acceptodds
Under review as a conference paper at ICLR 2027

MorphControl: Nuclear Fuel Rod Defect Image Generation via Morphology Prior Adaptation and Quality-Feedback Prior Regulation

Abstract

Surface defect datasets for nuclear fuel rods are typically scarce and class-imbalanced, while defects within the same category vary considerably in shape, scale, and local texture. These characteristics make it difficult for generative models to simultaneously preserve class consistency, intra-class diversity, and the structural fidelity of industrial backgrounds. Existing methods often rely on instance-level conditioning and therefore do not explicitly capture morphological patterns shared across samples of the same class. Moreover, fixed conditioning strengths cannot account for differences among defect categories, backgrounds, and stochastic sampling states, potentially leading to weak defect expression, excessive enhancement, or local artifacts. To address these limitations, we propose MorphControl, an industrial defect generation framework that combines defect-morphology-prior-driven multi-condition adaptation with quality-feedback prior regulation. During training, MorphControl attenuates defect regions in reference images to construct defect-suppressed background conditions while keeping the pretrained diffusion backbone frozen. It jointly learns class-level defect morphology priors from samples of the same category. Combining these priors with instance-specific industrial backgrounds, geometric layouts, and textual conditions enables class-shared morphological knowledge to be recombined under diverse generation conditions. A region-weighted diffusion objective and a background–morphology decorrelation constraint further strengthen fine-grained defect learning and reduce information coupling between conditions. During inference, MorphControl adaptively regulates the injection strength of the morphology prior according to regional feature consistency and artifact risk, and then selects the highest-quality candidate that satisfies the quality constraints. Experiments show that MorphControl outperforms existing methods in generation fidelity and intra-class diversity, achieving lower KID and higher IC-LPIPS. Using the generated samples for data augmentation further improves downstream defect detection in both mAP50 and mAP50:95, with larger gains for underrepresented defect categories. These results show that class-level defect morphology priors and inference-time quality-feedback regulation improve both synthesis quality and the utility of generated data for downstream defect detection.

Primary area:
generative models
Keywords:
Industrial Defect Generation, Nuclear Fuel Rods, Diffusion Models, Class-level Defect Morphology Priors, Quality-feedback Prior regulation

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_SbZGOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes MorphControl, a diffusion method to generate defect images for nuclear fuel rods with limited training data. It separates background textures from class-level defect shapes during training, and uses a feedback loop during inference to adjust defect strength. Experiments show better image quality than baselines and a 4.14 mAP50 boost on YOLOv8 defect detection.

Critical strengths

The ideas are reasonable and intuitive. Generated images clearly improve YOLOv8 detection accuracy, especially on rare scratch defects.

Critical weaknesses
  1. It is only tested on one private dataset and lacks results on public benchmarks like MVTec AD. Is this work specific to the image domain of nuclear fuel rod?

  2. I have a serious concern over novelty. Is this work a direct modification of "Defect-GAN: High-Fidelity Defect Synthesis for Automated Defect Inspection, WACV 2021"? It seems to me that the only difference is that this work uses diffusion models and Defect-GAN uses GAN.

Questions for authors

Please address the weaknesses listed above. My major concern is the novelty.

Reviewer_nAjiOpenReview ↗
Rating
1(−1)
Confidence
4(+1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a diffusion-based framework for generating surface defects on nuclear fuel rods called MorphControl. This framework utilizes class-level morphological priors and adaptive quality feedback, combined with background, geometric, and textual conditions, to improve the realism and diversity of defects. Experiments compare the quality of generated defects with existing methods and evaluate the effectiveness of synthetic data in downstream defect detection.

Critical strengths

This paper aims to address a practical challenge in industrial defect generation: the limited and imbalanced training data negatively impacts detection performance. Its main advantage lies in combining class-level morphological priors with adaptive quality feedback, thereby improving the realism and diversity of defects. Evaluation results also demonstrate the potential benefits of synthetic data for downstream defect detection.

Critical weaknesses

The main weakness of this approach lies in its limited technological innovation, as the framework primarily incorporates existing diffusion regulation and feedback mechanisms. The experimental validation is also insufficient, with incomplete ablation studies, and evaluation restricted to a single proprietary dataset. The source code is not publicly available, making it difficult to independently verify the reported results. These limitations make it difficult to assess the individual contributions and generalizability of the proposed method.

Questions for authors

The authors should provide a more complete ablation study to clarify the individual contributions of the morphology priors, region-weighted loss, and decorrelation constraint. The missing dataset specifications, baseline configurations, and evaluation details in the appendix should be addressed. The authors should also clarify whether the in-house dataset will be released and provide access to the source code to support independent verification. Validation on additional datasets would help establish the generalizability of the proposed method.

Reviewer_CfRgOpenReview ↗
Rating
2
Confidence
1
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes MorphControl, a diffusion-based framework for generating nuclear fuel-rod surface defect images. The method combines class-level morphology priors with background, geometric, and text conditions, and introduces a quality-feedback mechanism to adjust the morphology-prior strength during inference. Experiments on an in-house dataset show improvements in generation quality and diversity over several baselines. The generated images also improve downstream defect detection performance when used for data augmentation.

Critical strengths
  1. The method achieves better generation quality and diversity than the compared baselines, with consistent improvements in KID and IC-LPIPS across the three defect categories.

