Pith. sign in

REVIEW 4 major objections 4 minor 46 references

Keep the Needle, Prune the Haystack: Defect-Preserving Token Pruning for Efficient Zero-Shot Anomaly Detection

T0 review · 4 major / 4 minor · reviewed 2026-08-11 · deepseek-v4-flash

Pith's one-line read KeepAD prunes the vision transformer down to fewer than 20% of tokens while holding zero-shot anomaly detection accuracy within 2.7 AUROC points across industrial and medical benchmarks.

desk verdict Solid empirical paper on defect-preserving token pruning for zero-shot AD; averages hide per-dataset drops and some hand-tuning, but the ablations and physical-survivor metrics make it worth serious review. read the letter →

arxiv 2608.03681 v3 pith:T3L2GWZ3 submitted 2026-08-04 cs.CV

classification cs.CV
keywords zero-shotanomalydetectiontokenpruningvisiontransformerdefectlocalizationefficientinferenceself-distillationimage-adaptivebudgetCLIP
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper claims that token pruning inside a frozen vision transformer can be made safe for zero-shot anomaly detection by treating token selection as high-recall, defect-preserving routing rather than a generic efficiency trick. KeepAD keeps fewer than 20% of image tokens at inference and still holds average image-level and pixel-level AUROC degradation within 2.7 percentage points across six industrial and seven medical benchmarks, while reaching up to 7.9x throughput over the strongest CLIP-based baseline. This matters because anomaly evidence is scarce: most computation is spent on normal regions, but pruning can accidentally delete the only tokens that reveal a defect. The resolution is a two-stage router that is conservative early, aggressive late, and supervised by a dense-to-sparse self-distillation signal that costs nothing at inference.

What carries the argument

The central object is the two-stage progressive router inserted at layers 8 and 12 of the frozen backbone. At L8 the selector computes a prototype-free visual score, then applies coverage-preserving selection over non-overlapping $2\times2$ patch blocks (one representative per non-empty block) followed by a deterministic rescue of up to 16 diversity candidates and remaining high-score tokens; at L12 it combines the visual score with the sigmoid of the difference between cosine similarities to frozen abnormal and normal prototypes, standardizes the fused risks per image, and chooses whole $2\times2$ blocks under a budget controlled by the dispersion and upper-tail of those risks. The selectors are trained with a hard-topology, soft-gradient construction plus dense-to-sparse self-distillation, in which an unpruned teacher pass supplies detached deeper-layer anomaly responses that align the early routing; at inference only a single hard-pruned forward pass remains, and original-grid coordinates of survivors support nearest-survivor dense recovery.

What would settle it

A concrete check: on a new benchmark of small defects, stratify the complete-miss rate by ground-truth defect size. If defects spanning 1-4 tokens are missed in more than the roughly 2.7% of images the paper reports for its smallest-defect stratum at the final budget, the coverage floor is not preserving the smallest evidence.

Watch

Extended reading notes

Core claim

The central claim is that token pruning for zero-shot anomaly detection should be formulated as a risk-aware routing problem, not a classification-preserving compression problem. KeepAD does this with two stages inside a frozen CLIP ViT-L/14 backbone: at layer 8, before anomaly semantics are reliable, it uses coverage-preserving selection over local $2\times2$ patch blocks plus a deterministic rescue of high-score and spatially diverse tokens; at layer 12, it fuses visual salience with the difference between affinities to frozen normal and abnormal prototypes, then selects whole blocks under an image-adaptive budget that spends more compute on images with sharp or heavy-tailed anomaly evidence. Dense localization is recovered by mapping each dropped position to its nearest surviving token on the original grid, while image-level scores use only physical survivors so recovered positions cannot fake anomaly evidence. On six industrial and seven medical zero-shot anomaly detection benchmarks, the paper reports retention below 20%, average I-AUROC/P-AUROC degradation within 2.7 points, and up to 7.9x throughput over the strongest CLIP-based baseline.

Load-bearing premise

The result depends on the frozen normal and abnormal prototypes, trained on synthetic collage-and-crop defects from source-domain data only, transferring to unseen industrial and medical targets well enough that the inference-time calibrations keep working without target-domain tuning.

