimation

    Laurent Dinh, David Krueger, and Yoshua Bengio. Nice: Non-linear independent components estimation. arXiv preprint arXiv:1410.8516, 2014. 5

    arXiv 2014

  6. [6]

    Implicit generation and mod- eling with energy based models

    Yilun Du and Igor Mordatch. Implicit generation and mod- eling with energy based models. Advances in neural infor- mation processing systems, 32, 2019. 5

    work page 2019

  7. [7]

    Improved contrastive divergence training of energy- based models

    Yilun Du, Shuang Li, Joshua Tenenbaum, and Igor Mor- datch. Improved contrastive divergence training of energy- based models. In International Conference on Machine Learning, pages 2837–2848. PMLR, 2021. 5

    work page 2021

  8. [8]

    Denoising diffusion models for out-of-distribution detection

    Mark S Graham, Walter HL Pinaya, Petru-Daniel Tudosiu, Parashkev Nachev, Sebastien Ourselin, and Jorge Cardoso. Denoising diffusion models for out-of-distribution detection. In Proceedings of the IEEE/CVF CVPR , pages 2948–2957,

    work page

Show all 37 references
  1. [9]

    Out-of-distribution detection with a single unconditional diffusion model

    Alvin Heng, Harold Soh, et al. Out-of-distribution detection with a single unconditional diffusion model. Advances in NeurIPS, 37:43952–43974, 2024. 1, 2, 4, 6, 8

    2024

  2. [10]

    Denoising diffu- sion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffu- sion probabilistic models. Advances in NeurIPS , 33:6840– 6851, 2020. 1, 4

    2020

  3. [11]

    Auto-encoding varia- tional bayes, 2022

    Diederik P Kingma and Max Welling. Auto-encoding varia- tional bayes, 2022. 1

    2022

  4. [12]

    Why Normalizing Flows Fail to Detect Out-of-Distribution Data

    Polina Kirichenko, Pavel Izmailov, and Andrew G Wilson. Why Normalizing Flows Fail to Detect Out-of-Distribution Data. In NIPS, pages 20578–20589. Curran Associates, Inc.,

  5. [13]

    Information-theoretic diffusion

    Xianghao Kong, Rob Brekelmans, and Greg Ver Steeg. Information-theoretic diffusion. In The Eleventh Interna- tional Conference on Learning Representations, 2023. 5, 8

    2023

  6. [14]

    Learning multiple layers of features from tiny images

    Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 5

    2009

  7. [15]

    Pseudo numerical methods for diffusion models on manifolds

    Luping Liu, Yi Ren, Zhijie Lin, and Zhou Zhao. Pseudo numerical methods for diffusion models on manifolds. InIn- ternational Conference on Learning Representations , 2022. 1

    2022

  8. [16]

    Deep learning face attributes in the wild

    Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaoou Tang. Deep learning face attributes in the wild. In Proceedings of International Conference on Computer Vision (ICCV), 2015. 5

    2015

  9. [17]

    Unsupervised out-of-distribution detection with diffusion inpainting

    Zhenzhen Liu, Jin Peng Zhou, Yufan Wang, and Kilian Q Weinberger. Unsupervised out-of-distribution detection with diffusion inpainting. In International Conference on Ma- chine Learning, pages 22528–22538. PMLR, 2023. 1, 5

    2023

  10. [18]

    Multiscale score matching for out-of-distribution detection

    Ahsan Mahmood, Junier Oliva, and Martin Andreas Styner. Multiscale score matching for out-of-distribution detection. In International Conference on Learning Representations ,

  11. [19]

    Density of states estimation for out of distribution detection

    Warren Morningstar, Cusuh Ham, Andrew Gallagher, Balaji Lakshminarayanan, Alex Alemi, and Joshua Dillon. Density of states estimation for out of distribution detection. InInter- national Conference on Artificial Intelligence and Statistics, pages 3232–3240. PMLR, 2021. 5

    2021

  12. [20]

    Anomaly detection with conditioned denoising diffusion models

    Arian Mousakhan, Thomas Brox, and Jawad Tayyub. Anomaly detection with conditioned denoising diffusion models. In DAGM German Conference on Pattern Recog- nition, pages 181–195. Springer, 2024. 1, 2

    2024

  13. [21]

    Detecting out-of-distribution inputs to deep generative models using typicality

    Eric Nalisnick, Akihiro Matsukawa, Yee Whye Teh, and Bal- aji Lakshminarayanan. Detecting out-of-distribution inputs to deep generative models using typicality. arXiv preprint arXiv:1906.02994, 2019. 5

    1906 arXiv

  14. [22]

    Nalisnick, Akihiro Matsukawa, Yee Whye Teh, Dilan G¨or¨ur, and Balaji Lakshminarayanan

