Anomaly-Preference Image Generation

Dan Wang; Fuyun Wang; Hui Yan; Sujia Huang; Tong Zhang; Xin Liu; Xu Guo; Yuanzhi Wang; Zhen Cui

arxiv: 2605.02439 · v3 · 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 2026-07-01 00:33 UTC · model grok-4.3

classification 💻 cs.CV cs.LG

keywords anomaly generationpreference optimizationdiffusion modelsimage synthesisdenoising trajectorycapacity allocationanomaly detection

0 comments

The pith

Anomaly Preference Optimization generates realistic and diverse anomalous images by treating real anomalies as positive references in diffusion training.

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

The paper seeks to improve synthesis of anomalous images from scarce data, a key step for training models that handle rare events without overfitting or losing variety. It recasts the task as preference learning, where real anomalous samples supply positive signals that steer the model by measuring how its denoising path deviates from those references. A Time-Aware Capacity Allocation module then shifts emphasis from broad structure at high noise to fine detail at low noise, while inference uses hierarchical sampling to tune the realism-diversity balance. If the approach holds, generated anomalies become reliable enough to replace or augment scarce real data in downstream detection tasks.

Core claim

Anomaly Preference Optimization reformulates anomaly image synthesis as a preference learning problem. An implicit alignment step uses real anomalies as positive references to extract optimization signals directly from deviations along the denoising trajectory, removing the need for human annotations. A Time-Aware Capacity Allocation module dynamically assigns model capacity across diffusion timesteps, favoring structural diversity early and fidelity later. Hierarchical sampling at inference gives explicit control over coherence versus alignment. Experiments show the resulting method surpasses prior baselines on both realism and diversity metrics.

What carries the argument

Anomaly Preference Optimization, which derives training signals from denoising-trajectory deviations against real anomalies as positive references, together with Time-Aware Capacity Allocation across noise levels.

If this is right

Downstream anomaly detectors trained on the generated samples exhibit stronger generalization to unseen real anomalies.
Generation quality can be tuned at inference time by adjusting the hierarchical sampling schedule without retraining.
The method operates from limited seed data while avoiding the usual fidelity-diversity trade-off seen in prior diffusion approaches.
No human preference labels are required to obtain the alignment signal.

Where Pith is reading between the lines

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

The same preference-from-trajectory idea could transfer to other scarce-sample regimes such as medical lesion synthesis or rare-event video generation.
Combining the approach with existing anomaly detection pipelines might reduce the volume of real labeled anomalies needed for training.
If the capacity-allocation schedule proves robust, similar dynamic allocation could be tested in non-diffusion generative models.

Load-bearing premise

Real anomalies can be used directly as positive references to derive optimization signals from denoising trajectory deviations, and doing so improves both fidelity and diversity without human labels.

What would settle it

A side-by-side test in which anomaly detectors trained on samples from this method show no gain in generalization accuracy on held-out real anomaly datasets compared with detectors trained on samples from the strongest baseline generators.

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: 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 methods (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: Comparative analysis on the MVTec dataset demonstrates 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 established 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: Parameter sensitivity analysis of kmin. kmin = 4, where insufficient constraints (kmin < 4) impair structural fidelity despite preserving diversity, while excessive constraints (kmin > 4) reduce diversity without commensurate gains in realism. This behavior systematically confirms that our dynamic rank scheduling effectively regulates 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.

Desk Editor's Note private letter to a colleague

APO reframes anomaly synthesis as preference learning from real-anomaly trajectory deviations plus a time-aware module, but the abstract gives no loss, metrics, or ablations to show the signal is distinct or superior.

read the letter

The new piece here is treating anomaly generation as preference optimization where real anomalies act as positive references and optimization comes from denoising-trajectory deviations, plus the Time-Aware Capacity Allocation module that shifts capacity by noise level and the hierarchical sampler for coherence control.

That framing is reasonable for the fidelity-diversity problem in limited-data anomaly synthesis. Using existing anomalies to derive signals without fresh annotation is a practical angle, and splitting capacity across the diffusion timeline matches how diffusion actually works.

The soft spots are straightforward. The abstract states SOTA results in realism and diversity but supplies no equations for the preference term, no distance metric on trajectories, no baselines, and no numbers. Without those, it is impossible to tell whether the deviation signal produces a gradient different from ordinary denoising or simply restates reconstruction error. The stress-test concern lands: if the signal collapses, the preference reformulation is mostly nominal. The circularity risk is also live until the full derivation is checked.

This paper is aimed at people who need synthetic anomalies for training detectors in inspection or safety tasks. A reader working on domain-specific diffusion tweaks would find the mechanisms worth looking at even if the results require verification.

It deserves a serious referee. The problem is real, the approach is a coherent attempt, and the gaps are fixable with standard additions rather than fatal.

Referee Report

2 major / 1 minor

Summary. The manuscript proposes Anomaly Preference Optimization (APO), reformulating anomaly image generation as a preference learning problem. It uses real anomalies as positive references to derive optimization signals directly from denoising trajectory deviations without human annotation. It introduces a Time-Aware Capacity Allocation module that dynamically allocates model capacity across diffusion timesteps and a hierarchical sampling strategy to control the coherence-alignment trade-off at inference. The authors claim that APO achieves state-of-the-art performance in both realism and diversity over existing baselines.

