Pith. sign in

REVIEW 4 major objections 5 minor 60 references

Training an energy model on full-dimensional DINOv2 tokens stabilizes once an optimal-transport rectified flow reshapes the sampling geometry; the paper reports 98.6%/97.9% AUROC on MVTec-AD and 97.3%/99.0% on VisA

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

ReFP-AD uses rectified-flow preconditioning to make finite-step MCMC stable for energy-based anomaly detection on full-dimensional DINOv2 tokens, achieving strong AUROC on MVTec-AD and VisA.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection A credible geometric fix for token-space EBMs, with a decisive ablation but a checkpoint-selection step that needs multi-seed evidence before the headline numbers are trusted. the 4 major comments →

arxiv 2608.01793 v1 pith:J4MRZF7G submitted 2026-08-03 cs.LG

ReFP-AD: Rectified Flow Preconditioning for Energy-Based Anomaly Detection

classification cs.LG
keywords anomaly detectionenergy-based modelsrectified flowoptimal transportLangevin dynamicsMCMC preconditioningDINOv2 tokensunified multi-class detection
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

ReFP-AD claims that the reason energy-based models fail on modern vision-token features is not model capacity but sampling geometry. Standard energy-model training draws negative samples with short Markov-chain (Langevin) updates, and those dynamics implicitly assume an isotropic Euclidean metric; the 1536-dimensional DINOv2 token space is anisotropic and strongly correlated across dimensions, so finite-step sampling produces unstable negative phases and forces prior work to compress features into low-dimensional bottlenecks. The paper's solution is a learned rectified flow, coupled to data via optimal transport, that transports standardized tokens into a well-conditioned latent space, followed by an unconstrained energy model trained there with persistent contrastive divergence and preconditioned SGLD. Under a strict unified protocol — one shared flow and one shared energy model across all object categories — it reports 98.6%/97.9% image/pixel AUROC on MVTec-AD and 97.3%/99.0% on VisA, beating prior unified energy-based baselines by up to 10.8 points of image AUROC. If correct, the paper demonstrates that explicit density estimation can be competitive in full-dimensional foundation-token spaces, provided the representation geometry is conditioned before the density is learned.

Core claim

The central claim, stated on the paper's own terms, is that the key limitation in training energy-based models on foundation-token representations is sampling geometry, not capacity. In the raw DINOv2 token space, finite-step Langevin dynamics mix poorly — trajectories zig-zag along narrow dimensions and make little progress along elongated correlated directions — so persistent contrastive divergence destabilizes and the learned energy landscape collapses. ReFP-AD learns a geometric reparameterization: an optimal-transport-coupled rectified flow maps Z-score-standardized tokens into an approximately isotropic latent space whose checkpoint is chosen by MCMC-oriented diagnostics (conditioning,

What carries the argument

The central object is the OT-coupled rectified flow used purely as a preconditioner, not as a detector. Data tokens z and Gaussian targets u ~ N(0, τ²I) are paired through a log-domain Sinkhorn optimal-transport coupling, and a velocity network is trained with the rectified-flow matching objective on their linear interpolants; at inference, tokens are transported by integrating the learned ODE with a 10-step RK4 solver. The transported space u is then the home of an unconstrained residual-MLP energy model Eφ(u) trained with persistent contrastive divergence and preconditioned SGLD, where the diagonal preconditioner adds local RMSProp-style scaling on top of the global transport. A secondary

Load-bearing premise

The whole pipeline rests on a single checkpoint-selection step: the flow model is chosen by MCMC diagnostics computed on one fixed random 128-dimensional projection of the 1536-dimensional token space, on the assumption that this projection faithfully mirrors the conditioning and correlation structure of the full space — if that one random subspace is atypical, the preconditioning is mis-calibrated and the reported gains could shrink or vanish on other data or backbones.

