acceptodds
Withdrawn from ICLR 2027

DecoGen: Zero-Shot Cross-Style Industrial Anomaly Generation via Explicit Physical Semantic Constraint

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

In recent years, industrial anomaly image generation has shown immense potential in alleviating the scarcity of real defect data. However, existing few-shot generation methods typically rely on implicit and coarse-grained visual guidance lacking real-world physical grounding, which inherently entangles anomalous features with domain-specific background textures. This representation entanglement leads to severe texture contamination and semantic misalignment, making it exceedingly difficult to transfer known defects to novel product styles that lack real anomaly images. To address these issues, we propose DecoGen, a novel zero-shot cross-style anomaly transfer framework driven by explicit physical semantic constraints. Specifically, we construct a structured anomaly prompt bank to explicitly decompose defects into multi-dimensional, domain-invariant physical properties for flexible semantic combinations. Furthermore, a physical attribute aligner is designed to enforce rigorous disentangled alignment in the feature space, thoroughly decoupling the pure physical anomaly appearance from both the source background texture and its mask-derived spatial support. Experiments show that DecoGen yields high-fidelity and diverse cross-style anomalies. Crucially, our synthesized data significantly improves the performance of downstream anomaly detectors on unseen product styles, increasing the average image-level and pixel-level AUROC by 3.99% and 6.32%, respectively.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Zero-Shot Anomaly Generation, Defect Detection, Cross-Style Transfer

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_mcEYOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes DecoGen for cross-style industrial anomaly generation when only source-style defect examples and normal images from target styles are available. The method constructs a structured anomaly prompt bank through defect-region clustering and attribute descriptions. It combines appearance and spatial tokens, attribute- and prototype-level alignment losses, prototype aggregation at inference, and mask-guided diffusion sampling to transfer defects to unseen styles. The paper evaluates generation quality and downstream anomaly detection on ZJU Leaper and MVTec AD, together with component ablations.

Critical strengths

1.The paper tackles a practical problem in industrial anomaly detection: generating defects for new product styles when no real defect samples are available. DecoGen combines structured defect attributes, prototype alignment, and mask-guided generation to transfer defects across styles. I find this design interesting because it allows the generated defects to be controlled in terms of their characteristics and locations.

2.The experiments on real target data are convincing. Table 3 evaluates the benefit of synthetic augmentation across eight unseen styles and three detectors. Table 12 further reports positive average gains in both image-level and pixel-level AUROC for every target style. Although some individual results decline, the overall improvements across different styles and detectors support the practical value of the method.

Critical weaknesses

1.The paper lacks a direct comparison with existing generators under the same zero-shot cross-style setting, and the downstream adaptation setup needs further clarification. Table 3 only compares detector performance with and without DecoGen augmentation. It is also unclear whether Tables 1 and 2 use the same source-target splits, data access conditions, and number of generated samples. The current results show that DecoGen's synthetic data is useful, but they do not clearly establish its advantage over existing generators for cross-style defect generation.

2.The claimed disentanglement of physical defect mechanisms is not directly demonstrated. The attribute and prototype prediction losses encourage semantic consistency, but this does not necessarily mean that defect attributes are independent of source textures or mask geometry. The ablations in Table 4 mainly examine whether the differences between generated and normal images align with the given defect masks. They do not directly evaluate the accuracy of physical attributes, possible contamination from source textures, or the effect on detecting real target anomalies. It therefore remains unclear how much of the reported improvement comes from the proposed disentanglement mechanism.

Questions for authors

Q1. Do all the generators in Tables 1 and 2 follow the same target-anomaly-free setting as Table 3? Could the authors clarify the source-target splits, what images and masks each method can access, where the masks come from, and how many synthetic samples are generated? It would also help to explain whether the normal reference images overlap with the final test set.

Q2. How exactly are the generated anomalies and masks used in PatchCore and WinCLIP? PatchCore normally relies on a memory bank of normal features, while WinCLIP uses text matching or normal reference features. Were their training procedures, memory banks, or anomaly scoring methods modified? Also, do Base and +Ours use the same target normal images?

Q3. Could the authors compare other generators and the ablated variants in Table 4 on real target test data under the same zero-shot setting and synthesis budget? Is there also an independent evaluation of defect attribute accuracy and source-texture contamination, without relying on the attribute aligner used during training or the image-residual scoring method?

Reviewer_HVbmOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes DecoGen, a diffusion-based framework for generating anomalies on unseen industrial product styles using source-domain anomaly examples but no real target-domain anomalies. The method clusters masked source anomaly features into prototypes, represents each prototype through a structured bank of physical attributes, and learns appearance and mask-derived spatial tokens around a frozen latent diffusion backbone. An auxiliary attribute aligner is used to encourage attribute- and prototype-level consistency. At inference, source-style appearance tokens are aggregated or combined according to prototype similarity and applied to normal target images through mask-guided DDIM sampling.

Critical strengths
  1. The paper studies anomaly generation when a new product style has normal images but no observed anomalies. This cold-start setting is practically relevant and more challenging than conventional target-domain few-shot anomaly generation.

  2. DecoGen is evaluated through both generation metrics and downstream detection across multiple detector families and unseen textile styles. The reported improvements are generally consistent across styles and architectures, suggesting that the generated samples may provide useful supervision rather than benefiting only one detector.

Critical weaknesses
  1. The claimed physical-semantic alignment is not independently validated. Both the attribute aligner and prototype projector are learned jointly using labels produced by the same prompt-bank construction process. Consequently, the auxiliary task may learn prototype-specific artifacts or other shortcuts without ensuring that generated images satisfy the intended physical attributes. The component ablation primarily uses image–normal residual consistency inside a known mask, which measures localization but does not independently verify physical correctness or realism. The four manually selected semantic examples are insufficient to resolve this circularity.

