The reviewed record of science sign in
Pith

arxiv: 2607.01949 · v1 · pith:RRWTPGRS · submitted 2026-07-02 · cs.CV

LiZAD: A Lightweight Zero-Shot Anomaly Detection Framework for Industrial Manufacturing

Uzair Khan , Luigi Capogrosso , Muhammad Aqeel , Francesco Setti , Michele Magno , Marco Cristani This is my paper

Reviewed by Pith T0 review T1 audit T2 compute T3 formal T4 kernel 2026-07-03 15:49 UTCgrok-4.3pith:RRWTPGRSrecord.jsonopen to challenge →

classification cs.CV
keywords zero-shot anomaly detectionlightweight frameworkedge deploymentindustrial inspectionfeature alignmentDINOv3MobileCLIP2
0
0 comments X

The pith

LiZAD aligns DINOv3 visual features with MobileCLIP2 text embeddings through small projection heads to cut memory and latency for zero-shot anomaly detection on edge devices.

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

Modern factories change product designs often enough that collecting labeled data for every new item becomes impractical. Zero-shot anomaly detection can spot defects without target-specific training data, yet most current methods demand too much memory and compute for edge hardware. LiZAD pairs the spatially detailed features from DINOv3 with the compact text embeddings from MobileCLIP2 and routes both through low-memory trainable projection heads into one shared space. On four standard industrial datasets the resulting system uses 61.5 percent less memory and 74.6 percent fewer parameters than six prior zero-shot models while running three times faster, at the cost of a 6.4 percent average drop in pixel-level detection score. The same pipeline has already been placed on NVIDIA Jetson edge boards and run on an active production line.

Core claim

The paper claims that low-memory trainable projection heads can map dense DINOv3 visual features and efficient MobileCLIP2 text embeddings into a shared latent space that supports competitive pixel-level zero-shot anomaly detection, delivering 61.5 percent average memory reduction, 74.6 percent parameter reduction, and 3.02 times latency speedup relative to six state-of-the-art ZSAD models across VisA, BTAD, MPDD, and MVTec-AD while remaining deployable on Jetson NX and AGX hardware.

What carries the argument

Low-memory trainable projection heads that align DINOv3 visual features with MobileCLIP2 text embeddings into a shared latent space for anomaly scoring and localization.

If this is right

  • Real-time zero-shot inspection becomes practical on resource-limited factory hardware without per-product retraining.
  • The same alignment approach can be tested on other pairs of dense visual and compact text encoders for further efficiency gains.
  • Memory and parameter budgets on Jetson-class devices now suffice for continuous anomaly monitoring on changing production lines.
  • Pixel-level localization remains usable even after the large reductions in model size and latency.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The projection-head technique might transfer to other multimodal industrial tasks such as zero-shot classification or segmentation on the same hardware.
  • If the alignment proves stable across more varied lighting and texture conditions, the framework could reduce the need for custom data collection in additional manufacturing sectors.
  • Further compression of the projection heads themselves could be explored to reach even lower power budgets while monitoring any additional drop in localization accuracy.

Load-bearing premise

The small projection heads can align the two feature streams well enough to keep pixel-level anomaly localization within 6.4 percent of the best prior models.

What would settle it

An independent re-run on the same four datasets that measures P-AUROC more than 6.4 percent below the reported value when the projection heads are removed or replaced by fixed linear layers while keeping the identical DINOv3 and MobileCLIP2 backbones.

Figures

Figures reproduced from arXiv: 2607.01949 by Francesco Setti, Luigi Capogrosso, Marco Cristani, Michele Magno, Muhammad Aqeel, Uzair Khan.

