acceptodds
Under review as a conference paper at ICLR 2027

SwiftAD: Visibility-Guided Hierarchical Token Pruning for Efficient Zero-Shot 3D Anomaly Detection

29 views

Abstract

Zero-shot 3D anomaly detection must localize subtle defects without target-domain training data while remaining efficient enough for practical inspection. Existing CLIP-based methods render point clouds into multiple high-resolution views and process all visual tokens, even though many tokens correspond to empty backgrounds, regions without valid 3D support, or redundant normal surfaces. We present SwiftAD, an efficient framework centered on visibility-guided hierarchical token pruning. SwiftAD first converts rendering-derived 3D-to-2D correspondences into a geometric visibility mask and removes unsupported background tokens without additional learnable parameters. It then applies a lightweight anomaly-aware selector to retain informative tokens from the remaining object regions. We find that token pruning introduces a distributional shift in the visual features, weakening their alignment with the original CLIP text prototypes. Textual prototype adaptation therefore learns separate global and local prototypes for object-level recognition and point-level localization, respectively, and applies residual calibration to their anomaly prototypes. Experiments on MVTec3D-AD, Eyecandies, and Real3D-AD show a favorable accuracy–efficiency trade-off. SwiftAD preserves strong zero-shot performance while halving computational cost and improving throughput by 73.9% over the unpruned baseline.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Zero-shot, 3D anomaly detection, CLIP, Token pruning, Efficiency

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_r6PMOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

SwiftAD accelerates cross-dataset zero-shot 3D anomaly detection using a frozen CLIP backbone. Rendering-derived visibility first removes unsupported background tokens; a learned selector then retains informative foreground tokens. Decoupled global/local text prototypes and residual adaptation compensate for pruning-induced feature shifts. Evaluation on MVTec3D-AD, Eyecandies, and Real3D-AD demonstrates a strong accuracy-efficiency tradeoff. Training uses labeled source data, with no target-domain adaptation.

Critical strengths
  1. Substantial practical gains. Relative to the unpruned configuration, computation falls from 1830.65 to 911.60 GFLOPs and throughput rises from 2.57 to 4.47 FPS. Table 1 reports strong performance across three datasets, including 98.4% P-AUROC and 93.7% AUPRO on MVTec3D-AD and 81.8% P-AUROC on Real3D-AD. These results support the practical value of the approach.
  2. A coherent design supported by ablations. Geometric visibility offers an inexpensive, task-specific pruning signal. Tables 2, 3, and 6 support hierarchical pruning and branch-specific prototype adaptation, while multi-view visualizations illustrate complementary retained evidence. LFP improves localization and reduces computation further, with only a small object-level accuracy tradeoff relative to VGBP plus adaptation.
Critical weaknesses
  1. The learned selector is not fully specified, and performance attribution needs a stronger control. Appendix B gives loss weights, but the sparsity and anomaly-focus objectives remain described mainly in prose. Equation 13 defines a straight-through mask, yet its connection to hard token compaction and downstream task gradients is unclear. Table 2 isolates the effect of adding LFP to VGBP plus adaptation, but lacks an unpruned model with the same prototype adaptation, leaving the full accuracy gain over the unpruned reference partly confounded.
  2. The model-selection protocol and robustness comparisons need clarification. Tables 3 and 4 report target-dataset configuration sweeps without stating whether the default was chosen independently of target test performance. This is a missing protocol detail, not evidence of test-set tuning. Table 5 reports robustness under perturbations, but an unpruned control would help quantify additional sensitivity from pruning. The measured efficiency gains are useful; the timing boundary needs clarification to assess end-to-end benefit.
Questions for authors
  1. Attribution of the accuracy gain. Table 2 already shows the incremental effect of LFP over VGBP plus adaptation. Can the authors add an unpruned configuration using the same decoupled prompts and residual prototype adaptation, with matched supervision and training? This would clarify how much of the full model’s accuracy gain over the current unpruned reference comes from adaptation, alongside the computation savings from pruning.
  2. Target-independent model selection. In the labeled-source transfer setting described in Appendix A, were the pruning layers, keep ratio, resolution, and view count fixed using source-domain validation before target evaluation? Please distinguish the ablations in Tables 3 and 4 from the selection procedure and clarify whether the same selection rule applies across datasets.
  3. Selector training and gradient flow. Please provide the exact sparsity and anomaly-focus objectives, including patch-label generation and anomaly coverage. How is the straight-through mask used alongside hard top-k compaction, and which losses provide gradients to selected and discarded token scores?
  4. Defect preservation and robustness. Can the authors quantify retained defect tokens by defect size and compare the full model with an unpruned model using the same prototype adaptation under Table 5 perturbations? Table 2’s recovery after adaptation is informative, but the VGBP-only drop does not by itself prove that defect evidence was removed. These controls would help assess retention and additional sensitivity caused by pruning.
  5. Timing protocol. Section 4.2 and Appendix B specify an RTX 3090 and batch size 2. Do the FPS measurements use that inference batch size, what numerical precision and timing procedure are used, and do they include rendering and back-projection? Please separate selector overhead and report typical retained-token counts to make the measured efficiency gains easier to interpret.
2 comments