  2. Several central implementation and evaluation details are internally inconsistent or missing. Eq. (8) uses the condition as the query and the noisy latent as key/value, which appears incompatible with the stated frozen pretrained LDM architecture. The main text refers to a CLIP text encoder, whereas Appendix A.1 specifies a 32-layer BERT conditioner.

Questions for authors

From which distribution are the masks used for pattern8–pattern15 sampled after random-mask fallback is disabled? Please also specify the real target test-set sizes/splits and the exact loss or adaptation procedure by which synthetic anomaly–mask pairs are used with each downstream detector. These details are necessary to exclude target-anomaly leakage and assess the reported gains.

Reviewer_PM6EOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes DecoGen, a framework for transferring industrial defect appearances to new product styles that have normal reference images but no real target anomalies. It addresses the difficulty of separating a defect from the source product texture so that the defect can be synthesized on a different background. The method constructs a structured prompt bank describing defect attributes, learns separate appearance and mask embeddings, and adds attribute and prototype alignment during diffusion personalization. At inference, source prototype embeddings are aggregated or combined according to requested attributes, while masked latent composition helps preserve the target background. The evaluation covers ZJU Leaper and MVTec AD, including an explicit transfer setting with seven source textile styles and eight target styles. Results report improved generation diversity and downstream anomaly detection, with average cross-style image-level and pixel-level AUROC gains of 3.99 and 6.32 percentage points over unaugmented detectors. The framework provides useful evidence for transferring known defects to new styles, although the degree of physical disentanglement and the fairness of the central cross-style comparisons require stronger support.

Critical strengths
  1. The problem formulation addresses a concrete industrial cold-start scenario, and the method combines interpretable conditioning with a practical generation pipeline. The shared attribute schema makes the intended defects more inspectable than an unspecified learned token alone. Explicit rules for aggregating observed prototypes and composing attribute requests also help clarify how transfer operates without optimizing on target anomaly images.

  2. The reported downstream improvements extend across multiple detectors and target styles. In Table 3, all three detectors improve on average, and the per-style breakdown shows positive cross-detector mean gains for all eight target styles. Table 2 provides complementary results on two datasets: DecoGen achieves the strongest augmentation average on 13 of 14 metrics, with MVTec pixel AP explicitly remaining below DualAnoDiff. These results provide useful evidence beyond selected visual examples.

  3. The paper provides valuable methodological qualifications. It distinguishes generated image–mask consistency from real-image detection, explains that 200 outputs per target style reuse five normal backgrounds, and discloses single-seed uncertainty and individual performance regressions. The triplet-integrity checks and prototype metadata are also useful reproducibility measures. These details help delimit the evidence and avoid interpreting diversity scores or selected examples as guarantees of realism.

Critical weaknesses
  1. The comparison supporting the central cross-style claim needs a clearer connection to the broader generator benchmarks. Tables 1 and 2 include competing generators, but Table 3 compares only unaugmented detectors against DecoGen augmentation. The reported comparison protocol standardizes output resolution and metrics without establishing whether all generators share the same seven-to-eight style split, anomaly budget, masks, normal references, and sample count. Thus, Table 3 establishes utility relative to no augmentation, but does not directly establish superiority over alternative generators under that exact setting. The target-mask provenance also remains unclear despite the stated disabling of random-mask fallback, and the detector-specific adaptation procedures are underspecified. Matched generator comparisons, mask-source information, and explicit training/adaptation configurations would make the zero-target-anomaly evaluation substantially easier to assess.

  2. The evidence for physical disentanglement is weaker than the corresponding claims. Separate appearance and spatial inputs, together with attribute classification losses, do not themselves establish invariance to source texture or independence from mask geometry. Table 4 offers indirect evidence, but its localization scores explicitly measure residual-based consistency between generated images and paired normal references, rather than detector performance on real anomalies. Larger or more conspicuous changes could improve these proxies without establishing greater physical fidelity. Moreover, the stated prompt-bank removal also removes prototype identities, attribute supervision, and routing, preventing a clean attribution to structured semantics. Real-test-set component ablations and controlled attribute/background interventions, with matched token capacity and routing, would more directly support the proposed explanation.

  3. Several implementation descriptions require reconciliation. Sections 3.4 and 3.7 specify a CLIP text encoder, whereas the implementation settings report a frozen BERT conditioner; the actual conditioning path should be made consistent. More consequentially, the stated complete objective excludes an additional feedback loss, yet Section 4.4 and the sensitivity study use a generated-feedback weight, λgen, without defining which implementation operation it controls. A common one-factor reference configuration is also specified, but the λgen=0.10 and learning-rate=0.005 settings report different image-level AUROCs, 92.72 and 93.72, despite apparently corresponding to that reference. These may reflect documentation differences or nondeterminism rather than implementation errors, but resolved configurations and an explanation of the repeated reference would be needed to interpret the sensitivity results reliably.

  4. The robustness of smaller improvements remains uncertain. The paper explicitly acknowledges that controlled ablations use one seed, while the five repeated backgrounds per target style limit the independent background variation represented by 200 outputs. This matters especially for prototype-adaptation localization differences of approximately 0.01–0.07 points in Table 4 and small MVTec generation-metric margins in Table 1. Repeated source-example and normal-reference sampling, uncertainty estimates, and diversity comparisons controlling for background, mask area, and requested prototype would help distinguish consistent gains from sampling variation. The existing cross-style averages are encouraging, but do not resolve these finer-grained comparisons.

Questions for authors

See above

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

Related papers

Discussion (0)

Sign in to comment.