Editorial extensions

If this is right

  • At the most aggressive operating point, the model retains 14-18% of tokens and reaches 41.1 img/s, a 7.9x throughput gain over the strongest CLIP-based baseline with only 1.9-2.2 average AUROC loss on industrial transfer.
  • Ablations show that L8 pruning is the main efficiency lever: removing it costs 6.5-8.3 img/s, more than removing L12 pruning (2.7-3.2 img/s), while the L21 early exit adds 0.9-3.3 img/s with no AUROC change.
  • Replacing L8 coverage with global top-k increases complete misses from below 1% to 4.0-6.3% and costs 8.9-10.7 P-AUROC points, so the coverage floor is what makes aggressive early pruning safe.
  • The image-adaptive budget works per image: shuffling the same budget multiset across images raises the complete-miss rate by at least 4 points, so the gain comes from matching compute to evidence rather than from aggregate compute.
  • Dense-to-sparse self-distillation is the key training signal for the shallow router: without it, complete-miss rate jumps from below 1% to 4.3-6.9% and I-AUROC drops by 2.7-4.9 points.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • An implication the paper leaves implicit is that the coverage-plus-rescue design could transfer to other sparse-evidence tasks, such as medical lesion detection or video anomaly detection, where missing a small region is far costlier than wasting compute.
  • The paper's physical-survivor metrics DTR and CMR are a useful template for the field: post-recovery pixel AUROC can stay high even when every defect token was pruned, so any pruning method claiming to preserve dense predictions should report survivor-based recall, not just recovered maps.
  • A natural extension the authors do not pursue is to make the L8 keep fraction image-adaptive as well, since it is currently a user-chosen operating point (30-70%) rather than a learned decision.
  • The depth analysis suggests token separability is non-monotonic, so the L8/L12 breakpoints are a trade-off; one could test whether the same two-stage idea should place pruning layers per backbone rather than fixing them globally.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper proposes KeepAD, a two-stage token pruning framework for zero-shot anomaly detection built on a frozen CLIP ViT-L/14@336px detector. Stage 1 trains a dense anomaly detector with per-layer normal/anomaly prototypes; Stage 2 freezes it and trains lightweight selectors at layers 8 and 12. At L8, selection is coverage-preserving over non-overlapping 2x2 blocks with a deterministic rescue set; at L12, block-level selection uses a fusion of visual salience and frozen-prototype affinity under an image-adaptive budget. Dense-to-sparse self-distillation supervises the shallow router during training. Localization is recovered by nearest-survivor assignment on the original patch grid, while image-level scores aggregate only physical survivors. Experiments across six industrial and seven medical benchmarks report final token retention below 20% at the prune70 operating point, average I-AUROC/P-AUROC degradation within 2.7 percentage points, and up to 7.9x throughput over Bayes-PFL, the strongest CLIP-based baseline.

Significance. If the empirical claims hold, KeepAD is a useful and well-posed contribution to efficient zero-shot anomaly detection. The paper identifies an asymmetric pruning risk and an early-pruning paradox, and translates them into a concrete two-stage router with physical-survivor tracking. The evaluation is unusually thorough for this area: matched-budget controls for the adaptive budget (Table 5), physical defection-preservation metrics DTR/CMR, seed stability (Table 18), backbone generality (Table 9), comparison with generic token reducers in the same frozen detector (Table 6), and a medical-suite transfer study (Table 19) are all present, and the code is released. The main value is in demonstrating that anomaly-aware, coverage-preserving routing can largely preserve detection and localization while substantially reducing compute. The principal caveat is that some load-bearing components, especially the L12 controller constants and the frozen prototypes, are hand-tuned and only partially stress-tested under domain transfer.

