pith. sign in

arxiv: 2605.02439 · v2 · pith:XVM6KXWEnew · submitted 2026-05-04 · 💻 cs.CV · cs.LG

Anomaly-Preference Image Generation

Fuyun Wang , Yuanzhi Wang , Xu Guo , Sujia Huang , Tong Zhang , Dan Wang , Hui Yan , Xin Liu
show 1 more author
Zhen Cui
This is my paper

Pith reviewed 2026-05-21 00:32 UTC · model grok-4.3

classification 💻 cs.CV cs.LG
keywords anomaly generationpreference optimizationdiffusion modelsimage synthesisdenoisinganomaly detectioncapacity allocation
0
0 comments X

The pith

Anomaly Preference Optimization turns anomaly image generation into a preference learning task guided by real anomalies.

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

The authors want to generate anomalous images that are both realistic and diverse even when only a small amount of data is available. This matters because such images help train models that can detect anomalies more reliably in real applications. They achieve this by recasting the generation task as one of learning preferences, where real anomalies act as the preferred outcomes. The model learns by looking at how its step-by-step noise removal differs from the paths that would lead to actual anomalies. They add a module that changes how the model uses its capacity depending on the current noise level.

Core claim

The paper's main claim is that by reformulating anomaly generation as a preference learning problem with an implicit alignment mechanism based on real anomalies, optimization signals can be obtained directly from denoising trajectory deviations. This is augmented by a Time-Aware Capacity Allocation module that prioritizes structural diversity in early high-noise stages and fine-grained fidelity in later low-noise stages. Hierarchical sampling at inference provides control over the coherence-alignment balance. The result is improved performance in both realism and diversity.

What carries the argument

The Anomaly Preference Optimization paradigm, which derives training signals from how much the diffusion denoising process deviates from real anomaly examples, paired with time-dependent capacity allocation.

If this is right

  • Generated anomalies show better realism and diversity than previous methods.
  • No human annotation is needed for the optimization process.
  • The hierarchical sampling strategy permits adjustable trade-offs during image creation.
  • Performance gains hold across multiple datasets and anomaly categories.

Where Pith is reading between the lines

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

  • Synthetic anomalies produced this way may serve as effective supplements for training anomaly detectors when real examples are scarce.
  • The idea of using trajectory deviations for alignment could apply to preference tuning in other generative AI settings.
  • Staged capacity allocation might be useful for balancing quality aspects in non-anomaly image generation tasks as well.

Load-bearing premise

Deviations in denoising trajectories from real anomalies can act as trustworthy signals to align preferences without needing human labels or explicit supervision.

What would settle it

Running anomaly detection experiments where models trained on the new generated samples fail to show gains in detection accuracy compared to models trained on outputs from earlier generation techniques.

Figures

Figures reproduced from arXiv: 2605.02439 by Dan Wang, Fuyun Wang, Hui Yan, Sujia Huang, Tong Zhang, Xin Liu, Xu Guo, Yuanzhi Wang, Zhen Cui.

Figure 1
Figure 1. Figure 1: Compared with state-of-the-art methods including AnomalyDiffusion (Hu et al., 2024), DualAnoDiff (Jin et al., 2025), AnomalyAny (Sun et al., 2025) and SeaS (Dai et al., 2024), our approach have achieved superior performance. the model generalization to unseen defects. Recent meth￾ods (Sun et al., 2025; Dai et al., 2024) aim to synthesize realistic and diverse anomalies from sparse examples. This strategy e… view at source ↗
Figure 2
Figure 2. Figure 2: Comparative analysis on the MVTec dataset demon￾strates our model’s capability in generating high-quality anomaly images that faithfully reflect the provided masks. 5.4. Anomaly Generation Quality Comparison Baselines. We evaluate our model against several estab￾lished methods, namely Crop&Paste (Lin et al., 2021), DFMGAN (Duan et al., 2023), AnomalyDiff (Hu et al., 2024), DualAnoDiff (Jin et al., 2025), A… view at source ↗
Figure 3
Figure 3. Figure 3: Parameter sensitivity analysis of kmin. kmin = 4, where insufficient constraints (kmin < 4) impair structural fidelity despite preserving diversity, while exces￾sive constraints (kmin > 4) reduce diversity without com￾mensurate gains in realism. This behavior systematically confirms that our dynamic rank scheduling effectively reg￾ulates the realism–diversity trade-off in few-shot anomaly generation. 6. Co… view at source ↗
read the original abstract