Figure 1
Figure 1. Figure 1: The comparative analysis of LiZAD against six [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of LiZAD architecture. at the pixel level through an anomaly map Yˆ ∈ {0, 1} H×W , where higher values indicate a higher likelihood of anomaly. B. Overview of LiZAD Existing ZSAD methods are commonly based on CLIP￾style models (details in Section II-B). Although such models are effective for global image-text matching, their visual representations are primarily optimized for semantic alignment and… view at source ↗
Figure 4
Figure 4. Figure 4: Real-world deployment results of LiZAD. The first [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative comparison of ZSAD methods on differ [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
read the original abstract

In modern high-throughput industrial production lines, product configurations and visual characteristics frequently change, making it impractical to collect and annotate data for every new scenario. This dynamic setting makes Zero-Shot Anomaly Detection (ZSAD) particularly suitable, as it enables defect detection without requiring training on target-specific samples. Although recent ZSAD approaches show promising results, they are computationally intensive and thus unsuitable for deployment on resource-constrained devices. We propose LiZAD: a lightweight framework designed for real-time ZSAD specifically tailored for use on edge devices. The proposed approach pairs the dense and spatially aware visual features of DINOv3, crucial for precise pixel-level localization, with the highly computationally efficient text embeddings of MobileCLIP2. These features are then mapped into a shared latent space via low-memory trainable projection heads. Compared to six state-of-the-art ZSAD models, LiZAD achieves an average memory reduction of 61.5%, a parameter reduction of 74.6%, and a speedup of 3.02x in terms of latency. Despite substantial reductions in computational and memory costs, our approach maintains competitive anomaly detection performance, dropping the average P-AUROC by just 6.4% relative to the best state-of-the-art model across the VisA, BTAD, MPDD, and MVTec-AD datasets. Finally, it is successfully deployed on the NVIDIA Jetson NX and Jetson AGX edge devices and tested on the real production line of the Industrial Computer Engineering Laboratory (ICE Lab) at the University of Verona. The code is available at https://github.com/intelligolabs/LiZAD.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper introduces LiZAD, a lightweight zero-shot anomaly detection (ZSAD) framework for industrial settings that pairs DINOv3 visual features with MobileCLIP2 text embeddings, aligns them via small trainable projection heads, and deploys the result on edge hardware. It reports average reductions of 61.5% memory, 74.6% parameters, and 3.02x latency versus six prior ZSAD models, with only a 6.4% average drop in P-AUROC across VisA, BTAD, MPDD, and MVTec-AD, plus successful real-line deployment on NVIDIA Jetson devices; code is released.

Significance. If the efficiency-accuracy trade-off holds under scrutiny, the work directly addresses the deployment barrier for ZSAD on resource-limited industrial hardware. The open code repository is a clear strength that supports reproducibility and further validation.

major comments (2)
  1. [Experiments] Experiments (reported averages): the headline 6.4% P-AUROC drop and efficiency deltas are presented as single aggregate numbers without reported standard deviations, number of runs, or dataset-split details; this makes it impossible to judge whether the observed trade-off is statistically stable or sensitive to particular train/test partitions.
  2. [Method] Method (projection heads): the central performance claim rests on the assumption that the low-memory trainable heads align DINOv3 spatial tokens with MobileCLIP2 embeddings while preserving pixel-level anomaly localization; no ablation (e.g., removing the heads) or diagnostic (cosine-similarity heatmaps before/after projection) is provided to verify that the alignment step does not collapse the localization signal.
minor comments (2)
  1. [Method] Notation for the shared latent space dimension and the exact architecture of the projection heads should be stated explicitly (e.g., layer widths, activation functions) rather than left as “low-memory trainable projection heads.”
  2. [Deployment] Figure captions for the deployment results on Jetson devices should include the exact batch size, input resolution, and measured power draw to allow direct comparison with the reported latency numbers.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. We address each major comment below and will revise the manuscript to incorporate the requested analyses and statistical details.

read point-by-point responses

  1. Referee: [Experiments] Experiments (reported averages): the headline 6.4% P-AUROC drop and efficiency deltas are presented as single aggregate numbers without reported standard deviations, number of runs, or dataset-split details; this makes it impossible to judge whether the observed trade-off is statistically stable or sensitive to particular train/test partitions.