major comments (4)
  1. [§4.1 and Appendix A.6] The two cross-dataset directions use different Stage-2 checkpoints (epoch 7 for MVTec-to-VisA, epoch 10 for VisA-to-MVTec), and Appendix A.6 states that the Stage-2 data configuration uses the source-domain test split. Please clarify the checkpoint selection criterion and the exact split used for selector training. If the checkpoint epoch was chosen using target-domain test performance, or if the source-domain test split (with its real anomaly labels) was used for training, the reported numbers do not constitute a clean cross-dataset zero-shot evaluation. Please specify the selection rule, and if necessary rerun with a fixed epoch and with selector training restricted to the source-domain train split.
  2. [§3.2, Eq. (4), Appendix A.4 and F.2] The deep-stage router and its image-adaptive budget rely on constants rho_min, rho_max, c, T_b, kappa, on the frozen L12 prototypes from Stage 1, and on the independently calibrated sparse-mid gate q(I). These are listed in Table 7, but no sensitivity analysis is reported for them. The medical-transfer ablation in Table 19 shows that component conclusions do shift under domain transfer (replacing local coverage with global top-k changes P-AUROC/CMR by -8.3/+6.8; removing self-distillation by -1.4/+3.4). Please add a sensitivity study over the controller and gate constants on source validation and on the medical suite, and state explicitly whether the constants were fixed before any target-domain results were inspected.
  3. [Appendix B and Tables 1/2/3/19] Appendix B correctly warns that post-recovery P-AUROC cannot verify defect-token retention and that the 2x2 block coverage floor does not guarantee defect survival. However, the main benchmark tables report only I-AUROC and P-AUROC; DTR and CMR are reported only for the two transfer directions and as a medical-suite average. Since the central localization claim depends on physical defect preservation before recovery, please report per-benchmark DTR and CMR for the main operating points, and make clear in the abstract and results that pixel-level AUROC is measured after nearest-survivor recovery rather than on physical survivors alone.
  4. [Abstract and Table 2] The abstract's 'within 2.7 percentage points' is an average degradation across benchmarks and could be read as a per-dataset guarantee. The actual per-dataset drops at prune70 include -6.8 I-AUROC on BrainMRI and -3.9 P-AUROC on Endo. Please state the per-dataset ranges or add a clear qualification that individual datasets can degrade substantially more than the average.
minor comments (4)
  1. [§1 and Table 1] The '7.9x speedup over the strongest CLIP-based baseline' is a system-level comparison: it comes from comparing KeepAD at prune70 with Bayes-PFL, while the same pruning gives only about 1.48x over the paper's own dense detector. This is a legitimate comparison, but the abstract should state the baseline-specific nature of the 7.9x number.
  2. [Figure 2 and Appendix C.2] Figure 2 reports 'AUC' values, but the caption does not define what is being measured. The appendix later clarifies that these are token-level salience AUCs of a transplanted L8 scorer; the caption should be updated accordingly.
  3. [Table 2] Several medical datasets report only I-AUROC or only P-AUROC (for example, OCT17 and CVC-ClinicDB). Adding a column note or an em-dash convention would help readers understand the metric coverage.
  4. [Appendix A.3, Eq. (10)] The notation TopCell and TopScore is used without an explicit definition; please state that TopCell(min(16,R8), ...) selects the highest-scoring remaining token per coarse 8x8 cell and TopScore selects by global score, as described in the surrounding text.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: KeepAD's claims are measured benchmark results evaluated against ground-truth labels, and no load-bearing derivation reduces to its own inputs.

full rationale

The paper is an empirical systems paper. Its central claims—sub-20% token retention with average I/P-AUROC degradation within 2.7 points, and up to 7.9x throughput—are reported measurements on held-out target test sets (Tables 1–2), not quantities derived from fitted parameters. The routing components (coverage-preserving L8 selection, prototype-guided L12 block selection, image-adaptive budget) are defined independently of the evaluation metrics DTR, CMR, I-AUROC, and P-AUROC; these metrics are computed from ground-truth masks and physical survivor indices (Appendix B), so success is externally falsifiable. The frozen L12 prototypes from Stage 1 are reused in Stage 2, but this is parameter sharing within a single pipeline, not a logical reduction of the target claim to an input: the final evaluation uses target-domain ground-truth labels that the frozen prototypes were never fitted to. The paper's own component ablations and medical-transfer analysis (Tables 4–5, 17–19) show that individual components can help or hurt differentially under domain shift, which would be impossible if the results were forced by construction. The only self-citation (VisualAD, Hou et al. 2026) appears in related work as a baseline and is not load-bearing for the method's validity. Concerns about miscalibrated frozen prototypes on unseen targets are correctness/robustness risks, not circularity, because the paper does not claim a derivation that presupposes its conclusion. No qualifying circular step was found.