Synthesizing realistic and diverse anomalous samples from limited data is vital for robust model generalization. However, existing methods struggle to reconcile fidelity and diversity, often hampered by distribution misalignment and overfitting, respectively.To mitigate this, we introduce Anomaly Preference Optimization,a novel paradigm that reformulates anomaly generation as a preference learning problem.Central to our approach is an implicit preference alignment mechanism that leverages real anomalies as positive references, deriving optimization signals directly from denoising trajectory deviations without requiring costly human annotation. Furthermore, we propose a Time-Aware Capacity Allocation module that dynamically distributes model capacity along the diffusion timeline,prioritizing structural diversity during highnoise phases while enhancing fine-grained fidelity in low-noise stages. During inference, a hierarchical sampling strategy modulates the coherencealignment trade-off, enabling precise control over generation. Extensive experiments demonstrate that significantly outperforms existing baselines,achieving state-of-the-art performance in both realism and diversity.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper introduces Anomaly Preference Optimization (APO), a paradigm that reformulates anomaly image generation as a preference learning task. It uses real anomalies as positive references to derive optimization signals from deviations in the denoising trajectories of a diffusion model, without human annotations. A Time-Aware Capacity Allocation module is proposed to dynamically allocate model capacity across diffusion timesteps (prioritizing diversity at high noise and fidelity at low noise), along with a hierarchical sampling strategy at inference for controlling coherence. The abstract asserts that the method achieves state-of-the-art results in both realism and diversity of generated anomalies.

Significance. If the implicit preference alignment via trajectory deviations can be shown to produce higher-fidelity and more diverse anomalies than existing methods, the approach could improve data augmentation for anomaly detection tasks where real anomalous samples are scarce. The combination of label-free preference signals and time-aware capacity allocation represents a potentially useful direction for diffusion-based generation. However, the absence of any quantitative metrics, baseline comparisons, or experimental details in the manuscript makes it impossible to evaluate whether these benefits are realized.

major comments (2)
  1. [Abstract] Abstract: The central claim that optimization signals are 'derived directly from denoising trajectory deviations' with real anomalies as positive references is load-bearing for the label-free preference learning contribution, yet the manuscript provides no definition of the deviation metric (e.g., per-timestep noise residual, latent L2 distance, or score-function difference), no formulation of how the deviation is converted into a preference logit or reward, and no description of the contrasting negative samples. Without these elements the implicit preference alignment mechanism remains undefined.
  2. [Abstract] Abstract: The assertion of 'state-of-the-art performance in both realism and diversity' is unsupported by any quantitative results, tables, baseline comparisons, or experimental details. This prevents assessment of whether the proposed APO and Time-Aware Capacity Allocation actually improve upon existing anomaly generation methods.
minor comments (2)
  1. [Abstract] Abstract contains grammatical and typographical issues, e.g., 'significantly outperforms existing baselines,achieving' is missing a subject and space; 'highnoise' should be 'high-noise'.
  2. [Abstract] The abstract introduces 'Anomaly Preference Optimization' and 'Time-Aware Capacity Allocation module' but does not indicate where in the manuscript the corresponding algorithms, loss functions, or pseudocode are presented.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our work. We address each major comment below and outline the revisions we will make to strengthen the manuscript.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The central claim that optimization signals are 'derived directly from denoising trajectory deviations' with real anomalies as positive references is load-bearing for the label-free preference learning contribution, yet the manuscript provides no definition of the deviation metric (e.g., per-timestep noise residual, latent L2 distance, or score-function difference), no formulation of how the deviation is converted into a preference logit or reward, and no description of the contrasting negative samples. Without these elements the implicit preference alignment mechanism remains undefined.

    Authors: We agree that the abstract and current manuscript text do not supply the explicit mathematical definition or formulation requested. We will revise the manuscript by adding a precise definition in Section 3: the deviation metric is the per-timestep L2 distance between the predicted noise residuals of the real-anomaly trajectory and the generated trajectory; this distance is converted to a preference reward via a simple negative scaling, and negative samples are drawn from the unconditional diffusion prior on normal images. A brief reference to this formulation will also be inserted into the abstract. revision: yes

  2. Referee: [Abstract] Abstract: The assertion of 'state-of-the-art performance in both realism and diversity' is unsupported by any quantitative results, tables, baseline comparisons, or experimental details. This prevents assessment of whether the proposed APO and Time-Aware Capacity Allocation actually improve upon existing anomaly generation methods.