What would settle it

Run the full ReFP-AD pipeline on VisA five times, keeping every setting identical except the fixed random 128-dimensional projection used in the SGLD-fitness checkpoint criterion. If the selected flow checkpoint and the final image AUROC vary by more than a couple of points across the five runs, the diagnostic does not faithfully represent the 1536-dimensional sampling geometry, and the preconditioning is only as good as the luck of one projection.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Explicit density-based unified detection becomes competitive in full-dimensional foundation-token spaces: one shared flow and one shared EBM handle all MVTec-AD and VisA categories with only per-category Z-score standardization, and the unified model trails its per-category variant by less than one AUROC point.
  • Stable training needs only short Markov chains: image AUROC on MVTec-AD saturates at 20–40 preconditioned SGLD steps, with 80 steps giving no further gain, so finite-step MCMC ceases to be the bottleneck once geometry is fixed.
  • The recipe transfers across feature spaces: performance degrades only mildly when the backbone shrinks from 1536-D to 768-D and remains stable at 4096-D with a larger backbone, and localization is essentially invariant to input resolution from 224 to 448 pixels.
  • Transport alone carries much of the anomaly signal: scoring by the transported token magnitude ||u||₂ already reaches 96–97.5% image AUROC, and explicit energy learning mainly improves the hardest categories (up to +7.4 AUROC points on the worst VisA category).
  • Flow-trainability is not the right stopping rule: minimizing the flow-matching loss to convergence without the MCMC-fitness checkpoint collapses VisA image AUROC to 73.05%, so the transport must be selected by sampler behavior, not regression accuracy.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The paper's ordering — condition the geometry first, then learn the density — reads as a general template for scaling energy-based or likelihood-based models to any high-dimensional pretrained representation, including text and multimodal embeddings, not just vision tokens.
  • The large gap between pSGLD and plain SGLD after transport (97.3% vs 75.3% VisA image AUROC) suggests global isotropization is necessary but not sufficient; a testable simplification would fold the diagonal preconditioner into the flow itself so that the transport becomes the single conditioning mechanism.
  • The flow-magnitude result implies a cheap inference variant is within reach: keep the learned flow and replace the energy-gradient score with a lightweight functional of the transported tokens, trading some worst-category performance for a large cut in the 299.5 ms per-image latency the paper measures.
  • The SGLD-fitness criterion is applied only at checkpoint selection, so a natural next step — one the paper's own outlook gestures at — is to make the flow's training objective itself sampler-aware, optimizing for finite-step mixing rather than selecting for it afterwards.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper argues that the key obstacle to training unconstrained energy-based models (EBMs) on full-dimensional DINOv2 token features is sampling geometry, not model capacity. It proposes ReFP-AD, which (i) Z-score standardizes per-category tokens, (ii) learns an OT-coupled rectified flow that maps tokens into a well-conditioned latent space, with the flow checkpoint selected by an SGLD-fitness diagnostic, and (iii) trains a residual-MLP EBM using persistent contrastive divergence with preconditioned SGLD, scoring anomalies by the energy gradient norm. Under a strict unified protocol on MVTec-AD and VisA, it reports 98.6%/97.9% and 97.3%/99.0% Image/Pixel AUROC, outperforming a stabilized MPDR baseline by up to +10.8% Image AUROC. Core ablations show that removing the flow drops I-AUROC by 7.5 on MVTec-AD and 10.3 on VisA, and that the SGLD-fitness checkpoint selection is decisive (Suppl. C: selecting the final flow checkpoint instead of the F(u) optimum collapses VisA I-AUROC from 97.3% to 73.05%).

Significance. If the geometric-preconditioning claim holds, the paper makes a useful conceptual contribution: it demonstrates that explicit energy-based density models can be trained directly on dense foundation-token representations without dimensionality reduction, provided the sampling geometry is reshaped first. The paper is honest about limitations (inference cost, computational overhead, remaining gap to reconstruction-based SOTA in Suppl. B) and provides several targeted ablations that directly test the mechanism. Strengths include the decisive checkpoint-sensitivity control in Suppl. C, the backbone scalability study, and the additional datasets in Suppl. D. However, the central checkpoint-selection procedure relies on a single random 128-dimensional projection, and no multi-seed or projection-seed variability is reported. Since Suppl. C shows that checkpoint choice can change VisA I-AUROC by more than 24 points, this is a load-bearing uncertainty rather than a presentation issue. The paper would be substantially strengthened by multi-seed and multi-projection robustness results, and by a more transparent baseline-stabilization protocol.