Assumptions & free parameters 7 free parameters · 4 assumptions · 0 invented entities

No new physical entities are introduced. The paper's extra assumptions are all domain-modeling choices: synthetic-anomaly transferability, depth-wise salience as a routing signal, the nearest-survivor recovery rule, and any-overlap token labeling. The free-parameter count is moderate for a systems paper, with the main burden falling on the calibrated controller constants and inference-only gate.

free parameters (7)
  • L8/L12 pruning layers = L8 and L12
    Chosen by a fixed-scorer depth scan as a trade-off between separability and remaining computation, and then validated by end-to-end placement ablations. The choice is post hoc but the placement ablation supports it.
  • L12 controller constants c, T_b, kappa = c=0.55, T_b=0.04, kappa=0.65
    Appendix A.4 lists them; they map the evidence statistic to a keep fraction and are hand-set, though the matched-budget controls show adaptive allocation helps beyond the specific constants.
  • L12 budget bounds rho_min, rho_max = 0.25 and 0.50
    Hand-set bounds governing the realized L12 keep range; the paper does not report a sensitivity sweep over these bounds.
  • Stage-2 loss weights = 0.45/0.22/0.60/0.25
    Appendix A.5 gives the weights; the ablation shows each term contributes, but optimal weights are not derived.
  • Sparse-mid gate parameters = gate temperature 0.5, token fraction 0.05, mid-layer recipe
    Inference-only image-score calibration chosen on source data and reported in Appendix A; the paper states the gate is source-calibrated.
  • Checkpoint epoch selection = epoch 7 for MVTec to VisA, epoch 10 for VisA to MVTec
    Different epochs chosen per transfer direction, which weakens the claim of one shared recipe.
  • Top-3% tail fraction and temperatures = 0.03 and 0.12/0.20/0.08
    Evidence-tail and relaxation temperatures are hand-set constants. The E_var versus E_tail decomposition supports their role, but the values are not swept.
assumptions (4)
  • domain assumption The Stage-1 dense detector, trained on synthetic collage-and-crop anomalies from a source dataset, provides transferable anomaly prototypes and responses under the cross-dataset ZSAD protocol.
    The whole Stage-2 routing supervision relies on frozen L12 anomaly responses and prototypes; Section 3.1 and Appendix F.2 test this transfer empirically but do not prove it.
  • domain assumption Token salience at L8 and L12, as measured by the transplanted scorer, is a valid proxy for routing quality.
    Figure 2 and Appendix C.2 report Token AUC at different depths; the choice of pruning layers relies on this proxy.
  • domain assumption Nearest-survivor recovery of dropped positions is acceptable for pixel-level evaluation.
    Section 3.4 defines the recovery; the paper explicitly says recovery cannot restore deleted evidence and separates DTR/CMR from recovered P-AUROC.
  • domain assumption Any-overlap token labeling using resized ground-truth maps is a correct ground truth for defect-token recall.
    Appendix B defines DTR and CMR on any-overlap patch labels; this is a modeling choice that influences reported DTR/CMR values.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Keep the Needle, Prune the Haystack: Defect-Preserving Token Pruning for Efficient Zero-Shot Anomaly Detection." pith.science (2026). https://pith.science/paper/T3L2GWZ3

@misc{pith2026260803681,
  author       = {Pith},
  title        = {Pith review of: Keep the Needle, Prune the Haystack: Defect-Preserving Token Pruning for Efficient Zero-Shot Anomaly Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/T3L2GWZ3}},
  note         = {Machine review of arXiv:2608.03681}
}
abstract

Zero-shot visual anomaly detection has achieved remarkable progress, with recent vision-only approaches further improving performance while simplifying the inference pipeline. However, existing methods typically perform dense computation over all images and spatial tokens, despite the fact that normal samples dominate real-world scenarios and anomalies usually occupy only small regions. Token pruning offers a promising solution, but introduces an asymmetric pruning risk in anomaly detection: retaining normal tokens mainly incurs redundant computation, whereas removing anomalous tokens may eliminate the only evidence for detection and localization. This risk is particularly severe in early layers, where pruning provides the greatest computational benefit but anomaly semantics remain unreliable. We propose KeepAD, a defect-preserving token pruning framework that formulates token selection as high-recall, anomaly-aware routing. In shallow layers, KeepAD combines coverage-preserving selection over local $2\times2$ patch neighborhoods with deterministic anomaly rescue to reduce the risk of discarding subtle defects. In deeper layers, frozen normal and abnormal prototypes guide pruning under an image-adaptive token budget, aggressively removing low-risk normal tokens while preserving local anomaly evidence. Dense-to-sparse self-distillation further supervises early token routing without introducing additional inference overhead. Experiments on six industrial and seven medical zero-shot anomaly detection benchmarks show that KeepAD reduces the token retention ratio to below $20\%$, while limiting the average degradation in image-level and pixel-level AUROC to within $2.7$ percentage points. At the most aggressive operating point, KeepAD achieves a $7.9\times$ speedup over the strongest CLIP-based baseline.