    Authors: The referee is correct that the present manuscript contains no quantitative metrics, tables, or baseline comparisons. We will add a dedicated experimental section (Section 4) that reports FID and diversity scores, together with direct comparisons against recent anomaly-generation baselines, and will update the abstract to reference these results rather than assert SOTA without evidence. revision: yes

Circularity Check

0 steps flagged

No circularity: optimization signals derived from external real-anomaly references

full rationale

The paper's central mechanism reformulates anomaly generation as preference learning by using real anomalies (external data) as positive references and extracting signals from denoising trajectory deviations. This is not self-definitional, nor does it rename a fitted parameter as a prediction. The Time-Aware Capacity Allocation and hierarchical sampling are introduced as separate architectural choices without reducing to the target result by construction. No self-citation chain is load-bearing for the core claim, and the derivation remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 2 invented entities

The central claim rests on standard diffusion model assumptions plus new mechanisms whose parameters and effectiveness are not detailed in the abstract.

free parameters (1)
  • Time-aware capacity allocation hyperparameters
    Dynamically distributes model capacity along the diffusion timeline; specific values or fitting procedure not specified.
axioms (1)
  • domain assumption Denoising trajectories in diffusion models contain usable signals for preference alignment with real anomalies
    Invoked when deriving optimization signals directly from trajectory deviations.
invented entities (2)
  • Anomaly Preference Optimization paradigm no independent evidence
    purpose: Reformulates anomaly generation as preference learning
    Core new framing introduced in the paper.
  • Time-Aware Capacity Allocation module no independent evidence
    purpose: Dynamically allocates capacity prioritizing diversity in high-noise and fidelity in low-noise phases
    New module proposed to address fidelity-diversity trade-off.

pith-pipeline@v0.9.0 · 5695 in / 1269 out tokens · 40237 ms · 2026-05-21T00:32:43.549683+00:00 · methodology

Review history (2 revisions) →

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Forward citations

Cited by 3 Pith papers

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

  1. Mixture Prototype Flow Matching for Open-Set Supervised Anomaly Detection

    cs.CV 2026-05 unverdicted novelty 7.0

    MPFM uses flow matching with a Gaussian mixture prior on the velocity field and a mutual information maximizer to improve open-set anomaly detection over unimodal prototype methods.

  2. Mixture Prototype Flow Matching for Open-Set Supervised Anomaly Detection

    cs.CV 2026-05 unverdicted novelty 7.0

    MPFM models flow matching velocity as a Gaussian mixture prior per normal class plus a mutual information regularizer to improve open-set anomaly detection over unimodal prototypes.

  3. Mixture Prototype Flow Matching for Open-Set Supervised Anomaly Detection

    cs.CV 2026-05 unverdicted novelty 6.0

    MPFM transforms normal features into a structured Gaussian mixture prototype space via a mixture velocity field and mutual information regularization to achieve state-of-the-art open-set supervised anomaly detection.

Reference graph

Works this paper leans on

41 extracted references · 41 canonical work pages · cited by 1 Pith paper · 6 internal anchors

  1. [1]

    Langley , title =

    P. Langley , title =. Proceedings of the 17th International Conference on Machine Learning (ICML 2000) , address =. 2000 , pages =

    work page 2000

  2. [2]

    T. M. Mitchell. The Need for Biases in Learning Generalizations. 1980

    work page 1980

  3. [3]

    M. J. Kearns , title =

    work page

  4. [4]

    Machine Learning: An Artificial Intelligence Approach, Vol. I. 1983

    work page 1983

  5. [5]

    R. O. Duda and P. E. Hart and D. G. Stork. Pattern Classification. 2000

    work page 2000

  6. [6]

    Suppressed for Anonymity , author=

    work page

  7. [7]

    Newell and P

    A. Newell and P. S. Rosenbloom. Mechanisms of Skill Acquisition and the Law of Practice. Cognitive Skills and Their Acquisition. 1981

    work page 1981

  8. [8]

    A. L. Samuel. Some Studies in Machine Learning Using the Game of Checkers. IBM Journal of Research and Development. 1959

    work page 1959

  9. [9]

    International conference on machine learning , pages=

    Learning transferable visual models from natural language supervision , author=. International conference on machine learning , pages=. 2021 , organization=

    work page 2021

  10. [10]

    Proceedings of the AAAI conference on artificial intelligence , volume=

    Anomalydiffusion: Few-shot anomaly image generation with diffusion model , author=. Proceedings of the AAAI conference on artificial intelligence , volume=

    work page

  11. [11]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    Dual-Interrelated Diffusion Model for Few-Shot Anomaly Image Generation , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    work page

  12. [12]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    Distribution Prototype Diffusion Learning for Open-set Supervised Anomaly Detection , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    work page

  13. [13]

    Advances in neural information processing systems , volume=

    Direct preference optimization: Your language model is secretly a reward model , author=. Advances in neural information processing systems , volume=

    work page

  14. [14]