Dear Reviewer,

Thank you very much for your careful and constructive review. We sincerely appreciate your positive assessment of SwiftAD’s practical efficiency gains, coherent design, and comprehensive evaluation. Your comments helped us substantially improve the clarity, reproducibility, and analysis of the paper. All revisions in the resubmitted manuscript are marked in blue. We have also released anonymized code at https://anonymous.4open.science/r/SwiftAD-B754/.

  1. Selector objective, patch supervision, and gradient flow

We agree that the original description of the learned foreground selector was insufficiently detailed. We have clarified the selector losses, patch-label construction, hard token compaction, and straight-through gradient mechanism. For each rendered view, the pixel-level anomaly mask is converted to a patch-level label by adaptive max pooling onto the 24×2424 \times 24 patch grid. A patch is labeled anomalous if it contains at least one anomalous pixel. After VGBP compaction, the corresponding patch labels are gathered using the same token indices, so selector supervision is always aligned with the compact visible-token sequence.

In Stage 2, the objective combines the task loss, the sparsity loss (weight 0.1), the anomaly-focus loss (weight 0.5), and the prototype-anchor regularizer (weight 0.5). The sparsity loss contains an MSE penalty between the soft retained-token ratio and the target keep ratio, together with a score-spread term preventing score collapse. The anomaly-focus loss combines positive-class-weighted patch-level BCE, anomaly coverage, and normal-sample regularization. We use a positive-token weight of 5.0, anomaly-coverage weight of 1.0, score-spread coefficient of 0.05, and straight-through temperature of 0.7.

For forward inference, LFP uses an exact sample-wise hard top-kk mask and physically compacts the selected tokens before they enter the remaining Transformer blocks; thus, the measured acceleration comes from genuine sequence shortening rather than soft masking. During training, we use a sigmoid surrogate around the sample-wise top-kk threshold in the backward pass. Therefore, selected tokens receive task-loss gradients through the straight-through mask, while the selector scores for both selected and discarded tokens are directly supervised by the sparsity and anomaly-focus objectives. Discarded tokens are not processed by downstream Transformer layers after hard compaction.

  1. Attribution of PA and pruning

We agree that an unpruned model with identical PA is necessary to isolate the accuracy contribution of adaptation. We therefore added the unpruned + PA control in Table 2, using the same source-domain supervision, decoupled prompts, and residual prototype adaptation as the full model.

VGBPLFPPAO-AUROCAPP-AUROCAUPROGFLOPsFPS
✗✗✗75.192.194.382.21830.652.57
✗✗✓84.095.198.393.31830.652.57
✓✗✓84.395.298.493.81165.123.70
✓✓✓84.595.298.293.4911.604.47

The results show that PA independently improves O-AUROC by 8.9 points over the unpruned baseline, while the full model further improves O-AUROC to 84.5 and reduces computation by approximately half. Thus, PA mainly compensates for cross-domain and pruning-induced representation mismatch, whereas VGBP and LFP provide the principal computational savings.

  1. Target-independent model selection

We confirm that there is no target-domain leakage. No target-domain sample, annotation, validation metric, or test result is used to optimize the prompt learners, residual prototypes, foreground selector, or final configuration. The pruning layers, keep ratio, resolution, view count, optimizer, learning rates, and loss weights are determined using source-domain-only adaptation and fixed before target evaluation. The target-domain ablations in Tables 3 and 4 are diagnostic analyses rather than target-guided model selection. They are reported to characterize sensitivity, not to choose the final operating point.

