acceptodds
Withdrawn from ICLR 2027

SevAnoDiff: Severity-Aware Diffusion Adaptation for Zero-Shot Industrial Anomaly Generation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Industrial anomaly detection relies on diverse anomaly images with pixel-level annotations, yet real industrial anomalies are rare and costly to collect. Existing anomaly generation methods mainly synthesize isolated samples while not modeling how the anomaly changes across various severity stages. First, we construct SevADSet, an industrial anomaly dataset with four-stage ordinal trajectories from normal images to increasingly visible anomaly stages. Second, based on SevADSet, we propose SevAnoDiff, a diffusion-based method for zero-shot severity-aware anomaly generation. An Appearance Prior Adapter first learns transferable anomaly appearance from observed anomalies. A Severity Progression Module then models adjacent-stage changes using stage tokens, mask trajectories, and previous-stage visual references. Without using anomalous images or masks from the target domain, SevAnoDiff generates realistic anomaly image-mask pairs with clear severity progression on target domain normal images. Extensive experiments on MVTec-AD and VisA demonstrate the effectiveness of SevAnoDiff in anomaly generation and its utility for anomaly detection and localization.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Industrial anomaly detection, anomaly generation, diffusion models, image generation, industrial inspection

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_RNAnOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Motivation: Most industrial anomaly generation methods synthesize isolated defects without explicitly modeling how the visual manifestation of the same anomaly progresses from subtle to severe. The paper studies ordered visual severity in a zero-shot target-domain setting.

Contribution and validation: The authors construct SevADSet from AnomVerse references, retaining 2,476 four-stage image-mask trajectories (normal, weak, reference, severe). SevAnoDiff adapts Stable Diffusion v1.5 Inpainting in two stages: an Appearance Prior Adapter (APA) trains cross-attention LoRA for anomaly appearance, and a Severity Progression Module (SPM) combines self-attention LoRA, learned stage tokens, mask-trajectory encoding, and previous-stage visual references. Evaluation on MVTec-AD and VisA reports generation quality/diversity, VLM-based severity progression accuracy (SPA/PSA), and downstream U-Net detection/localization results; the reported SPA values are 89.11% and 88.63%, respectively.

Critical strengths
  1. The paper targets a meaningful and underexplored problem: producing coherent stages of the same industrial anomaly rather than independent defective images. Its careful distinction between ordinal visual severity and actual physical damage makes the research question reasonably well specified. SevADSet may also serve as a useful starting point for studying controlled anomaly progression.

  2. The empirical study includes both generation and downstream detection/localization, with clear gains from the complete two-stage pipeline. On MVTec-AD, SPA rises from 6.00% for the diffusion baseline to 38.67% with APA and 89.11% with APA+SPM (Table 3). The condition ablations and qualitative examples provide credible initial evidence that the combined model improves ordinal consistency.

Critical weaknesses
  1. The necessity of the proposed severity progression mechanism is not sufficiently established by controlled comparisons. At inference, the generator is provided with an already ordered three-stage mask trajectory sampled from a region bank, while successive stages are generated recursively using previous outputs (Sec. 3.2.3 and Appendix A.3). The model also uses stage tokens and a fixed denoising strength of 0.85. Although these conditions jointly achieve high SPA, the presented ablations do not disentangle learned appearance-severity control from predefined geometry and ordinary conditional editing. For example, removing the stage token still leaves 76.00% SPA on MVTec-AD and 70.00% on VisA (Table 4), suggesting that other conditions already convey substantial stage information. Denoising strength is not itself a severity measure, but a systematic sweep is needed as a competing control. Crucially, the paper lacks matched baselines combining severity-specific prompts, strength variation, the same masks, and recursive inpainting; existing methods such as Anomagic are not adapted to the new trajectory task. Without these experiments, the incremental technical advance over standard diffusion editing remains uncertain.

  2. The proposed architectural division and the new evaluation protocol receive insufficient validation. APA updates cross-attention LoRA, whereas SPM updates self-attention LoRA (Appendix A.3.1), but no attention-placement or parameter-matched ablation explains why this separation is necessary. The learned stage tokens are trained through reconstruction/denoising objectives, without an explicit ordinal constraint or counterfactual test demonstrating that token identity alone controls severity when masks and references are held fixed. Moreover, SevADSet constructs three of four stages using multimodal image editing and filters them through VLM and human review; the paper reports no clear inter-rater agreement or independent validation of how well its VLM-based SPA tracks human severity judgments across anomaly types. The main comparison table omits SPA for prior methods because they generate individual samples, leaving few directly comparable trajectory-generation baselines. These omissions weaken both the mechanistic interpretation and the strength of the claimed benchmark-level results.