    Authors: We agree that the current presentation lacks statistical rigor. In the revised manuscript we will report results from five independent runs (different random seeds for projection-head training), include mean and standard deviation for all P-AUROC and efficiency metrics, and explicitly document the train/test splits and any cross-validation procedure used on VisA, BTAD, MPDD, and MVTec-AD. revision: yes

  2. Referee: [Method] Method (projection heads): the central performance claim rests on the assumption that the low-memory trainable heads align DINOv3 spatial tokens with MobileCLIP2 embeddings while preserving pixel-level anomaly localization; no ablation (e.g., removing the heads) or diagnostic (cosine-similarity heatmaps before/after projection) is provided to verify that the alignment step does not collapse the localization signal.

    Authors: We acknowledge the absence of direct verification. The revised version will contain (i) an ablation that removes the projection heads and measures the resulting drop in P-AUROC and localization quality, and (ii) side-by-side cosine-similarity heatmaps (and anomaly maps) computed before and after the projection step on representative images to demonstrate that spatial structure is retained. revision: yes

Circularity Check

0 steps flagged

No circularity; empirical claims rest on public-dataset benchmarks

full rationale

The manuscript describes an engineering framework that pairs off-the-shelf DINOv3 and MobileCLIP2 encoders with small trainable projection heads, then reports measured memory, parameter, latency, and P-AUROC numbers on four standard public datasets. No equations, uniqueness theorems, or self-citations are invoked to derive the performance deltas; the deltas are obtained by direct measurement against external baselines. The projection-head alignment step is an empirical design choice whose effectiveness is assessed by the same held-out test sets, not by any reduction to quantities defined inside the paper itself.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The framework rests on the domain properties of two external pre-trained models and the effectiveness of learned projection heads; no new physical entities or ad-hoc constants are introduced.

free parameters (1)
  • projection head weights
    Trainable low-memory layers are fitted to align the two feature spaces; their values are not reported as fixed constants.
axioms (2)
  • domain assumption DINOv3 supplies dense spatially aware visual features suitable for pixel-level localization
    Invoked directly as the reason for choosing DINOv3.
  • domain assumption MobileCLIP2 supplies highly computationally efficient text embeddings
    Invoked as the source of efficiency in the paired design.

pith-pipeline@v0.9.1-grok · 5849 in / 1388 out tokens · 31657 ms · 2026-07-03T15:49:15.865087+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

40 extracted references · 40 canonical work pages · 1 internal anchor

  1. [1]

    Deep Industrial Image Anomaly Detection: A Survey,

    J. Liuet al., “Deep Industrial Image Anomaly Detection: A Survey,” Machine Intelligence Research, vol. 21, no. 1, pp. 104–135, 2024

    work page 2024

  2. [2]

    Towards Real Unsupervised Anomaly Detection Via Confident Meta-Learning,

    M. Aqeelet al., “Towards Real Unsupervised Anomaly Detection Via Confident Meta-Learning,” inIEEE/CVF International Conference on Computer Vision (ICCV), 2025

    work page 2025

  3. [3]

    KairosAD: A SAM-Based Model for Industrial Anomaly Detection on Embedded Devices,

    U. Khanet al., “KairosAD: A SAM-Based Model for Industrial Anomaly Detection on Embedded Devices,” inInternational Conference on Image Analysis and Processing (ICIAP), 2026

    work page 2026

  4. [4]

    Diffusion-Based Image Generation for In- Distribution Data Augmentation in Surface Defect Detection,

    L. Capogrossoet al., “Diffusion-Based Image Generation for In- Distribution Data Augmentation in Surface Defect Detection,” inIn- ternational Conference on Computer Vision Theory and Applications (VISAPP), 2024

    work page 2024

  5. [5]

    Leveraging Latent Diffusion Models for Training- Free in-Distribution Data Augmentation for Surface Defect Detection,

    F. Girellaet al., “Leveraging Latent Diffusion Models for Training- Free in-Distribution Data Augmentation for Surface Defect Detection,” inInternational Conference on Content-Based Multimedia Indexing (CBMI), 2024

    work page 2024

  6. [6]

    Towards Total Recall in Industrial Anomaly Detection,

    K. Rothet al., “Towards Total Recall in Industrial Anomaly Detection,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022

    work page 2022

  7. [7]