major comments (4)
  1. [Section 3.3 and Suppl. C] Checkpoint selection is decisive but depends on a single random 128-D projection. The transported tokens u are projected onto a fixed random orthogonal subspace of dimension k=128, and the checkpoint t* is chosen by minimizing F(u) in Eq. (3) subject to the Spearman guardrail. Suppl. C shows that this choice is not incremental: using the final flow checkpoint (min LRF) instead of min F(u) drops VisA I-AUROC from 97.3% to 73.05%. Yet no variance over projection seeds is reported. A single random projection is one draw from a distribution; the reported numbers could be a lucky draw. Please report F(t) and the downstream AUROC across several projection seeds, or otherwise demonstrate that the selected t* is stable with respect to the projection choice.
  2. [Section 3.3, Eq. (3)] The random-projection diagnostic is not merely noisily estimated; it is systematically biased in a way that undermines the conditioning penalty. For a 1536-D covariance with spread eigenvalues, the condition number of a random k-dimensional projection is generally much smaller than the full-space condition number, because projection mixes extreme eigen-directions into the bulk. Thus P_kappa in Eq. (3) does not upper-bound, or even faithfully estimate, the anisotropy of the actual 1536-D space in which the EBM is trained. The claim that the projection 'preserves second-order structure in expectation' is true for the covariance matrix in expectation, but condition number and mean-squared off-diagonal correlation are nonlinear functions, and it does not follow that their expectations are preserved. This matters because checkpoint selection is driven by F(u). Please justify the representativ
  3. [Tables 1-5 and Suppl. D] All reported results appear to be single runs with no error bars or seed counts. The pipeline is stochastic at several levels: flow training, SGLD negative phases, replay buffer dynamics, and checkpoint selection. Given the decisive checkpoint sensitivity documented in Suppl. C, single-run numbers are insufficient to support the claimed margins, e.g., the +10.8% improvement over MPDR† on VisA and the -10.3% ablation drop on VisA. Please report means and standard deviations over at least three independent runs, and ideally across projection seeds as well. This is a reproducibility requirement, not a polish issue.
  4. [Section 4.2, Table 1] The MPDR† baseline requires clarification. The text states that the official MPDR implementation was 'numerically unstable and resulted in near-random performance' and that the authors tuned only SGLD hyperparameters to obtain 86.5% I-AUROC on VisA. Since the headline improvement is computed against this stabilized baseline, the reader needs to know: the exact hyperparameters used, the search range and selection criterion, whether the same validation/test folds were used for tuning, and what the untuned official result was. Without these details, the +10.8% comparison is not firmly interpretable.
minor comments (5)
  1. [Table 1] The HGAD row appears to contain a formatting error: '97.997.1' should likely be '97.9 / 97.1' for MVTec-AD / VisA pixel AUROC. Please fix.
  2. [Section 3.3] The coefficients 1, 1/2, 1/4 in Eq. (3) and the guardrail rho >= 0.6 are described as fixed across all experiments, but no evidence is provided that they were not selected using the same datasets on which the final numbers are reported. If they were chosen after observing MVTec-AD/VisA behavior, the final comparison inherits selection bias. Please state how these constants were chosen or provide a small sensitivity analysis.
  3. [Figure 2] The PCA visualization shows PC-1 and PC-2 capturing only 5.2% and 4.7% of the variance, respectively. This makes it hard to visually support the claim of isotropization. Consider showing eigenvalue spectra or condition-number estimates in addition to, or instead of, the two-dimensional PCA scatter.
  4. [Section 3.1 / 4.1] The protocol is called 'strict unified', but per-category standardization (Eq. 1) and balanced replay-buffer initialization use category identity. The paper states this assumption, but it should be more prominent in the definition of the unified protocol, since it means test-time category labels are required.
  5. [Section 4.3, Table 3] The SGLD-step ablation shows a very sharp transition between K=10 and K=20 (54.7% to 98.2% I-AUROC). A brief discussion of the failure mode at K=10 beyond 'chains fail to mix' would be helpful, especially for readers who want to reproduce the sensitivity.

Circularity Check

0 steps flagged

No significant circularity: the EBM scores are not fitted from the reported AUROC; the flow-checkpoint heuristic is model selection, not a circular reduction.

full rationale

The paper's derivation chain is self-contained against external benchmarks. The preprocessing (Eq. 1), flow objective (Eq. 2), checkpoint-selection criterion (Eq. 3), EBM contrastive objective (Eq. 6), and gradient-norm score (Eq. 7) are all defined independently of the reported AUROC values. The checkpoint-selection functional F(u) is computed on unlabeled transported tokens and used only to pick a flow-training checkpoint; it is not fitted to the test labels or to the final AUROC, so the later detection scores are not forced by construction. The flow's Gaussian target makes the transported space well-conditioned by design, but that is the proposed mechanism, not a circular justification; the paper's claim is that this geometric preconditioning enables stable PCD, which is tested by ablations (removing the flow, varying SGLD steps, removing pSGLD) against the same protocol. The single random 128-D projection used in Eq. 3 is a potential robustness weakness (the condition-number estimate can be biased and seed-dependent), but that is a statistical-reliability concern, not a circular-equivalence concern: the reported anomaly scores still come from an independently trained EBM on full-dimensional tokens. The only self-citation is [26] (SubspaceAD) in related work, and it is not load-bearing. No step in the paper reduces to its own input by definition; therefore no circularity is identified.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