  2. The generated samples improve downstream defect detection performance, particularly for underrepresented defect categories, demonstrating the practical value of the proposed approach.

Critical weaknesses
  1. The proposed DMPA largely builds upon existing techniques, including IP-Adapter-style multi-branch cross-attention, a frozen SDXL backbone, and region-weighted diffusion loss (Sec. 3.3). Similarly, QFPR applies a conventional PID controller to adjust conditioning strength (Sec. 3.4), making the overall methodological contribution relatively incremental.

  2. The class-level morphology prior is implemented as a bank of learnable tokens Pk∈RM×dP_k \in \mathbb{R}^{M \times d}, optimized using samples from the same category (Eq. 5–6). However, no comparison with standard class embeddings or analysis of the learned representations is provided to demonstrate that these tokens capture meaningful morphological information.

  3. QFPR adjusts a single scalar conditioning strength within a narrow range of [0.80, 1.20] over only three feedback iterations (Sec. 3.4 and Appendix B.2). Since the final output is selected based on CLIP feature consistency and an artifact threshold, it remains unclear whether PID control offers advantages over simple grid search or fixed candidate selection.

  4. Table 2 only compares DMPA with and without QFPR, leaving the individual effects of defect suppression, morphology priors, region-weighted loss, and decorrelation loss unexplored. Moreover, all experiments are conducted on a single in-house nuclear fuel-rod dataset with three defect categories, limiting evidence of generalizability to other industrial domains.

Questions for authors
  1. How does the proposed class-level morphology prior differ from standard learnable class embeddings? Is there any evidence that it captures meaningful morphological information?

  2. How much of the performance improvement comes from the PID controller itself rather than generating and selecting multiple candidates?

  3. Can the authors provide ablation results for the morphology prior, region-weighted loss, and background–morphology decorrelation loss to clarify their individual contributions?

  4. How well does the proposed method generalize to other industrial defect datasets beyond the nuclear fuel-rod dataset?

1 comment

Thank you for the clarifications, and I maintain my original score after considering the rebuttal.

Reviewer_fDGkOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes MorphControl for generating nuclear fuel rod surface defect images with limited and class-imbalanced training data. Its main components are DMPA and QFPR. DMPA combines learnable class-level morphology priors with separately encoded background, geometric, and textual conditions and injects them into a frozen SDXL backbone. During inference, QFPR adjusts the morphology-prior injection strength based on regional feature consistency and an artifact metric, then selects a candidate satisfying the quality constraints. The authors evaluate the method on 1,467 images of artificially introduced physical defects spanning three categories. Compared with the selected baselines, MorphControl achieves a lower average KID and a higher average IC-LPIPS. Augmenting the training set with 600 manually verified generated images per category improves the average mAP50 and mAP50:95 of YOLOv8s by 4.14 and 2.22 percentage points, respectively.

Critical strengths

1.The generated data not only improves the reported generation metrics but also enhances the average detection performance on the real-collected test set, providing preliminary evidence for the value of data augmentation. The authors also report the performance drop in high-IoU metrics for the pit category, presenting the results objectively. 2.The research problem holds significant practical value, and the method design is highly tailored to the task characteristics. Specifically, the class-shared representations, region-weighted training, and background conditions effectively address core challenges such as class imbalance, tiny defect learning, and background preservation in industrial scenarios, respectively.

Critical weaknesses

1.The current experiments do not sufficiently isolate the gains contributed by the core design, requiring additional ablation results. In Table 2, comparing a fixed intensity with a "generate multiple times and select the best" strategy fails to disentangle the contribution of feedback control from that of the additional search budget—or whether the improvement simply stems from the luck of generating more images. An ablation study is needed to isolate this factor. Furthermore, the key internal modules of DMPA lack independent ablation. Additionally, Table 1 does not include the most directly relevant specialized defect generation methods. It is necessary to incorporate the specialized defect generation models mentioned in the Related Work (e.g., DFMGAN, AnomalyDiffusion) as comparison baselines, supplementing quantitative metrics and qualitative visualizations to conduct a horizontal comparison with MorphControl. The current results are insufficient to demonstrate that the proposed mechanism has a clear advantage over simple alternatives or existing methods. 2.The paper augments the training data with manually verified generated images and their detection annotations. However, it does not specify whether the manual verification process involved discarding images or correcting bounding boxes, nor does it report the image retention rate or human labor costs. In real-world industrial data augmentation scenarios, it is impossible to manually verify and filter every generated image. Therefore, the current results only demonstrate that data augmentation is effective when "humans are in the loop," but they do not prove that a fully automated pipeline can achieve comparable gains. We suggest that the authors first clarify the specific details of the manual operations. Subsequently, they should add a set of downstream comparative experiments: retaining QFPR's automatic quality control while skipping manual image screening, and clearly stating how the annotations were obtained. This setting is more aligned with real-world industrial deployment. This would help disentangle the contributions of image screening, annotation correction, and the generation method itself.

Questions for authors
  1. How many candidates are generated and evaluated per final output by the fixed-strength baseline and QFPR in Table 2? Are there comparisons using the same candidate budget and selection rule to distinguish the contribution of feedback control from that of additional generation and candidate selection?
  2. Have you evaluated downstream detection using generated data without manual image screening or bounding-box correction, while retaining QFPR's automatic quality control? If so, please report the results and clarify how the detection annotations are obtained in this fully automated setting.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.