    Collaborative Discrepancy Optimization for Reliable Image Anomaly Localization,

    Y . Caoet al., “Collaborative Discrepancy Optimization for Reliable Image Anomaly Localization,”IEEE Transactions on Industrial Infor- matics, vol. 19, no. 11, pp. 10 674–10 683, 2023

    work page 2023

  8. [8]

    WinCLIP: Zero-/Few-Shot Anomaly Classification and Segmentation,

    J. Jeonget al., “WinCLIP: Zero-/Few-Shot Anomaly Classification and Segmentation,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023

    work page 2023

  9. [9]

    Efficient Visual Anomaly Detection at the Edge: Enabling Real-Time Industrial Inspection on Resource-Constrained De- vices,

    A. Stropeniet al., “Efficient Visual Anomaly Detection at the Edge: Enabling Real-Time Industrial Inspection on Resource-Constrained De- vices,”arXiv preprint arXiv:2603.20288, 2026

    work page arXiv 2026

  10. [10]

    DINOv3

    O. Siméoniet al., “DINOv3,”arXiv preprint arXiv:2508.10104, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  11. [11]

    MobileCLIP2: Improving Multi-Modal Reinforced Training,

    F. Faghriet al., “MobileCLIP2: Improving Multi-Modal Reinforced Training,”Transactions on Machine Learning Research, 2025

    work page 2025

  12. [12]

    TinyGLASS: Real-Time Self-Supervised In-Sensor Anomaly Detection,

    P. Bonazziet al., “TinyGLASS: Real-Time Self-Supervised In-Sensor Anomaly Detection,”arXiv preprint arXiv:2603.16451, 2026

    work page arXiv 2026

  13. [13]

    Enhancing Safety and Privacy in Industry 4.0: The ICE Laboratory Case Study,

    F. Cunicoet al., “Enhancing Safety and Privacy in Industry 4.0: The ICE Laboratory Case Study,”IEEE Access, vol. 12, pp. 154 570–154 599, 2024

    work page 2024

  14. [14]

    Reconstruction by inpainting for visual anomaly detection,

    V . Zavrtaniket al., “Reconstruction by inpainting for visual anomaly detection,”Pattern Recognition, vol. 112, p. 107706, 2021

    work page 2021

  15. [15]

    Self-Supervised Predictive Convolutional Attentive Block for Anomaly Detection,

    N.-C. Risteaet al., “Self-Supervised Predictive Convolutional Attentive Block for Anomaly Detection,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022

    work page 2022

  16. [16]

    Pni : Industrial anomaly detection using position and neighborhood information,

    J. Baeet al., “Pni : Industrial anomaly detection using position and neighborhood information,” inIEEE/CVF International Conference on Computer Vision (ICCV), 2023

    work page 2023

  17. [17]

    PANDA: Adapting Pretrained Features for Anomaly Detection and Segmentation,

    T. Reisset al., “PANDA: Adapting Pretrained Features for Anomaly Detection and Segmentation,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021

    work page 2021

  18. [18]

    PaDiM: A Patch Distribution Modeling Framework for Anomaly Detection and Localization,

    T. Defardet al., “PaDiM: A Patch Distribution Modeling Framework for Anomaly Detection and Localization,” inInternational Conference on Pattern Recognition Workshops (ICPRW), 2021

    work page 2021

  19. [19]

    Fastflow: Unsupervised anomaly detection and localization via 2d normalizing flows

    J. Yuet al., “FastFlow: Unsupervised Anomaly Detection and Local- ization via 2D Normalizing Flows,”arXiv preprint arXiv:2111.07677, 2021

    work page arXiv 2021

  20. [20]

    Same Same but DifferNet: Semi-Supervised Defect Detection With Normalizing Flows,

    M. Rudolphet al., “Same Same but DifferNet: Semi-Supervised Defect Detection With Normalizing Flows,” inIEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2021

    work page 2021

  21. [21]