The method is empirical and borrows its main components (rectified flow, pSGLD, PCD, OT coupling) from prior work. Its own contributions are the application to token-space EBMs and the checkpoint-selection heuristic, which rests on several hand-set coefficients and the assumption that a random 128-D projection captures full-space MCMC geometry. No new physical or mathematical entities are introduced.

free parameters (6)
  • target Gaussian temperature tau = 0.1
    Sets the scale and contraction-expansion balance of the rectified flow target; chosen by hand, not derived.
  • Sinkhorn entropic regularization epsilon = 0.01
    Controls the OT coupling approximation quality; fixed without sensitivity analysis.
  • SGLD-fitness coefficients (P_kappa, P_corr, P_tail weights) = 1, 1/2, 1/4
    Hand-selected weights in Eq. (3) assert relative impact of conditioning, correlation, and tails on short-run MCMC; they drive checkpoint selection and thus final results.
  • Spearman rank correlation guardrail rho_min = 0.6
    Threshold to preserve topological structure during flow checkpoint selection; chosen once, not justified by theory.
  • EBM gradient penalty lambda and energy penalty alpha = 10.0, 0.1
    Shape the energy landscape and the gradient-norm score; selected by tuning.
  • diagnostic projection dimension k = 128
    Random subspace used for covariance estimation during checkpoint selection; not varied or justified beyond estimator variance.
axioms (5)
  • domain assumption DINOv2 token representations capture the visual normal manifold at patch level.
    Used implicitly in Section 3.1; the method would be meaningless if tokens did not encode normality.
  • domain assumption A rectified flow trained only on normal tokens transports anomalous tokens into regions distinguishable by energy gradient norms at test time.
    No theoretical guarantee; Section 3.5 relies on this for scoring.
  • ad hoc to paper The SGLD-fitness score F(u) is a valid proxy for downstream EBM anomaly-detection performance.
    Eq. (3) is constructed by hand; Suppl. C shows it correlates with final AUROC better than flow loss, but the correlation is empirical and dataset-specific.
  • domain assumption Finite-step PCD with pSGLD provides a sufficient approximation of the EBM negative phase for anomaly scoring.
    Standard EBM practice; the paper does not analyze approximation error.
  • domain assumption Per-category Z-score standardization is an acceptable use of category identity in a 'unified' protocol.
    Section 3.1; the protocol uses category labels at test time for standardization, which affects the strength of the unified claim.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of ReFP-AD: Rectified Flow Preconditioning for Energy-Based Anomaly Detection." pith.science (2026). https://pith.science/paper/J4MRZF7G

@misc{pith2026260801793,
  author       = {Pith},
  title        = {Pith review of: ReFP-AD: Rectified Flow Preconditioning for Energy-Based Anomaly Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/J4MRZF7G}},
  note         = {Machine review of arXiv:2608.01793}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Unified anomaly detection requires modeling highly heterogeneous normal data without access to anomalous samples. While foundation models like DINOv2 provide rich token representations, leveraging these spaces for explicit density estimation remains challenging. Energy-Based Models (EBMs) offer a principled formulation, but their training in high-dimensional token spaces is unstable due to anisotropy and strong cross-dimensional correlations, which degrades finite-step Markov Chain Monte Carlo (MCMC) sampling. We identify this instability as fundamentally geometric and introduce ReFP-AD (Rectified Flow Preconditioning for Anomaly Detection), which learns a geometric reparameterization that maps high-dimensional embeddings into a well-conditioned latent space via an optimal transport (OT)-coupled rectified flow. This preconditioning enables stable persistent contrastive divergence with preconditioned Stochastic Gradient Langevin Dynamics (SGLD) in full-dimensional token spaces. Anomaly scores are then derived from the learned energy landscape using gradient norms. Under a strict unified protocol on the MVTec-AD and VisA datasets, ReFP-AD achieves 98.6%/97.9% Image/Pixel AUROC on MVTec-AD and 97.3%/99.0% on VisA, outperforming prior unified EBM baselines by up to +10.8% in Image AUROC. Ablation experiments demonstrate that geometric reparameterization is critical for finite-step MCMC and accurate anomaly localization in high-dimensional token spaces. Code is available at https://github.com/CLendering/ReFP-AD

Figures

Figures reproduced from arXiv: 2608.01793 by Camile Lendering, Egor Bondarev, Erkut Akdag, Joaqu\'in Figueira.