Figures

Figures reproduced from arXiv: 2608.03681 by the authors.

Figure 1
Figure 1. Accuracy–throughput trade-off. At competi￾tive accuracy, KeepAD achieves up to 7.9× and 1.7× the throughput of Bayes-PFL and VisualAD, respectively; throughput is measured on a single NVIDIA RTX 5090 GPU. Efficient inference is particularly important for practical ZSAD systems. Anomaly evidence is scarce at two levels: anomalous images are rare in deployment streams, and the decisive evidence within an anomalous ima… view at source ↗
Figure 2
Figure 2. Depth-dependent token separability. Results are obtained with CLIP ViT-L/14@336px. L8 and L12 balance token separability with remaining computation. Token pruning for anomaly detection consequently faces two core challenges. (i) Asymmetric pruning risk. Re￾taining redundant normal tokens mainly increases compu￾tation, whereas removing an anomaly-bearing token may irreversibly erase evidence required for detection an… view at source ↗
Figure 3
Figure 3. Overview of KeepAD. Stage 1 trains a dense ViT anomaly detector, which is then frozen. Stage 2 learns an L8 coverage-preserving selector and an L12 prototype-guided selector with an image-adaptive budget. The dense teacher and differentiable masks are used only for selector training. At inference, hard-selected tokens are propagated, their original-grid coordinates support dense recovery, and image-level scoring use… view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Layer-wise KeepAD routing at prune70. Each row [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Class-conditional L12 keep-ratio distributions at prune70. The adaptive controller spends more on anoma￾lous images (50.2% vs. 45.3% mean keep) and keeps most normal images on a lower budget, yielding the main speedup. Dashed lines mark class means. Adaptive budgets fo…
Figure 6
Figure 6. Figure 6: Within-image salience-score distributions across encoder depth. Probability densities of within-image normalized salience scores are shown for normal and defect tokens at every even encoder layer from L2 to L22. The two panels report MVTec AD→VisA (top) and VisA→MVTec …
Figure 7
Figure 7. Figure 7: Same-pipeline pruning-ratio curves. I-AUROC (left) and P-AUROC (right) are plotted against the final total percentage of tokens pruned for both transfer directions. The common grid contains 0%, 10%, 20%, 30%, 40%, 50%, 60%, 70%, 80%, 85%, and 90%; every point runs the …
Figure 8
Figure 8. Figure 8: Exploratory accuracy–throughput Pareto frontiers after the serial latency rerun. I-AUROC (top) and P-AUROC (bottom) are plotted against batch-1 throughput for both transfer directions. Points are colored by local block geometry and jointly vary the pruning-layer pair, …
Figure 9
Figure 9. Figure 9: Final-budget accuracy–throughput trajectories after the serial latency rerun. I-AUROC (top) and P-AUROC (bottom) are shown for both transfer directions. Percentage annotations denote target final keep ratios; the solid lines connect the evaluated budget sequence and th…
Figure 10
Figure 10. Figure 10: Image-adaptive budget behavior on MVTec AD→VisA (prune70). The realized L12 keep ratio rises smoothly with the image-level evidence score E(I) com￾puted from normalized L12 token risks, confirming that the controller spends more computation on evidence-rich im￾ages. A…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

46 extracted references · 1 linked inside Pith

  1. [1]

    European Conference on Computer Vision (ECCV) , year =

    An Image is Worth 1/2 Tokens After Layer 2: Plug-and-Play Inference Acceleration for Large Vision-Language Models , author =. European Conference on Computer Vision (ECCV) , year =

