acceptodds
Withdrawn from ICLR 2027

TPFreeDAS: Trainable-Parameter-Free DINO-Guided Anomaly Synthesis for Few-Shot Anomaly Detection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Few-shot anomaly detection commonly relies on synthetic anomalies to compensate for the scarcity of real anomalous samples and annotations. However, CutPaste-based synthesis faces three progressively finer compatibility conditions: valid foreground placement, a policy suited to image structure, and perturbations compatible with local material or functional regions. To address these conditions, we propose TPFreeDAS, a trainable-parameter-free DINO-guided anomaly synthesis framework for few-shot anomaly detection that directly reuses the frozen DINO representations of existing DINO-based detectors without introducing an additional feature backbone. Specifically, Foreground-Constrained Anomaly Synthesis (FCAS) derives an instance-level foreground mask from frozen DINO patch responses and restricts anomaly placement to valid spatial regions. Structure-Adaptive Synthesis Routing (SASR) characterizes image structure using foreground coverage, feature dispersion, and boundary complexity, and adaptively selects an appropriate synthesis path. For selected structured categories, Region-Aware and Quality-Gated Synthesis (RAQG) constructs pseudo-structural regions, generates anomalies within compatible regions, and filters candidates according to feature changes inside and outside the intended anomaly mask. TPFreeDAS changes only training-time synthesis, adding no learnable synthesis parameters or inference-time overhead. On PCBMarket, it reduces 4-shot per-image synthesis latency by 35.51% relative to our CutPaste implementation and exceeds FoundAD by up to percentage points in I-AUROC and in PRO; it improves most reported VisA and MVTec AD metrics. All code and model weights will be released.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Few-Shot Anomaly Detection

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_5r3VOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes TPFreeDAS, a trainable-parameter-free anomaly synthesis module, to address the mismatch between synthetic anomalies and target objects in few-shot anomaly detection. The module uses frozen DINO features to identify foreground regions, select synthesis strategies, generate candidate regions, and filter synthetic samples. It is designed to replace CutPaste during the training stage of existing anomaly detectors without introducing additional computational overhead during inference. The paper conducts experiments on three datasets: VisA, MVTec AD, and PCBMarket.

Critical strengths
  1. The proposed anomaly synthesis module operates only during training and reuses the detector's existing frozen DINO representations, introducing no additional inference overhead. Replacing CutPaste across multiple detection frameworks demonstrates a certain degree of transferability.

  2. The paper reports results under three shot settings, across three datasets, and with three support set seeds, and provides component ablation, weight sensitivity, and efficiency analyses. The improvements on PCBMarket are particularly notable.

  3. The method decomposes synthetic anomaly quality into three levels: foreground location, image structure, and local regions. This problem decomposition demonstrates a certain degree of systematic consideration.

Critical weaknesses
  1. Key components of the algorithm are not fully specified: the predefined set of structured categories CstrC_{str}, the region–anomaly type compatibility rules SregS_{reg} , the Gate thresholds and acceptance criteria, the number of clusters, and the PCA dimensionality all lack executable definitions. These components directly determine which samples are generated or rejected, making the current paper impossible to reproduce.

  2. The method consists of multiple fixed weights, quantiles, edge normalization constants, and category routing thresholds, yet the paper does not explain how these hyperparameters are selected when only normal samples are available, or whether target test results are used for tuning. The module ablation studies are mainly limited to a single VisA 2-shot setting, lacking cross-dataset quantitative evaluations and module-level visualizations corresponding to the three motivations.

  3. The paper provides no comparisons of similarity to real defects or comparisons across different anomaly generators to demonstrate that the synthetic samples are more natural and compatible. Improvements in final detection performance alone are insufficient to validate the claimed visual mechanisms of each synthesis module.

Questions for authors
  1. Please provide the default values and selection criteria for CstrC_{str}, SregS_{reg} , Gate, and all PCA/K-Means and anomaly-type rules, and clarify how the development and validation sets are constructed.

  2. If the manually defined list of structured categories is removed and the same automatic rules are applied to all categories, how does performance change across the three datasets?