Significance. If the preference signals derived from trajectory deviations can be shown to be both distinct from standard diffusion objectives and demonstrably superior, the approach would provide an annotation-free method for synthesizing diverse, realistic anomalies, which could improve data augmentation for anomaly detection and model robustness in computer vision.

major comments (2)

[Abstract] Abstract: the central claim that APO 'derives optimization signals directly from denoising trajectory deviations' using real anomalies as positive references provides no explicit loss term, distance metric on trajectories, or conditioning strategy. This formulation is load-bearing for the assertion that the resulting preference gradient is distinct from the standard denoising objective and yields the reported SOTA gains in realism and diversity.
[Abstract] Abstract and Method description: the Time-Aware Capacity Allocation module is described only at the level of 'dynamically distributes model capacity along the diffusion timeline' with no equations, allocation rule, or ablation showing its contribution to the fidelity-diversity trade-off; without this, the module's role in the overall performance claim cannot be evaluated.

minor comments (1)

[Abstract] Abstract: the sentence 'Extensive experiments demonstrate that significantly outperforms existing baselines' is grammatically incomplete (subject missing).

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on the clarity of our core technical claims. We agree that the abstract and method descriptions require greater specificity to allow proper evaluation of the contributions. We will revise the manuscript accordingly.

read point-by-point responses

Referee: [Abstract] Abstract: the central claim that APO 'derives optimization signals directly from denoising trajectory deviations' using real anomalies as positive references provides no explicit loss term, distance metric on trajectories, or conditioning strategy. This formulation is load-bearing for the assertion that the resulting preference gradient is distinct from the standard denoising objective and yields the reported SOTA gains in realism and diversity.

Authors: We agree that the abstract would benefit from greater specificity on this central claim. In the revised version, we will include an explicit description of the loss term used to capture the denoising trajectory deviations, the distance metric applied to the trajectories, and the conditioning strategy that uses real anomalies as positive references. This will better demonstrate how the preference gradient is derived and its distinction from the standard diffusion objective. revision: yes
Referee: [Abstract] Abstract and Method description: the Time-Aware Capacity Allocation module is described only at the level of 'dynamically distributes model capacity along the diffusion timeline' with no equations, allocation rule, or ablation showing its contribution to the fidelity-diversity trade-off; without this, the module's role in the overall performance claim cannot be evaluated.

Authors: We concur with the referee that the current description of the Time-Aware Capacity Allocation module is insufficiently detailed. We will revise the method section to provide the full equations for the capacity allocation rule, which dynamically adjusts based on the diffusion timestep to balance diversity and fidelity. We will also include an ablation study quantifying its contribution to the reported performance gains in realism and diversity. revision: yes

Circularity Check

0 steps flagged

No circularity detected from available text

full rationale

The abstract and description introduce APO as reformulating anomaly generation via preference signals from denoising deviations using real anomalies as references, plus a Time-Aware Capacity Allocation module and hierarchical sampling. No equations, self-citations, or derivations are quoted that reduce any prediction or central result to its own inputs by construction. The approach is presented as adding independent mechanisms (implicit alignment without annotation, dynamic capacity allocation) whose validity would be tested against external baselines rather than forced by definition. This matches the default expectation of a self-contained paper.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 3 invented entities

Review is abstract-only so the ledger lists only concepts explicitly named in the abstract; no free parameters or external benchmarks are described.

axioms (1)

domain assumption Real anomalies can serve as positive references for deriving optimization signals from denoising trajectory deviations without human annotation.
This is stated as central to the implicit preference alignment mechanism.

invented entities (3)

Anomaly Preference Optimization no independent evidence
purpose: Reformulate anomaly generation as a preference learning problem
New paradigm introduced to address fidelity-diversity trade-off.
Time-Aware Capacity Allocation module no independent evidence
purpose: Dynamically distribute model capacity along the diffusion timeline prioritizing diversity at high noise and fidelity at low noise
Proposed module to address distribution misalignment and overfitting.
hierarchical sampling strategy no independent evidence
purpose: Modulate the coherence-alignment trade-off during inference
Inference-time control mechanism.

pith-pipeline@v0.9.1-grok · 5695 in / 1389 out tokens · 52337 ms · 2026-07-01T00:33:35.680495+00:00 · methodology

Review history (3 revisions) →

discussion (0)

Forward citations

Cited by 3 Pith papers

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

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.
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.
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

9 extracted references · 8 canonical work pages · cited by 1 Pith paper · 5 internal anchors

[1]

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

Bai, Y ., Jones, A., Ndousse, K., Askell, A., Chen, A., Das- Sarma, N., Drain, D., Fort, S., Ganguli, D., Henighan, T., et al. Training a helpful and harmless assistant with rein- forcement learning from human feedback.arXiv preprint arXiv:2204.05862, 2022a. Bai, Y ., Kadavath, S., Kundu, S., Askell, A., Kernion, J., Jones, A., Chen, A., Goldie, A., Mirho...