    Eric T. Nalisnick, Akihiro Matsukawa, Yee Whye Teh, Dilan G¨or¨ur, and Balaji Lakshminarayanan. Do deep generative models know what they don’t know? In ICLR 2019,, 2019. 2

    2019

  15. [23]

    Reading digits in natural images with unsupervised feature learning

    Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bis- sacco, Baolin Wu, Andrew Y Ng, et al. Reading digits in natural images with unsupervised feature learning. In NIPS workshop on deep learning and unsupervised feature learn- ing, page 4. Granada, 2011. 5

    2011

  16. [24]

    Improved denoising diffusion probabilistic models

    Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International conference on machine learning, pages 8162–8171. PMLR,

  17. [25]

    Osada, T

    G. Osada, T. Takahashi, and T. Nishide. Understanding like- lihood of normalizing flow and image complexity through the lens of out-of-distribution detection. Proceedings of the AAAI Conference on Artificial Intelligence , 38(19):21492– 21500, 2024. 2

    2024

  18. [26]

    Likelihood ratios for out-of-distribution detec- tion

    Jie Ren, Peter J Liu, Emily Fertig, Jasper Snoek, Ryan Poplin, Mark Depristo, Joshua Dillon, and Balaji Lakshmi- narayanan. Likelihood ratios for out-of-distribution detec- tion. Advances in neural information processing systems , 32, 2019. 5

    2019

  19. [27]

    A simple fix to mahalanobis distance for improving near-ood detection

    Jie Ren, Stanislav Fort, Jeremiah Liu, Abhijit Guha Roy, Shreyas Padhy, and Balaji Lakshminarayanan. A simple fix to mahalanobis distance for improving near-ood detection. arXiv preprint arXiv:2106.09022, 2021. 2

    2021 arXiv

  20. [28]

    Variational inference with normalizing flows

    Danilo Rezende and Shakir Mohamed. Variational inference with normalizing flows. In International conference on ma- chine learning, pages 1530–1538. PMLR, 2015. 5

    2015

  21. [29]

    Reconstruction- free anomaly detection with diffusion models via direct la- tent likelihood evaluation

    Shunsuke Sakai and Tatsuhito Hasegawa. Reconstruction- free anomaly detection with diffusion models via direct la- tent likelihood evaluation. arXiv preprint arXiv:2504.05662,

  22. [30]

    N ´u˜nez, and Jordi Luque

    Joan Serr `a, David ´Alvarez, Vicenc ¸ G´omez, Olga Slizovskaia, Jos´e F. N ´u˜nez, and Jordi Luque. Input complexity and out-of-distribution detection with likelihood-based genera- tive models. In International Conference on Learning Rep- resentations, 2020. 1, 5

    2020

  23. [31]

    Score-based generative modeling through stochastic differential equa- tions

    Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Ab- hishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equa- tions. In ICLR, 2021. 2, 3, 5

    2021

  24. [32]

    A connection between score matching and denoising autoencoders

    Pascal Vincent. A connection between score matching and denoising autoencoders. Neural computation, 23(7):1661– 1674, 2011. 3

    2011

  25. [33]

    Can your generative model detect out-of-distribution covari- ate shift? In ECCV, pages 184–201

    Christiaan Viviers, Amaan Valiuddin, Francisco Caetano, Lemar Abdi, Lena Filatova, Fons van der Sommen, et al. Can your generative model detect out-of-distribution covari- ate shift? In ECCV, pages 184–201. Springer, 2025. 2

    2025

  26. [34]

    Diffusion models for medical anomaly detection

    Julia Wolleb, Florentin Bieder, Robin Sandk ¨uhler, and Philippe C Cattin. Diffusion models for medical anomaly detection. In MICCAI, pages 35–45. Springer, 2022. 1, 2

    2022

  27. [35]

    {V AEBM}: A symbiosis between variational autoencoders and energy-based models

    Zhisheng Xiao, Karsten Kreis, Jan Kautz, and Arash Vahdat. {V AEBM}: A symbiosis between variational autoencoders and energy-based models. In International Conference on Learning Representations, 2021. 5

    2021

  28. [36]

    Generalized out-of-distribution detection: A survey

    Jingkang Yang, Kaiyang Zhou, Yixuan Li, and Ziwei Liu. Generalized out-of-distribution detection: A survey. Inter- national Journal of Computer Vision , 132(12):5635–5662,

  29. [37]

    Rethinking reconstruction autoencoder-based out-of-distribution detection

    Yibo Zhou. Rethinking reconstruction autoencoder-based out-of-distribution detection. In Proceedings of the IEEE/CVF CVPR, pages 7379–7387, 2022. 2

    2022

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.