Questions for authors
  1. Can the authors isolate the new severity-control capability from standard diffusion controls and the supplied mask trajectories? Please compare against matched SD Inpainting and reasonably adapted prior zero-shot generators (especially Anomagic) using the same input images, anomaly types, stage masks, recursive schedule, and generation budget, while varying denoising strength and severity prompts. A controlled test with fixed masks/references and swapped stage tokens, together with a separate mask-trajectory-only test, would clarify which factor actually produces the severity ordering.

  2. What evidence supports the specific APA-cross-attention / SPM-self-attention division and the claimed learned severity semantics? Could the authors report parameter-matched attention-placement ablations (cross-only, self-only, joint, or swapped), plus fixed-input stage-token interventions and unseen-mask-trajectory tests? For the SevADSet/SPA evaluation, please also specify per-category evaluation sample counts, independent human-rating agreement, and the relationship between perceived severity, defect appearance, and mask area. These clarifications could materially affect confidence in the central conclusion.

Reviewer_P4QqOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper investigates severity-aware industrial anomaly generation, extending conventional synthesis of isolated anomaly samples to multi-stage severity trajectories within the same image context and anomaly type. The authors construct SevADSet, where real anomalies serve as the reference stage (S2S_2), while weaker and more severe variants are synthesized using the open-source Qwen-Image-Edit model. Built upon Stable Diffusion Inpainting, SevAnoDiff employs an Appearance Prior Adapter to learn anomaly appearance, followed by a Severity Progression Module that models adjacent-stage transitions using stage tokens, multi-stage mask encoding, and visual references from the preceding stage.

Critical strengths

The paper goes beyond simply generating anomalies by exploring how to control their progression from mild to severe while preserving object consistency and anomaly identity. Constructing severity trajectories with real anomalies as intermediate reference stages provides a more grounded visual basis for severity modeling than defining severity levels without real anomaly references.

Critical weaknesses
  1. Lack of comparison with Qwen-Image-Edit. The authors employ Qwen-Image-Edit to synthesize high-quality severity progression data for training the Severity Progression Module. However, if Qwen-Image-Edit is already capable of generating such high-quality anomalies, why not directly use it as a zero-shot anomaly generation model to produce synthetic training data for downstream anomaly detection? A direct comparison with Qwen-Image-Edit is necessary to justify the effectiveness and necessity of the proposed framework.
  2. Dependence on the teacher model. Since the training data are synthesized using Qwen-Image-Edit and subsequently used to fine-tune Stable Diffusion v1.5, the proposed approach can essentially be viewed as a form of knowledge distillation from a stronger image editing model. Consequently, the generation quality may be inherently constrained by the capabilities and limitations of Qwen-Image-Edit. The authors should clarify what advantages their framework provides beyond transferring the teacher model's generation capabilities. Furthermore, have the authors explored using a newer and more powerful model, such as Qwen-Image-2.1[1], for training data synthesis? If so, does it lead to further improvements?
  3. Limited competitiveness against AnoStyler. The downstream anomaly detection results on MVTec-AD and VisA do not demonstrate a convincing advantage over AnoStyler[2]. More importantly, AnoStyler achieves comparable performance without requiring such an elaborate dataset construction pipeline.
  4. I suggest including real anomaly images for direct visual comparison. Furthermore, the generated anomalies appear substantially different from real anomalies. The five generic anomaly types mentioned in Sec. 4.1 used for zero-shot generation also differ considerably from the actual defect categories in MVTec-AD and VisA. In particular, these datasets contain logical and structural anomalies, such as cable swaps, which are not adequately covered by the proposed anomaly types. These discrepancies raise significant concerns about the realism and diversity of the synthesized anomalies and their ability to represent real industrial defects.

[1] https://huggingface.co/Qwen/Qwen-Image-2.1

[2] So, Yulim, and Seokho Kang. "Anostyler: Text-driven localized anomaly generation via lightweight style transfer." Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 40. No. 18. 2026.

Questions for authors
  1. Evaluation of distributional similarity. I recommend adopting the Kernel Inception Distance (KID) or FID metric to evaluate the distributional discrepancy between generated and real anomalies, providing further evidence of the realism and distributional fidelity of the synthesized samples.
  2. Extension to few-shot anomaly generation. How would the proposed method perform under the same few-shot generation protocol as AnomalyDiffusion? The framework appears readily adaptable to the few-shot setting. Under a unified evaluation protocol, would it remain competitive with existing few-shot methods such as SeaS[3] and DualAnoDiff[4], as well as training-free approaches such as O2MAG[5] and TF-IDG[6]?
  3. I am particularly concerned about the use of Qwen-Image-Edit to synthesize training data for the Severity Progression Module, followed by fine-tuning Stable Diffusion v1.5. This process appears conceptually similar to knowledge distillation from Qwen-Image-Edit. Could the authors provide more detailed examples of the training data synthesized by Qwen-Image-Edit across different severity levels? Furthermore, do Gemini's severity assessments align with human judgments and preferences?