Figure 1
Figure 1. Figure 1: Overview of ReFP-AD. Standardized DINOv2 tokens in z space are transported to a better-conditioned latent space u via an OT-coupled rectified flow, with checkpoint selection guided by MCMC-oriented geometric diagnostics. An unconstrained EBM Eϕ(u) is trained using PCD with pSGLD, and anomalies are scored at inference by the gradient norm ∥∇uEϕ(u)∥2 under a fixed finite-step sampling budget. preconditioning… view at source ↗
Figure 2
Figure 2. Figure 2: PCA projection of rectified flow trajectories. Axes correspond to the first two principal components (PC-1: 5.2% variance, PC-2: 4.7% variance). The transport map transforms the correlated, heavy-tailed token distribution (z) into a well-conditioned latent space (u) with reduced anisotropy and correlations, improving finite-step Langevin dynamics. Effect of Rectified Flow Preconditioning. The fundamental h… view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative localization results on MVTec-AD and VisA datasets. From top to bottom: input image, ground-truth mask, predicted anomaly map, and overlay. The proposed model ReFP-AD produces well-localized anomaly maps that closely align with defect regions, while maintaining low scores on normal backgrounds. localization of both subtle texture and structural defects, while suppressing background noise by con… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

60 extracted references · 49 canonical work pages · 3 internal anchors

  1. [1]

    In: 2022 IEEE International Conference on Image Processing (ICIP)

    Akcay,S.,Ameln,D.,Vaidya,A.,Lakshmanan,B.,Ahuja,N.,Genc,U.:Anomalib:A deep learning library for anomaly detection. In: 2022 IEEE International Conference on Image Processing (ICIP). pp. 1706–1710. IEEE (2022) 24

  2. [2]

    In: Proceedings of the IEEE/CVF winter conference on applications of computer vision

    Batzner, K., Heckler, L., König, R.: Efficientad: Accurate visual anomaly detection at millisecond-level latencies. In: Proceedings of the IEEE/CVF winter conference on applications of computer vision. pp. 128–138 (2024) 3

  3. [3]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Bergmann, P., Fauser, M., Sattlegger, D., Steger, C.: Mvtec ad–a comprehensive real-world dataset for unsupervised anomaly detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9592–9600 (2019) 3, 8, 24, 27

  4. [4]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9650–9660 (2021) 2, 3

  5. [5]

    Advances in neural information processing systems26(2013) 6

    Cuturi, M.: Sinkhorn distances: Lightspeed computation of optimal transport. Advances in neural information processing systems26(2013) 6

  6. [6]

    In: 2025 IEEE/CVF Winter Con- ference on Applications of Computer Vision (WACV)

    Damm, S., Laszkiewicz, M., Lederer, J., Fischer, A.: Anomalydino: Boosting patch- based few-shot anomaly detection with dinov2. In: 2025 IEEE/CVF Winter Con- ference on Applications of Computer Vision (WACV). pp. 1319–1329. IEEE (2025) 3, 8 16 C. Lendering et al

  7. [7]

    In: International conference on pattern recognition

    Defard, T., Setkov, A., Loesch, A., Audigier, R.: Padim: a patch distribution modeling framework for anomaly detection and localization. In: International conference on pattern recognition. pp. 475–489. Springer (2021) 3, 10

  8. [8]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Deng,H.,Li,X.:Anomalydetectionviareversedistillationfromone-classembedding. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9737–9746 (2022) 21

  9. [9]

    arXiv preprint arXiv:2012.01316 (2020) 2

    Du, Y., Li, S., Tenenbaum, J., Mordatch, I.: Improved contrastive divergence training of energy based models. arXiv preprint arXiv:2012.01316 (2020) 2

  10. [10]

    Advances in neural information processing systems32(2019) 2

    Du, Y., Mordatch, I.: Implicit generation and modeling with energy based models. Advances in neural information processing systems32(2019) 2

  11. [11]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Gao, R., Nijkamp, E., Kingma, D.P., Xu, Z., Dai, A.M., Wu, Y.N.: Flow contrastive estimation of energy-based models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7518–7528 (2020) 4

  12. [12]

    Energy-Based Anomaly Detection and Localization

    Genc, E.U., Ahuja, N., Ndiour, I.J., Tickoo, O.: Energy-based anomaly detection and localization. arXiv preprint arXiv:2105.03270 (2021) 10

  13. [13]

    arXiv preprint arXiv:1912.03263 (2019) 4, 8

    Grathwohl, W., Wang, K.C., Jacobsen, J.H., Duvenaud, D., Norouzi, M., Swersky, K.: Your classifier is secretly an energy based model and you should treat it like one. arXiv preprint arXiv:1912.03263 (2019) 4, 8

  14. [14]

    In: Proceedings of the IEEE/CVF winter conference on applications of computer vision

    Gudovskiy, D., Ishizaka, S., Kozuka, K.: Cflow-ad: Real-time unsupervised anomaly detection with localization via conditional normalizing flows. In: Proceedings of the IEEE/CVF winter conference on applications of computer vision. pp. 98–107 (2022) 4, 10, 24