The only source-dataset-dependent setting is the number of training epochs. Since the source datasets differ in size, their fitting and convergence speeds differ; the epoch count is therefore determined solely by source-domain training behavior. It is never selected according to target-domain performance.

  1. Defect preservation and robustness

To assess whether pruning preserves anomaly evidence of different spatial scales, we added a defect-size analysis in Table 5 and Fig. 6. Table 5 stratifies anomalous samples according to the ground-truth defect area and reports object-level and point-level metrics for each size interval, while Fig. 6 visualizes the relationship between defect area, sample count, and anomaly-detection performance. The results show that localization remains strong even for small defects: for the 0–99 pixel bin, SwiftAD achieves 97.5% P-AUROC and 88.8% AUPRO. Since this bin contains only eight anomalous samples, we explicitly caution against over-interpreting its image-level metrics. Across the substantially larger 100–2999 pixel range, P-AUROC remains between 98.5% and 99.2%.

We further provide representative qualitative visualizations for small, medium, and large defects in Appendix G. These examples show the rendered views, view-wise anomaly maps, fused 3D predictions, and ground-truth masks. They illustrate that the retained tokens preserve localized anomaly responses for sparse defects, while also supporting coherent multi-view fusion for larger anomalous regions.

In addition, Table 7 compares SwiftAD with the matched unpruned + PA control under Gaussian noise, calibration shifts, and mixed perturbations. The results show that pruning has negligible influence under calibration shifts. Under strong geometric noise, pruning introduces some additional object-level sensitivity; however, point-level localization remains strong. For example, under the most severe mixed condition, SwiftAD retains 95.7% P-AUROC and 87.9% AUPRO, compared with 97.0% and 88.0% for the unpruned + PA control. We have explicitly acknowledged this limitation in the revised conclusion.

  1. Timing protocol and retained-token statistics

We appreciate the request for a clearer timing boundary. Appendix F now provides the hardware setting, precision, batch size, warm-up, timed sample count, synchronization procedure, stage-wise latency, selector overhead, and token-retention statistics.

Training uses batch size 2, whereas latency is measured with batch size 1 to report per-object online latency. Measurements use one RTX 3090 GPU, FP32 inference, ten warm-up batches, and 100 timed objects with CUDA synchronization at stage boundaries.

The reported 4.47 FPS is cached-view online throughput: it includes device transfer, visibility-mask construction, visual encoding, selector scoring, 2D anomaly scoring, 3D back-projection, post-processing, and image-level scoring. Multi-view rendering itself is excluded because it is a shared preprocessing cost for both the unpruned baseline and SwiftAD; both methods use the same renderer, nine views, and input resolution. Hence, the reported 73.9% gain should be interpreted as an online encoder-side improvement within the same rendering-based CLIP pipeline, rather than as a claim of rendering acceleration.

Finally, Appendix F reports that VGBP retains 71.8% of the original patch tokens, while LFP retains 35.9% on average, corresponding to 206.7 tokens per view. The selector scorer is separately listed for transparency but is not double-counted, since it is part of the visual encoder.

We thank the reviewer again for these insightful suggestions. We believe the additional controls, detailed selector specification, source-only selection clarification, robustness analysis, and transparent timing protocol significantly strengthen the revised paper.

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

This manuscript proposes SwiftAD, an efficient zero-shot 3D anomaly detection method. Built upon multi-view rendering and CLIP, the proposed approach reduces redundant computation through Visibility-Guided Background Token Pruning (VGBP) and Learnable Foreground Pruning (LFP). It further introduces Prototype Adaptation (PA) to mitigate feature distribution shifts caused by token pruning. Experiments on MVTec3D-AD, Eyecandies, and Real3D-AD demonstrate an approximately 50% reduction in computational cost and a 73.9% improvement in throughput, while maintaining competitive anomaly detection performance.