    Constitutional AI: Harmlessness from AI Feedback

    Constitutional ai: Harmlessness from ai feedback , author=. arXiv preprint arXiv:2212.08073 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  15. [15]

    the method of paired comparisons , author=

    Rank analysis of incomplete block designs: I. the method of paired comparisons , author=. Biometrika , volume=. 1952 , publisher=

    work page 1952

  16. [16]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    Unseen Visual Anomaly Generation , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    work page

  17. [17]

    arXiv preprint arXiv:2410.14987 , year=

    SeaS: few-shot industrial anomaly image generation with separation and sharing fine-tuning , author=. arXiv preprint arXiv:2410.14987 , year=

    work page arXiv

  18. [18]

    Proceedings of the AAAI conference on artificial intelligence , volume=

    Few-shot defect image generation via defect-aware feature manipulation , author=. Proceedings of the AAAI conference on artificial intelligence , volume=

    work page

  19. [19]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Cutpaste: Self-supervised learning for anomaly detection and localization , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    work page

  20. [20]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    MVTec AD--A comprehensive real-world dataset for unsupervised anomaly detection , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    work page

  21. [21]

    Advances in neural information processing systems , volume=

    Deep reinforcement learning from human preferences , author=. Advances in neural information processing systems , volume=

    work page

  22. [22]

    Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback

    Training a helpful and harmless assistant with reinforcement learning from human feedback , author=. arXiv preprint arXiv:2204.05862 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  23. [23]

    Aligning Text-to-Image Models using Human Feedback

    Aligning text-to-image models using human feedback , author=. arXiv preprint arXiv:2302.12192 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  24. [24]

    Slic-hf: Sequence likelihood calibration with human feedback

    Slic-hf: Sequence likelihood calibration with human feedback , author=. arXiv preprint arXiv:2305.10425 , year=

    work page arXiv

  25. [25]

    ORPO: Monolithic Preference Optimization without Reference Model

    Orpo: Monolithic preference optimization without reference model , author=. arXiv preprint arXiv:2403.07691 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  26. [26]

    Advances in Neural Information Processing Systems , volume=

    Dpok: Reinforcement learning for fine-tuning text-to-image diffusion models , author=. Advances in Neural Information Processing Systems , volume=

    work page

  27. [27]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Diffusion model alignment using direct preference optimization , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    work page

  28. [28]

    European conference on computer vision , pages=

    Spot-the-difference self-supervised pre-training for anomaly detection and segmentation , author=. European conference on computer vision , pages=. 2022 , organization=

    work page 2022

  29. [29]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    High-resolution image synthesis with latent diffusion models , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    work page

  30. [30]

    Denoising Diffusion Implicit Models

    Denoising diffusion implicit models , author=. arXiv preprint arXiv:2010.02502 , year=

    work page internal anchor Pith review Pith/arXiv arXiv 2010

  31. [31]

    Adam: A Method for Stochastic Optimization

    Adam: A method for stochastic optimization , author=. arXiv preprint arXiv:1412.6980 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  32. [32]

    2021 IEEE International Conference on Multimedia and Expo (ICME) , pages=

    Few-shot defect segmentation leveraging abundant defect-free training samples through normal background regularization and crop-and-paste operation , author=. 2021 IEEE International Conference on Multimedia and Expo (ICME) , pages=. 2021 , organization=

    work page 2021

  33. [33]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Draem-a discriminatively trained reconstruction embedding for surface anomaly detection , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

    work page

  34. [34]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Self-supervised predictive convolutional attentive block for anomaly detection , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    work page

  35. [35]

    IEEE Access , volume=

    Cfa: Coupled-hypersphere-based feature adaptation for target-oriented anomaly localization , author=. IEEE Access , volume=. 2022 , publisher=

    work page 2022

  36. [36]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Anomaly detection via reverse distillation from one-class embedding , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    work page

  37. [37]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Towards total recall in industrial anomaly detection , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    work page

  38. [38]

    The Twelfth International Conference on Learning Representations , year=

    Musc: Zero-shot industrial anomaly classification and segmentation with mutual scoring of the unlabeled images , author=. The Twelfth International Conference on Learning Representations , year=

    work page

  39. [39]

    arXiv preprint arXiv:2108.00462 , year=

    Explainable deep few-shot anomaly detection with deviation networks , author=. arXiv preprint arXiv:2108.00462 , year=

    work page arXiv

  40. [40]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Catching both gray and black swans: Open-set supervised anomaly detection , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    work page

  41. [41]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Prototypical residual networks for anomaly detection and localization , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    work page