  15. [15]

    Advances in Neural Information Processing Systems36, 10721–10740 (2023) 21

    Guo, J., Lu, S., Jia, L., Zhang, W., Li, H.: Recontrast: Domain-specific anomaly detection via contrastive reconstruction. Advances in Neural Information Processing Systems36, 10721–10740 (2023) 21

  16. [16]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Guo, J., Lu, S., Zhang, W., Chen, F., Li, H., Liao, H.: Dinomaly: The less is more philosophy in multi-class unsupervised anomaly detection. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 20405–20415 (2025) 21

  17. [17]

    Advances in Neural Information Processing Systems37, 71162–71187 (2024) 21

    He, H., Bai, Y., Zhang, J., He, Q., Chen, H., Gan, Z., Wang, C., Li, X., Tian, G., Xie, L.: Mambaad: Exploring state space models for multi-class unsupervised anomaly detection. Advances in Neural Information Processing Systems37, 71162–71187 (2024) 21

  18. [18]

    In: Proceedings of the AAAI conference on artificial intelligence

    He, H., Zhang, J., Chen, H., Chen, X., Li, Z., Chen, X., Wang, Y., Wang, C., Xie, L.: A diffusion-based framework for multi-class anomaly detection. In: Proceedings of the AAAI conference on artificial intelligence. vol. 38, pp. 8472–8480 (2024) 4

  19. [19]

    arXiv preprint arXiv:2503.21622 (2025) 23, 24

    Heckler-Kram, L., Neudeck, J.H., Scheler, U., König, R., Steger, C.: The mvtec ad 2 dataset: Advanced scenarios for unsupervised anomaly detection. arXiv preprint arXiv:2503.21622 (2025) 23, 24

  20. [20]

    arXiv preprint arXiv:1903.03704 (2019) 4, 6

    Hoffman, M., Sountsov, P., Dillon, J.V., Langmore, I., Tran, D., Vasudevan, S.: Neutra-lizing bad geometry in hamiltonian monte carlo using neural transport. arXiv preprint arXiv:1903.03704 (2019) 4, 6

  21. [21]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Jeong, J., Zou, Y., Kim, T., Zhang, D., Ravichandran, A., Dabeer, O.: Winclip: Zero-/few-shot anomaly classification and segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 19606–19616 (2023) 3

  22. [22]

    Advances in Neural Information Processing Systems35, 15433–15445 (2022) 3

    Jiang, X., Liu, J., Wang, J., Nie, Q., Wu, K., Liu, Y., Wang, C., Zheng, F.: Softpatch: Unsupervised anomaly detection with noisy data. Advances in Neural Information Processing Systems35, 15433–15445 (2022) 3

  23. [23]

    Advances in neural information processing systems33, 20578–20589 (2020) 4 ReFP-AD: Flow-Preconditioned EBMs for Anomaly Detection 17

    Kirichenko, P., Izmailov, P., Wilson, A.G.: Why normalizing flows fail to detect out-of-distribution data. Advances in neural information processing systems33, 20578–20589 (2020) 4 ReFP-AD: Flow-Preconditioned EBMs for Anomaly Detection 17

  24. [24]

    CR Acad Sci (Paris)146, 530 (1908) 2

    Langevin, P.: On the theory of brownian motion. CR Acad Sci (Paris)146, 530 (1908) 2

  25. [25]

    Predicting structured data1(0) (2006) 2, 4

    LeCun, Y., Chopra, S., Hadsell, R., Ranzato, M., Huang, F., et al.: A tutorial on energy-based learning. Predicting structured data1(0) (2006) 2, 4

  26. [26]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Lendering, C., Akdag, E., Bondarau, E.: Subspacead: Training-free few-shot anomaly detection via subspace modeling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 28557–28566 (2026) 3

  27. [27]

    In: Proceedings of the AAAI conference on artificial intelligence

    Li, C., Chen, C., Carlson, D., Carin, L.: Preconditioned stochastic gradient langevin dynamics for deep neural networks. In: Proceedings of the AAAI conference on artificial intelligence. vol. 30 (2016) 2, 6, 7, 13

  28. [28]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Li, X., Zhang, Z., Tan, X., Chen, C., Qu, Y., Xie, Y., Ma, L.: Promptad: Learning prompts with only normal samples for few-shot anomaly detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16838–16848 (2024) 3

  29. [29]