Reviewer_8gtVOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The proposes TPFreeDAS, a trainable-parameter-free anomaly synthesis framework for few-shot anomaly detection. Built upon FoundAD, it improves CutPaste-based data augmentation through three modules: Foreground-Constrained Anomaly Synthesis (FCAS), Structure-Adaptive Synthesis Routing (SASR), and Region-Aware and Quality-Gated Synthesis (RAQG). These modules use frozen DINO features to guide anomaly placement, select synthesis strategies, and filter synthetic samples. Experiments on multiple benchmarks demonstrate improvements in anomaly detection and localization performance.

Critical strengths
  1. The paper is generally clear and easy to follow.
  2. The experimental results demonstrate the effectiveness of the proposed anomaly synthesis strategy across multiple benchmarks.
Critical weaknesses
  1. Limited novelty and methodological contribution. The paper mainly improves the CutPaste augmentation strategy in FoundAD, while the three proposed modules largely combine existing techniques and are engineering-oriented. The authors should further clarify the key novelty and broader contributions of this work.
  2. Insufficient experimental comparisons. The paper is essentially an anomaly data augmentation method, but it mainly compares against CutPaste and lacks comparisons with more advanced training-free anomaly generation methods such as AnomalyAny[1] and O2MAG[2]. Although the method is evaluated with several detection frameworks, the main results rely on FoundAD, making it difficult to assess its generalizability and whether the strong performance comes from the proposed augmentation strategy or the FoundAD baseline. Could the authors provide broader comparisons and more evidence of the method's generalizability?

[1] Sun H, Cao Y, Dong H, et al. Unseen visual anomaly generation[C]//2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2025: 25508-25517. [2] Rao H, Wang Z, Si C, et al. One-to-More: High-fidelity training-free anomaly generation with attention control[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2026: 28492-28501.

Questions for authors

See weaknesses

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

This paper proposes a trainable-parameter-free DINO-guided framework (TPFreeDAS) to improve anomaly synthesis validity for few-shot anomaly detection. Three modules are introduced, including Foreground Constrained Anomaly Synthesis (FCAS) which estimates placement support from DINO patch responses, Structure-Adaptive Synthesis Routing (SASR) which selects between foreground-constrained and CutPaste paths and Region-Aware and Quality-Gated Synthesis (RAQG) constrains local placement and validates feature changes for selected structured categories. Experiments are conducted on three anomaly detection datasets, including MVTec AD, VisA and PCBMarket.

Critical strengths
  • The motivation is good. An anomaly synthesis model is proposed to address the problem of insufficient anomaly images for anomaly detection.

  • Some ablation studies are provided to facilitate the understanding of how the performance benefits from various components, including FCAS, SASR and RAQG.

Critical weaknesses
  • The entire framework is overly engineered, appearing as a collection of stacked heuristic modules. As shown in Figure 2, including too many blocks reduces the practicality and reproducibility of this work.

  • RAQG only contains some simple modules, such as low-rank PCA, and K-Means. It is unclear whether it can partition well workpieces with complex structures, such as PCB boards, into regions. It is also unclear how the anomaly patterns within the anomaly mask are derived.

  • Visualization results of some key components, such as FCAS, SASR and RAQG, are missing for different image categories; furthermore, as listed in Table 4, these components do not yield a significant performance improvement, particularly with regard to pixel-level metrics, e.g., P-AUROC and PRO.

  • There are too many hyperparameters, and for some of them, it is unclear how they should be set in practice, e.g., the image-adaptive quantile threshold in Eq. (3), the edge-density normalization factor in Eq. (6), the synthesis selection threshold delta for Eq. (7), etc.

  • The experimental results fail to convincingly demonstrate the effectiveness of the proposed method. As listed in Table 2, compared to the baseline method FoundAD, the improvement is marginal on the mainstream VisA and MVTec AD datasets, and the performance even declines on some metrics.

Questions for authors
  1. Can RAQG partition well workpieces with complex structures, such as PCB boards, into regions?

  2. How are the anomaly patterns within the anomaly mask derived?

  3. How should the hyperparameters be set in practice, for example, the image-adaptive quantile threshold in Eq. (3), the edge-density normalization factor in Eq. (6), and the synthesis selection threshold delta for Eq. (7)?

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

Related papers

Discussion (0)

Sign in to comment.