Critical strengths
  1. This manuscript addresses redundant token computation in multi-view rendering by introducing a hierarchical pruning strategy based on geometric visibility information. The proposed approach effectively reduces computational costs while maintaining competitive anomaly detection performance, demonstrating practical value.
  2. The proposed method achieves competitive anomaly detection and localization results across three 3D anomaly detection benchmarks. The manuscript also provides relatively comprehensive evaluations through component ablations, parameter analyses, and robustness experiments, demonstrating a favorable trade-off between detection accuracy and computational efficiency.
Critical weaknesses
  1. This manuscript largely follows the multi-view rendering and CLIP-based detection paradigm established by PointAD [1] and related methods. Its main contribution lies in combining existing geometric visibility information with token pruning techniques, resulting in relatively limited methodological novelty. Furthermore, prior work [2] has explored zero-shot 3D anomaly detection directly on native point clouds, but this alternative approach is not discussed in the manuscript. Given that computational efficiency is the primary motivation of this work, the authors should further clarify the advantages and limitations of rendering-based approaches compared with native point cloud-based methods.
  2. The current experiments are insufficient to fully demonstrate the necessity of the key components. As shown in Table 2, VGBP+PA achieves an O-AUROC of 83.9%, outperforming the complete model, which achieves 83.4%. This suggests that LFP does not provide additional improvements in object-level anomaly detection performance. Furthermore, the manuscript lacks an ablation experiment using PA alone without token pruning, as well as comparisons with simpler pruning strategies, such as random pruning and attention-based pruning. Consequently, the current experiments do not clearly distinguish the actual sources of performance improvements or sufficiently establish the unique advantages of the learnable foreground pruning module over existing approaches.

[1] Qihang Zhou, Jiangtao Yan, Shibo He, Wenchao Meng, and Jiming Chen, “PointAD: Comprehending 3D Anomalies from Points and Pixels for Zero-Shot 3D Anomaly Detection,” in Advances in Neural Information Processing Systems (NeurIPS), vol. 37, 2024, pp. 84866–84896. [2] Kaiqiang Li, Gang Li, Mingle Zhou, Min Li, Delong Han, and Jin Wan, “Back to Point: Exploring Point-Language Models for Zero-Shot 3D Anomaly Detection,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2026, pp. 14167–14177.

Questions for authors
  1. Could the authors provide an additional ablation experiment using only Prototype Adaptation (PA) without any token pruning to clarify the independent contributions of each component?
  2. Does LFP offer clear advantages over simpler pruning strategies, such as random pruning, attention-score-based pruning, or feature-similarity-based pruning? Could the authors provide comparisons under the same token budget?
  3. Prior work [1] has explored zero-shot anomaly detection directly on native point clouds. Could the authors explain their motivation for continuing to adopt multi-view rendering and discuss the differences between these two approaches in terms of computational efficiency and anomaly localization performance?
  4. Do the reported efficiency metrics account for the entire pipeline, including rendering, visibility computation, and 3D back-projection? Could the authors provide a detailed analysis of end-to-end inference latency?

[1] Kaiqiang Li, Gang Li, Mingle Zhou, Min Li, Delong Han, and Jin Wan, “Back to Point: Exploring Point-Language Models for Zero-Shot 3D Anomaly Detection,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2026, pp. 14167–14177.

2 comments

Dear Reviewer,

Thank you very much for your careful and constructive review. We sincerely appreciate your positive assessment of the practical value of visibility-guided hierarchical pruning, as well as your thoughtful suggestions regarding methodological positioning, component attribution, and efficiency evaluation. These comments have been very helpful for improving the paper. All revisions in the resubmitted manuscript are highlighted in blue. We have also released the anonymized code at https://anonymous.4open.science/r/SwiftAD-B754/.

  1. Methodological contribution and attribution of PA, VGBP, and LFP

We agree that the original manuscript did not sufficiently isolate the effects of prototype adaptation (PA), geometric pruning, and learnable foreground pruning (LFP). We have therefore expanded the component ablation in Table 2 by adding an unpruned + PA control. This setting uses exactly the same source-domain training protocol, decoupled prompts, and residual prototype adaptation as the full model, but does not use token pruning.

VGBPLFPPAO-AUROCAPP-AUROCAUPROGFLOPsFPS
✗✗✗75.192.194.382.21830.652.57
✗✗✓84.095.198.393.31830.652.57
✓✗✓84.395.298.493.81165.123.70
✓✓✓84.595.298.293.4911.604.47

The new results clarify the roles of the components: PA independently improves cross-domain transfer: compared with the unpruned baseline, unpruned + PA improves O-AUROC from 75.1 to 84.0. VGBP + PA reduces computation by 36.4% while preserving or slightly improving performance. LFP further reduces computation from 1165.12 to 911.60 GFLOPs and increases throughput from 3.70 to 4.47 FPS. Compared with VGBP + PA, it improves O-AUROC by 0.2 points and preserves AP, while incurring only small point-level trade-offs (0.2 P-AUROC and 0.4 AUPRO).