    Advances in Neural Information Processing Systems38, 87834–87899 (2026) 4

    Li, Z., Huang, Q., Zhu, Y., Yang, L., Mohammadi Amiri, M., van Stein, N., van Leeuwen, M.: Scalable, explainable and provably robust anomaly detection with one-step flow matching. Advances in Neural Information Processing Systems38, 87834–87899 (2026) 4

  30. [30]

    arXiv preprint arXiv:2210.02747 (2022) 4

    Lipman, Y., Chen, R.T., Ben-Hamu, H., Nickel, M., Le, M.: Flow matching for generative modeling. arXiv preprint arXiv:2210.02747 (2022) 4

  31. [31]

    Advances in neural information processing systems33, 21464–21475 (2020) 4

    Liu, W., Wang, X., Owens, J., Li, Y.: Energy-based out-of-distribution detection. Advances in neural information processing systems33, 21464–21475 (2020) 4

  32. [32]

    arXiv preprint arXiv:2209.03003 (2022) 2, 4, 6

    Liu, X., Gong, C., Liu, Q.: Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003 (2022) 2, 4, 6

  33. [33]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Liu, Z., Zhou, Y., Xu, Y., Wang, Z.: Simplenet: A simple network for image anomaly detection and localization. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 20402–20411 (2023) 12, 14

  34. [34]

    INP-Former++: Advancing Universal Anomaly Detection via Intrinsic Normal Prototypes and Residual Learning

    Luo, W., Yao, H., Cao, Y., Chen, Q., Gao, A., Shen, W., Yu, W.: Inp-former++: Advancing universal anomaly detection via intrinsic normal prototypes and residual learning. arXiv preprint arXiv:2506.03660 (2025) 21

  35. [35]

    Advances in neural information processing systems28(2015) 2

    Ma, Y.A., Chen, T., Fox, E.: A complete recipe for stochastic gradient mcmc. Advances in neural information processing systems28(2015) 2

  36. [36]

    In: DAGM German Conference on Pattern Recognition

    Mousakhan, A., Brox, T., Tayyub, J.: Anomaly detection with conditioned denoising diffusion models. In: DAGM German Conference on Pattern Recognition. pp. 181–

  37. [37]

    Nalisnick, E., Matsukawa, A., Teh, Y.W., Gorur, D., Lakshminarayanan, B.: Do deep generative models know what they don’t know? arXiv preprint arXiv:1810.09136 (2018) 4

  38. [38]

    MCMC Should Mix: Learning Energy-Based Model with Neural Transport Latent Space MCMC

    Nijkamp, E., Gao, R., Sountsov, P., Vasudevan, S., Pang, B., Zhu, S.C., Wu, Y.N.: Mcmc should mix: Learning energy-based model with neural transport latent space mcmc. arXiv preprint arXiv:2006.06897 (2020) 4

  39. [39]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Nijkamp, E., Hill, M., Han, T., Zhu, S.C., Wu, Y.N.: On the anatomy of mcmc-based maximum likelihood learning of energy-based models. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 34, pp. 5272–5280 (2020) 4

  40. [40]

    Advances in Neural Information Processing Systems32(2019) 2, 12

    Nijkamp, E., Hill, M., Zhu, S.C., Wu, Y.N.: Learning non-convergent non-persistent short-run mcmc toward energy-based model. Advances in Neural Information Processing Systems32(2019) 2, 12

  41. [41]

    arXiv preprint arXiv:2304.07193 (2023) 2, 3 18 C

    Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023) 2, 3 18 C. Lendering et al

  42. [42]

    Park, N., Kim, S.: How do vision transformers work? In: International Confer- ence on Learning Representations (2022),https://openreview.net/forum?id= D78Go4hVcxO2

  43. [43]

    Bernoulli2(4), 341–363 (Decem- ber 1996)

    Roberts, G.O., Tweedie, R.L.: Exponential convergence of Langevin distri- butions and their discrete approximations. Bernoulli2(4), 341–363 (Decem- ber 1996). https : / / doi . org / 10 . 2307 / 3318418, https : / / projecteuclid . org/journals/bernoulli/volume- 2/issue- 4/Exponential- convergence- of- Langevin - distributions - and - their - discrete - a...