work page internal anchor Pith review Pith/arXiv arXiv
[2]

Seas: few-shot industrial anomaly image generation with separation and sharing fine-tuning.arXiv preprint arXiv:2410.14987,

Dai, Z., Zeng, S., Liu, H., Li, X., Xue, F., and Zhou, Y . Seas: few-shot industrial anomaly image generation with separation and sharing fine-tuning.arXiv preprint arXiv:2410.14987,

work page arXiv
[3]

ORPO: Monolithic Preference Optimization without Reference Model

Hong, J., Lee, N., and Thorne, J. Orpo: Monolithic pref- erence optimization without reference model.arXiv preprint arXiv:2403.07691,

work page internal anchor Pith review Pith/arXiv arXiv
[4]

Kingma, D. P. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980,

work page internal anchor Pith review Pith/arXiv arXiv
[5]

Lee, K., Liu, H., Ryu, M., Watkins, O., Du, Y ., Boutilier, C., Abbeel, P., Ghavamzadeh, M., and Gu, S. S. Aligning text- to-image models using human feedback.arXiv preprint arXiv:2302.12192,

work page internal anchor Pith review Pith/arXiv arXiv
[6]

Pang, G., Ding, C., Shen, C., and Hengel, A. v. d. Ex- plainable deep few-shot anomaly detection with deviation networks.arXiv preprint arXiv:2108.00462,

work page arXiv
[7]

Denoising Diffusion Implicit Models

Song, J., Meng, C., and Ermon, S. Denoising diffusion implicit models.arXiv preprint arXiv:2010.02502,

work page internal anchor Pith review Pith/arXiv arXiv 2010
[8]

Zhao, Y ., Joshi, R., Liu, T., Khalman, M., Saleh, M., and Liu, P. J. Slic-hf: Sequence likelihood calibration with human feedback.arXiv preprint arXiv:2305.10425,

work page arXiv
[9]

10 Anomaly-Preference Image Generation A. Ablation Study on Regularization Coefficientβ The regularization coefficient β governs the fundamental trade-off in our constrained optimization framework, determining the extent to which the model adapts to few-shot anomalies while preventing overfitting. As shown in Tab. 6,β exhibits a pronounced effect on gener...

2000

[1] [1]

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

Bai, Y ., Jones, A., Ndousse, K., Askell, A., Chen, A., Das- Sarma, N., Drain, D., Fort, S., Ganguli, D., Henighan, T., et al. Training a helpful and harmless assistant with rein- forcement learning from human feedback.arXiv preprint arXiv:2204.05862, 2022a. Bai, Y ., Kadavath, S., Kundu, S., Askell, A., Kernion, J., Jones, A., Chen, A., Goldie, A., Mirho...

work page internal anchor Pith review Pith/arXiv arXiv

[2] [2]

Seas: few-shot industrial anomaly image generation with separation and sharing fine-tuning.arXiv preprint arXiv:2410.14987,

Dai, Z., Zeng, S., Liu, H., Li, X., Xue, F., and Zhou, Y . Seas: few-shot industrial anomaly image generation with separation and sharing fine-tuning.arXiv preprint arXiv:2410.14987,

work page arXiv

[3] [3]

ORPO: Monolithic Preference Optimization without Reference Model

Hong, J., Lee, N., and Thorne, J. Orpo: Monolithic pref- erence optimization without reference model.arXiv preprint arXiv:2403.07691,

work page internal anchor Pith review Pith/arXiv arXiv

[4] [4]

Kingma, D. P. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980,

work page internal anchor Pith review Pith/arXiv arXiv

[5] [5]

Lee, K., Liu, H., Ryu, M., Watkins, O., Du, Y ., Boutilier, C., Abbeel, P., Ghavamzadeh, M., and Gu, S. S. Aligning text- to-image models using human feedback.arXiv preprint arXiv:2302.12192,

work page internal anchor Pith review Pith/arXiv arXiv

[6] [6]

Pang, G., Ding, C., Shen, C., and Hengel, A. v. d. Ex- plainable deep few-shot anomaly detection with deviation networks.arXiv preprint arXiv:2108.00462,

work page arXiv

[7] [7]

Denoising Diffusion Implicit Models

Song, J., Meng, C., and Ermon, S. Denoising diffusion implicit models.arXiv preprint arXiv:2010.02502,

work page internal anchor Pith review Pith/arXiv arXiv 2010

[8] [8]

Zhao, Y ., Joshi, R., Liu, T., Khalman, M., Saleh, M., and Liu, P. J. Slic-hf: Sequence likelihood calibration with human feedback.arXiv preprint arXiv:2305.10425,

work page arXiv

[9] [9]

10 Anomaly-Preference Image Generation A. Ablation Study on Regularization Coefficientβ The regularization coefficient β governs the fundamental trade-off in our constrained optimization framework, determining the extent to which the model adapts to few-shot anomalies while preventing overfitting. As shown in Tab. 6,β exhibits a pronounced effect on gener...

2000