We therefore clarify the intended role of LFP in the revision: it is primarily designed to achieve a more favorable accuracy–efficiency trade-off, rather than to guarantee uniform improvements on every individual object-level and point-level metric. PA compensates for representation mismatch, VGBP removes geometrically unsupported background tokens, and LFP removes redundant foreground tokens while preserving the localization evidence needed for practical inspection.

To further test whether LFP is better than simpler selection rules, we added an equal-budget comparison in Table 4. All methods use VGBP and PA, follow the same source-domain training protocol, and retain the same 50% foreground-token budget.

Foreground selection strategyP-AUROCAUPROO-AUROCAP
Random91.366.255.482.8
Feature similarity89.357.946.678.3
Attention score96.485.980.693.5
LFP (ours)98.293.484.595.2

LFP consistently outperforms random, feature-similarity, and attention-based pruning under the same token budget. In particular, compared with attention-based selection, LFP improves O-AUROC by 3.9 points and AUPRO by 7.5 points. This result demonstrates that the learned selector is not merely an acceleration module: its anomaly-aware supervision helps retain more informative foreground evidence for both detection and localization.

  1. Relationship to native point-cloud methods

We thank the reviewer for recommending BTP [Li et al., CVPR 2026], which we find highly inspiring. We have added a discussion of this line of work in the revised Related Work and added a corresponding outlook in the Conclusion.

SwiftAD is built on the multi-view rendering + CLIP paradigm. This paradigm leverages mature 2D vision-language semantics and supports dense anomaly-map generation, multi-view aggregation, and 3D back-projection. However, processing all tokens from multiple high-resolution rendered views introduces substantial redundancy. SwiftAD addresses this issue within the rendering-based CLIP pipeline: compared with the unpruned baseline, it reduces computation from 1830.65 to 911.60 GFLOPs and improves throughput from 2.57 to 4.47 FPS.

BTP follows a complementary native point-cloud paradigm that directly models 3D points without rendering them into multi-view images. Because the two methods use different input representations and evaluation protocols, we do not regard their reported efficiency or accuracy as directly comparable, and therefore discuss BTP qualitatively rather than including it in the main comparison table. More broadly, we believe that the two paradigms have complementary strengths. Rendering-based methods can leverage strong visual-semantic priors from vision-language models and explicit 3D-to-2D correspondences, which are useful for visibility-aware pruning and multi-view anomaly localization. Native point-cloud methods preserve geometric information directly and avoid rendering overhead. We especially agree that, for discrete and irregular point clouds, token fusion may be a more suitable direction than directly transferring image-style token pruning. Rather than discarding point tokens independently, a point-based model could fuse redundant tokens from geometrically uniform or semantically similar local neighborhoods, while preserving anomaly-sensitive boundary and local-structure information. Such geometry-aware token fusion could reduce computational redundancy while better maintaining the topology and local continuity of discrete point clouds. We have added this perspective as a promising future direction in the revised Conclusion.

  1. Timing boundary and efficiency measurement

We thank the reviewer for highlighting the need to clarify the timing boundary. We acknowledge that the original description could more clearly distinguish the rendering cost from the online inference cost.

The reported 4.47 FPS and 73.9% throughput improvement refer to the cached-view online inference pipeline, i.e., after the multi-view images have been rendered. The timing includes device transfer, visibility-mask construction, visual encoding, selector scoring, 2D anomaly scoring, 3D back-projection, post-processing, and image-level scoring. It is measured on one RTX 3090 GPU using FP32 inference, batch size one, ten warm-up batches, and CUDA-synchronized timing over 100 objects. The detailed stage-wise timing and retained-token statistics are provided in Appendix F.