  2. [2]

    Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP) , pages =

    Stop Looking for Important Tokens in Multimodal Language Models: Duplication Matters More , author =. Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP) , pages =

  3. [3]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

    DivPrune: Diversity-based Visual Token Pruning for Large Multimodal Models , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

  4. [4]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

    VisionZip: Longer is Better but Not Necessary in Vision Language Models , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

  5. [5]

    Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , pages =

    LLaVA-PruMerge: Adaptive Token Reduction for Efficient Large Multimodal Models , author =. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , pages =

  6. [6]

    International Conference on Learning Representations (ICLR) , year =

    VisionTrim: Unified Vision Token Compression for Training-Free MLLM Acceleration , author =. International Conference on Learning Representations (ICLR) , year =

  7. [7]

    International Conference on Learning Representations (ICLR) , year =

    Prune Redundancy, Preserve Essence: Vision Token Compression in VLMs via Synergistic Importance-Diversity , author =. International Conference on Learning Representations (ICLR) , year =

  8. [8]

    International Conference on Learning Representations (ICLR) , year =

    Dynamic-LLaVA: Efficient Multimodal Large Language Models via Dynamic Vision-Language Context Sparsification , author =. International Conference on Learning Representations (ICLR) , year =

Show all 46 references
  1. [9]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

    ATP-LLaVA: Adaptive Token Pruning for Large Vision Language Models , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

  2. [10]

    Rethinking Visual Token Reduction in

    Xu, Rui and Wang, Yunke and Luo, Yong and Du, Bo , journal =. Rethinking Visual Token Reduction in. 2026 , doi =

  3. [11]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume =

    Instruction-Guided Cross-Modal Clustering for Training-Free Visual Token Pruning in Vision-Language Models , author =. Proceedings of the AAAI Conference on Artificial Intelligence , volume =. 2026 , doi =

  4. [12]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

    Hybrid Token Compression for Vision-Language Models , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

  5. [13]

    International Conference on Learning Representations (ICLR) , year =

    LearnPruner: Rethinking Attention-based Token Pruning in Vision Language Models , author =. International Conference on Learning Representations (ICLR) , year =

  6. [14]

    Don't Just Chase ``Highlighted Tokens'' in

    Zou, Xin and Lu, Di and Wang, Yizhou and Yan, Yibo and Lyu, Yuanhuiyi and Zheng, Xu and Zhang, Linfeng and Hu, Xuming , booktitle =. Don't Just Chase ``Highlighted Tokens'' in

  7. [15]

    and Okuno, Tomoyuki and Nakata, Yohei and Keutzer, Kurt and Zhang, Shanghang , booktitle =

    Zhang, Yuan and Fan, Chun-Kai and Ma, Junpeng and Zheng, Wenzhao and Huang, Tao and Cheng, Kuan and Gudovskiy, Denis A. and Okuno, Tomoyuki and Nakata, Yohei and Keutzer, Kurt and Zhang, Shanghang , booktitle =

  8. [16]

    International Conference on Learning Representations (ICLR) , year =

    An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale , author =. International Conference on Learning Representations (ICLR) , year =

  9. [17]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

    WinCLIP: Zero-/Few-Shot Anomaly Classification and Segmentation , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

  10. [18]

    2023 , doi =

    Chen, Xuhai and Han, Yue and Zhang, Jiangning , journal =. 2023 , doi =

  11. [19]

    arXiv preprint arXiv:2311.00453 , year =

    CLIP-AD: A Language-Guided Staged Dual-Path Model for Zero-shot Anomaly Detection , author =. arXiv preprint arXiv:2311.00453 , year =

  12. [20]

    International Conference on Learning Representations (ICLR) , year =

    AnomalyCLIP: Object-agnostic Prompt Learning for Zero-shot Anomaly Detection , author =. International Conference on Learning Representations (ICLR) , year =

  13. [21]

    European Conference on Computer Vision (ECCV) , pages =

    AdaCLIP: Adapting CLIP with Hybrid Learnable Prompts for Zero-Shot Anomaly Detection , author =. European Conference on Computer Vision (ECCV) , pages =

  14. [22]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

    Bayesian Prompt Flow Learning for Zero-Shot Anomaly Detection , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