[3] Dai, Zhewei, et al. "Seas: few-shot industrial anomaly image generation with separation and sharing fine-tuning." 2025 IEEE/CVF International Conference on Computer Vision (ICCV). IEEE, 2025.

[4] Jin, Ying, et al. "Dual-interrelated diffusion model for few-shot anomaly image generation." 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2025.

[5] Rao, Haoxiang, et al. "One-to-More: High-fidelity training-free anomaly generation with attention control." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2026.

[6] Xu, Ruyi, et al. "Training-free industrial defect generation with diffusion models." 2025 IEEE/CVF International Conference on Computer Vision (ICCV). IEEE, 2025.

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

This paper proposes SevAnoDiff for industrial anomaly generation with ordered visual severity, defined relative to the same defect type and image context. It introduces SevADSet, containing 2,476 four-stage trajectories anchored by real anomalous images, with normal, weaker, and stronger variants constructed through image editing and filtering. SevAnoDiff then adapts Stable Diffusion Inpainting in two stages. An Appearance Prior Adapter first learns defect appearance, and a Severity Progression Module learns adjacent-stage changes using stage tokens, mask-trajectory conditioning, and visual references from the preceding stage. At inference, source-domain mask trajectories and target normal images guide recursive generation, followed by refinement of the visible anomaly masks. The method does not use anomalous images or masks from the target MVTec AD and VisA datasets. Experiments assess generation quality, severity ordering through a vision-language-model evaluator, and downstream detection and localization using the synthesized image–mask pairs. The results show substantial gains in ordering consistency over controlled baselines and competitive downstream performance. The main unresolved questions concern the validity of the severity metric and how much ordered progression itself contributes to detector improvements.

Critical strengths
  1. The task formulation extends independent anomaly synthesis to consistent variations of a defect within a fixed image context. SevADSet provides a concrete resource for this setting, retaining a real anomaly and its original mask as an anchor rather than synthesizing every stage. Dataset construction includes automatic filtering and independent checks by two experienced reviewers. The paper also explicitly distinguishes visual severity from physical severity, appropriately delimiting the task.

  2. The architecture separates learning anomaly appearance from learning adjacent-stage changes, with complementary spatial and visual conditions. The ablations provide useful support: Section 4.4 specifies that the Base and APA comparisons use the same masks and recursive inference as the full model. SPA increases from 38.67% to 89.11% on MVTec-AD and from 36.67% to 88.63% on VisA when adding SPM (Table 3), supporting a contribution beyond prescribing progressively changing regions.

  3. The evaluation includes real-image downstream tests rather than relying entirely on generated-image metrics. Table 2 reports the strongest image AUROC and pixel AP among the compared zero-shot methods on both datasets, with additional pixel-level gains on MVTec-AD. The downstream protocol specifies a 500-sample training budget, allocated nearly equally across the three stages. Per-category results and extensive qualitative examples further help assess where the method performs well and where localization remains difficult.

Critical weaknesses
  1. The severity metric needs stronger validation and more complete reporting. The evaluation protocol specifies anonymous presentation, random permutations, and majority voting, but does not specify the number of evaluated trajectories per category, sampling procedure, full evaluator prompt, or treatment of indistinguishable stages. Generating 1,000 trajectories per category in Section 4.1 does not establish that all enter SPA evaluation. Human verification during dataset construction is valuable, but does not establish agreement between Gemini rankings and independent human judgments of generated target-domain trajectories. A blinded agreement study, per-category evaluation counts, and confidence intervals would clarify what the reported approximately 89% SPA measures and how reliably it captures visual ordering.

  2. The downstream experiments do not fully isolate the benefit of ordered progression from improved sample quality or additional adaptation. Although the same-mask recursive comparisons in Table 3 are meaningful, adding SPM also changes trainable capacity, conditioning, and training targets. Meanwhile, the downstream detector receives individual images and binary masks, without trajectory order. Equal-budget comparisons between reference-stage-only samples, a mixture of stages, and independently generated stages with matched model capacity would help identify which aspect produces the gain. An ablation disrupting stage correspondence during generator training could further test the role of ordering. Performance on subtle real anomalies would also connect the results more directly to the stated motivation.

  3. The mask-refinement description is ambiguous in a way that could matter for localization supervision. Section 3.2.3 describes refinement within a dilated cumulative region, and Algorithm 1 passes all masks through the current stage and the object region to the refinement operation. However, the detailed refinement procedure restricts the Lab-difference-based mask extraction to the support of the current editable region. By Equation (6), that region at stage 3 is the union of M2 and M3, whereas the cumulative region also includes M1. These supports need not coincide if masks are not nested, so an earlier visible anomaly outside the current editable region could be omitted under the latter interpretation. This may be only a presentation issue. Clarifying the implemented support, nesting assumptions, threshold rules, and quantitative mask agreement against independent annotations would resolve the concern.

Questions for authors

None

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

Related papers

Discussion (0)

Sign in to comment.