Multi-view rendering itself is not included in the reported FPS because it is a shared preprocessing cost for both the unpruned baseline and SwiftAD: both methods use the same renderer, the same nine views, and the same input resolution. SwiftAD is designed to reduce the redundant visual-token computation that occurs after rendering, rather than to accelerate rendering itself. Therefore, the reported 73.9% gain should be interpreted as an encoder-side/online inference improvement within the same rendering-based CLIP paradigm. We agree that, when rendering is performed online and dominates the application latency, the full point-cloud-to-result speedup will be smaller than the cached-view speedup reported in the paper. We have clarified this timing boundary in our response and view rendering acceleration, as well as native point-cloud processing, as important complementary directions for future work.

We thank the reviewer again for these valuable comments. We believe that the additional PA-only control, equal-budget foreground-selection comparisons, discussion of BTP, and explicit timing-boundary clarification and cached-view online latency analysis significantly improve the clarity, rigor, and transparency of the revised paper.

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

The paper accelerates multi-view CLIP-based 3D anomaly detection through a combination of visibility-guided background pruning and learned foreground token selection. A prototype adaptation mechanism compensates for representation changes caused by pruning. The method learns from a source dataset and transfers to a disjoint target dataset without target training; zero-shot therefore refers to the target-domain setting, rather than absence of training. Evaluation on three 3D anomaly detection benchmarks reports improved detection/localization alongside substantially reduced inference cost. The paper also studies pruning positions, token budgets, adaptation choices, and robustness to acquisition perturbations.

Critical strengths
  • The method addresses a practical cost in multi-view 3D anomaly detection. Using geometric visibility to remove background tokens before learned foreground selection is a sensible hierarchy. The view-count ablation also helps place this trade-off in context.

  • The paper recognizes that token removal changes the representation used for anomaly scoring, rather than treating pruning as a transparent acceleration. Table 2 demonstrates the severe degradation without adaptation, and the additional adaptation comparisons in the appendix provide useful support for addressing this mismatch. The evaluation includes both detection and localization, which is important because a faster classifier alone would not establish the utility of the system.

Critical weaknesses
  • The accuracy contribution of prototype adaptation is not separated from the pruning strategy. In Table 2, the results strongly establish the need for adaptation, but the table does not include an unpruned model with the same adaptation. Appendix C compares adaptation alternatives within the pruned pipeline, which answers a different question. An unpruned adapted control, followed by equal-budget foreground selection controls, would clarify how much of the improved accuracy comes from adaptation and what the learned selection contributes beyond acceleration.

  • The pruning configuration appears sensitive in a way that affects confidence in the chosen operating point. At pruning layers (6, 12), Table 3 reports object AUROC of 83.4 for retention ratio 0.5 but only 53.5 for 0.75, despite the latter retaining more tokens. The broad discussion of early and late pruning does not explain this particular difference. Since the setting assumes no target-domain training, it is important to specify how the layer positions and retention ratio are selected using source-side information, and whether this large difference is reproducible. This is a question about selection reliability, not evidence of target leakage.

Questions for authors
  • How does the unpruned model perform with the same prototype adaptation and source-training protocol? With adaptation held fixed, how does learned foreground selection compare with a simple foreground selection rule at the same retained-token budget?
  • What explains the large change between retention ratios 0.5 and 0.75 at layers (6, 12)? How were the final configuration and stopping criteria selected without target labels, and is the difference stable across training seeds?
2 comments

Dear Reviewer,

Thank you very much for your careful and constructive review. Your comments have been highly valuable in helping us improve the paper. In the revised submission, all modifications are highlighted in blue. We have also organized and released the anonymized code at https://anonymous.4open.science/r/SwiftAD-B754/.

  1. Separating the contributions of prototype adaptation, pruning, and LFP

We agree that the original manuscript did not sufficiently separate the accuracy contribution of prototype adaptation (PA) from that of pruning. We therefore added an unpruned + PA control to Table 2. This control uses exactly the same source-domain training protocol, decoupled prompts, and residual prototype adaptation as the full model, but does not apply token pruning.

VGBPLFPPAO-AUROCAPP-AUROCAUPROGFLOPsFPS
✗✗✗75.192.194.382.21830.652.57
✗✗✓84.095.198.393.31830.652.57
✓✗✓84.395.298.493.81165.123.70
✓✓✓84.595.298.293.4911.604.47

The new results clarify the role of each component. PA alone improves O-AUROC from 75.1 to 84.0, showing that prototype adaptation is important for cross-domain transfer and for compensating the representation mismatch caused by the source/target gap. VGBP + PA preserves or slightly improves performance while reducing computation from 1830.65 to 1165.12 GFLOPs. Adding LFP further reduces the cost to 911.60 GFLOPs and increases throughput from 3.70 to 4.47 FPS, while maintaining or slightly improving image-level performance. Therefore, PA mainly addresses representation/calibration mismatch, whereas hierarchical pruning provides the principal efficiency gain with only a minor performance trade-off.