  44. [44]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Roth, K., Pemula, L., Zepeda, J., Schölkopf, B., Brox, T., Gehler, P.: Towards total recall in industrial anomaly detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14318–14328 (2022) 2, 3, 14

  45. [45]

    In: Proceedings of the IEEE/CVF winter conference on applications of computer vision

    Rudolph, M., Wandt, B., Rosenhahn, B.: Same same but differnet: Semi-supervised defect detection with normalizing flows. In: Proceedings of the IEEE/CVF winter conference on applications of computer vision. pp. 1907–1916 (2021) 4

  46. [46]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Salehi,M.,Sadjadi,N.,Baselizadeh,S.,Rohban,M.H.,Rabiee,H.R.:Multiresolution knowledge distillation for anomaly detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14902–14912 (2021) 10

  47. [47]

    In: European conference on computer vision

    Sträter, L.P., Salehi, M., Gavves, E., Snoek, C.G., Asano, Y.M.: Generalad: Anomaly detection across domains by attending to distorted features. In: European conference on computer vision. pp. 448–465. Springer (2024) 3

  48. [48]

    Advances in neural information processing systems30(2017) 6

    Tarvainen, A., Valpola, H.: Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems30(2017) 6

  49. [49]

    In: Proceedings of the 25th international conference on Machine learning

    Tieleman, T.: Training restricted boltzmann machines using approximations to the likelihood gradient. In: Proceedings of the 25th international conference on Machine learning. pp. 1064–1071 (2008) 2, 7

  50. [50]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Wang, C., Zhu, W., Gao, B.B., Gan, Z., Zhang, J., Gu, Z., Qian, S., Chen, M., Ma, L.: Real-iad: A real-world multi-view dataset for benchmarking versatile industrial anomaly detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22883–22892 (2024) 3, 23

  51. [51]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Wang, X., Chai, L., Chen, J.: Debiasing trace guidance: Top-down trace distilla- tion and bottom-up velocity alignment for unsupervised anomaly detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 22989–22998 (2025) 4

  52. [52]

    In: European Conference on Computer Vision

    Yao, X., Li, R., Qian, Z., Wang, L., Zhang, C.: Hierarchical gaussian mixture normalizing flow modeling for unified anomaly detection. In: European Conference on Computer Vision. pp. 92–108. Springer (2024) 9, 10, 24, 27, 28

  53. [53]

    Advances in Neural Information Processing Systems36, 49445–49466 (2023) 2, 4, 9, 10, 24, 27, 28

    Yoon, S., Jin, Y.U., Noh, Y.K., Park, F.: Energy-based models for anomaly detection: A manifold diffusion recovery approach. Advances in Neural Information Processing Systems36, 49445–49466 (2023) 2, 4, 9, 10, 24, 27, 28

  54. [54]

    Advances in Neural Information Processing Systems 35, 4571–4584 (2022) 21

    You, Z., Cui, L., Shen, Y., Yang, K., Lu, X., Zheng, Y., Le, X.: A unified model for multi-class anomaly detection. Advances in Neural Information Processing Systems 35, 4571–4584 (2022) 21

  55. [55]

    arXiv preprint arXiv:2111.07677 (2021) 4, 10, 24

    Yu, J., Zheng, Y., Wang, X., Li, W., Wu, Y., Zhao, R., Wu, L.: Fastflow: Unsuper- vised anomaly detection and localization via 2d normalizing flows. arXiv preprint arXiv:2111.07677 (2021) 4, 10, 24

  56. [56]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Zavrtanik, V., Kristan, M., Skočaj, D.: Draem-a discriminatively trained reconstruc- tion embedding for surface anomaly detection. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 8330–8339 (2021) 10 ReFP-AD: Flow-Preconditioned EBMs for Anomaly Detection 19

  57. [57]

    IEEE transactions on pattern analysis and machine intelligence (2025) 4

    Zhang, H., Wang, Z., Zeng, D., Wu, Z., Jiang, Y.G.: Diffusionad: Norm-guided one-step denoising diffusion for anomaly detection. IEEE transactions on pattern analysis and machine intelligence (2025) 4

  58. [58]

    arXiv preprint arXiv:2312.07495 (2023) 21

    Zhang, J., Chen, X., Wang, Y., Wang, C., Liu, Y., Li, X., Yang, M.H., Tao, D.: Exploring plain vit reconstruction for multi-class unsupervised anomaly detection. arXiv preprint arXiv:2312.07495 (2023) 21

  59. [59]

    arXiv preprint arXiv:2310.18961 (2023) 3

    Zhou, Q., Pang, G., Tian, Y., He, S., Chen, J.: Anomalyclip: Object-agnostic prompt learning for zero-shot anomaly detection. arXiv preprint arXiv:2310.18961 (2023) 3

  60. [60]

    In: European conference on computer vision

    Zou, Y., Jeong, J., Pemula, L., Zhang, D., Dabeer, O.: Spot-the-difference self- supervised pre-training for anomaly detection and segmentation. In: European conference on computer vision. pp. 392–408. Springer (2022) 3, 8, 27 Supplementary Material for ReFP-AD: Rectified Flow Preconditioning for Energy-Based Anomaly Detection Camile Lendering, Erkut Akda...

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.