  15. [23]

    arXiv preprint arXiv:2601.05552 , year=

    One Language-Free Foundation Model Is Enough for Universal Vision Anomaly Detection , author=. arXiv preprint arXiv:2601.05552 , year=

  16. [24]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

    VisualAD: Language-Free Zero-Shot Anomaly Detection via Vision Transformer , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

  17. [25]

    2026 , doi =

    Deng, Huilin and Luo, Hongchen and Zhai, Wei and Guo, Yanming and Cao, Yang and Kang, Yu , journal =. 2026 , doi =

  18. [26]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

    Towards Total Recall in Industrial Anomaly Detection , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

  19. [27]

    International Conference on Pattern Recognition (ICPR) Workshops , pages =

    PaDiM: A Patch Distribution Modeling Framework for Anomaly Detection and Localization , author =. International Conference on Pattern Recognition (ICPR) Workshops , pages =

  20. [28]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

    Anomaly Detection via Reverse Distillation from One-Class Embedding , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

  21. [29]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

    MVTec AD -- A Comprehensive Real-World Dataset for Unsupervised Anomaly Detection , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

  22. [30]

    European Conference on Computer Vision (ECCV) , pages =

    SPot-the-Difference Self-Supervised Pre-training for Anomaly Detection and Segmentation , author =. European Conference on Computer Vision (ECCV) , pages =

  23. [31]

    IEEE International Symposium on Industrial Electronics (ISIE) , pages =

    VT-ADL: A Vision Transformer Network for Image Anomaly Detection and Localization , author =. IEEE International Symposium on Industrial Electronics (ISIE) , pages =

  24. [32]

    Computers in Industry , volume =

    Mixed Supervision for Surface-Defect Detection: From Weakly to Fully Supervised Learning , author =. Computers in Industry , volume =

  25. [33]

    2007 , doi =

    Weakly Supervised Learning for Industrial Optical Inspection , author =. 2007 , doi =

  26. [34]

    Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) , pages =

    Zero-shot versus Many-shot: Unsupervised Texture Anomaly Detection , author =. Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) , pages =

  27. [35]

    Advances in Neural Information Processing Systems (NeurIPS) , pages =

    DynamicViT: Efficient Vision Transformers with Dynamic Token Sparsification , author =. Advances in Neural Information Processing Systems (NeurIPS) , pages =

  28. [36]

    International Conference on Learning Representations (ICLR) , year =

    Not All Patches are What You Need: Expediting Vision Transformers via Token Reorganizations , author =. International Conference on Learning Representations (ICLR) , year =

  29. [37]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

    A-ViT: Adaptive Tokens for Efficient Vision Transformer , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

  30. [38]

    International Conference on Learning Representations (ICLR) , year =

    Token Merging: Your ViT But Faster , author =. International Conference on Learning Representations (ICLR) , year =

  31. [39]

    European Conference on Computer Vision (ECCV) , pages =

    Adaptive Token Sampling for Efficient Vision Transformers , author =. European Conference on Computer Vision (ECCV) , pages =

  32. [40]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

    Zero-TPrune: Zero-Shot Token Pruning through Leveraging of the Attention Graph in Pre-Trained Transformers , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

  33. [41]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

    Patch Slimming for Efficient Vision Transformers , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

  34. [42]

    2022 , doi =

    Kong, Zhenglun and Dong, Peiyan and Ma, Xiaolong and Meng, Xin and Niu, Wei and Sun, Mengshu and Shen, Xuan and Yuan, Geng and Ren, Bin and Tang, Hao and Qin, Minghai and Wang, Yanzhi , booktitle =. 2022 , doi =

  35. [43]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume =

    Evo-ViT: Slow-Fast Token Evolution for Dynamic Vision Transformer , author =. Proceedings of the AAAI Conference on Artificial Intelligence , volume =. 2022 , doi =

  36. [44]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

    SparseViT: Revisiting Activation Sparsity for Efficient High-Resolution Vision Transformer , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

  37. [45]

    Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , pages =

    Dynamic Token Pruning in Plain Vision Transformers for Semantic Segmentation , author =. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , pages =

  38. [46]

    Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) , pages =

    Revisiting Token Pruning for Object Detection and Instance Segmentation , author =. Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) , pages =

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.