We also added an equal-budget foreground-token selection comparison in Table 4. All strategies use VGBP and PA, follow the same source-domain training protocol, and retain the same 50% foreground-token budget.

Foreground selection strategyP-AUROCAUPROO-AUROCAP
Random91.366.255.482.8
Feature similarity89.357.946.678.3
Attention score96.485.980.693.5
LFP (ours)98.293.484.595.2

LFP consistently outperforms random, feature-similarity, and attention-based pruning under the same token budget. In particular, compared with attention-based selection, LFP improves O-AUROC by 3.9 points and AUPRO by 7.5 points. These results show that LFP is not merely an acceleration mechanism: its anomaly-aware supervision helps preserve more informative foreground tokens for both detection and localization.

  1. Keep-ratio sensitivity, PA fitting, reproducibility, and target-independent selection

We appreciate the reviewer’s concern regarding the substantial difference between keep ratios of 0.50 and 0.75. This behavior is stable and reproducible using the released code, configurations, and corresponding interfaces/checkpoints. We further added a diagnostic analysis in Appendix D.

The diagnostic results show that the lower performance at r=0.75r=0.75 is not caused by loss of defect evidence. Both settings preserve nearly all annotated defect tokens: 99.66% ± 1.79% at r=0.50r=0.50 and 99.99% ± 0.21% at r=0.75r=0.75. Their anomaly–normal logit separations are also nearly identical (5.5230 versus 5.5198). Thus, the degradation cannot be explained simply by discarding anomaly-related tokens or by reduced early score separation. More specifically, token pruning changes not only the number of tokens processed by the subsequent visual blocks, but also the distribution of the retained sequence. Changing the keep ratio alters the relative composition of normal-surface, boundary, and anomaly-related tokens, as well as the gradient statistics received by the local adaptation branch. Consequently, the optimization landscape and fitting/convergence speed of PA change with the token budget.

In our experiments, we intentionally keep the optimizer, learning rate, loss weights, residual parameterization, and other training settings fixed across keep ratios, such that the comparison isolates the effect of the token budget rather than introducing ratio-specific tuning. Under this fixed source-only protocol, the PA configuration that is well calibrated for r=0.50r=0.50 cannot fully absorb the changed optimization dynamics at r=0.75r=0.75. This is also reflected by the larger local PA residual norm at r=0.75r=0.75 (0.051112 versus 0.036535), while the global residual remains similar (0.064218 versus 0.066879). We therefore attribute the non-monotonic behavior primarily to different PA fitting/calibration regimes induced by different token budgets, rather than to a simple feature-level loss of defect evidence.

Importantly, we do not use target-domain labels to compensate for this effect. Ratio-specific tuning based on target results would violate our zero-shot transfer setting. The default pruning layers, keep ratio, view number, resolution, optimizer settings, and loss weights are fixed using source-domain-only adaptation before target evaluation. The target-domain ablations are reported only for analysis; they are not used to select the final configuration.

Finally, there is no target-domain leakage in training or model selection. No target sample, target annotation, target validation metric, or target test result is used to optimize the prompt learners, residual prototypes, or foreground selector. All learnable components are adapted only using labeled source-domain data. The only dataset-dependent training setting is the number of epochs: because the source datasets differ in size, their fitting/convergence speeds differ, and the epoch count is determined solely from the corresponding source-domain training process. It is never chosen based on target-domain performance.

We thank the reviewer again for these valuable suggestions. We believe that the added controls, same-budget comparisons, Appendix D diagnostics, and released code substantially improve the attribution, reproducibility, and transparency of SwiftAD.

Comments on the paper

1 comment

Dear Reviewers,

Thank you for your time and effort in reviewing this submission. Given the mixed opinions among reviewers, I would like to kindly encourage you to engage in the discussion, consider the authors’ responses, and update your ratings as appropriate.

Thank you again for your valuable contributions!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.