    CFLOW-AD: Real-Time Unsupervised Anomaly Detection with Localization via Conditional Normalizing Flows,

    D. Gudovskiyet al., “CFLOW-AD: Real-Time Unsupervised Anomaly Detection with Localization via Conditional Normalizing Flows,” in IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2022

    work page 2022

  22. [22]

    Exploiting Multimodal Latent Diffusion Models for Accurate Anomaly Detection in Industry 5.0,

    L. Capogrossoet al., “Exploiting Multimodal Latent Diffusion Models for Accurate Anomaly Detection in Industry 5.0,” inItal-IA, 2024

    work page 2024

  23. [23]

    CutPaste: Self-Supervised Learning for Anomaly Detection and Localization,

    C.-L. Liet al., “CutPaste: Self-Supervised Learning for Anomaly Detection and Localization,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021

    work page 2021

  24. [24]

    DRAEM - A Discriminatively Trained Recon- struction Embedding for Surface Anomaly Detection,

    V . Zavrtaniket al., “DRAEM - A Discriminatively Trained Recon- struction Embedding for Surface Anomaly Detection,” inInternational Conference on Computer Vision (ICCV), 2021

    work page 2021

  25. [25]

    SimpleNet: A Simple Network for Image Anomaly Detection and Localization,

    Z. Liuet al., “SimpleNet: A Simple Network for Image Anomaly Detection and Localization,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023

    work page 2023

  26. [26]

    A Unified Anomaly Synthesis Strategy with Gradient Ascent for Industrial Anomaly Detection and Localization,

    W. Chenet al., “A Unified Anomaly Synthesis Strategy with Gradient Ascent for Industrial Anomaly Detection and Localization,” inEuropean Conference on Computer Vision (ECCV), 2024

    work page 2024

  27. [27]

    Learning Transferable Visual Models From Natural Language Supervision,

    A. Radfordet al., “Learning Transferable Visual Models From Natural Language Supervision,” inInternational Conference on Machine Learn- ing (ICML), 2021

    work page 2021

  28. [28]

    arXiv preprint arXiv:2305.17382 , year=

    X. Chenet al., “A Zero-/Few-Shot Anomaly Classification and Seg- mentation Method for CVPR 2023 V AND Workshop Challenge Tracks 1&2: 1st Place on Zero-shot AD and 4th Place on Few-shot AD,”arXiv preprint arXiv:2305.17382, 2023

    work page arXiv 2023

  29. [29]

    AnomalyCLIP: Object-agnostic Prompt Learning for Zero-shot Anomaly Detection,

    Q. Zhouet al., “AnomalyCLIP: Object-agnostic Prompt Learning for Zero-shot Anomaly Detection,” inInternational Conference on Learning Representations (ICLR), 2023

    work page 2023

  30. [30]

    AdaCLIP: Adapting CLIP with Hybrid Learnable Prompts for Zero-Shot Anomaly Detection,

    Y . Caoet al., “AdaCLIP: Adapting CLIP with Hybrid Learnable Prompts for Zero-Shot Anomaly Detection,” inEuropean Conference on Com- puter Vision (ECCV), 2024

    work page 2024

  31. [31]

    Bayesian Prompt Flow Learning for Zero-Shot Anomaly Detection,

    Z. Quet al., “Bayesian Prompt Flow Learning for Zero-Shot Anomaly Detection,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025

    work page 2025

  32. [32]

    PaSTe: Improving the Efficiency of Visual Anomaly Detection at the Edge,

    M. Baruscoet al., “PaSTe: Improving the Efficiency of Visual Anomaly Detection at the Edge,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2025

    work page 2025

  33. [33]

    A SAM-guided Two-stream Lightweight Model for Anomaly Detection,

    C. Liet al., “A SAM-guided Two-stream Lightweight Model for Anomaly Detection,”ACM Transactions on Multimedia Computing, Communications, and Applications, vol. 21, no. 2, pp. 1–23, 2025

    work page 2025

  34. [34]

    A Machine Learning-Oriented Survey on Tiny Machine Learning,

    L. Capogrossoet al., “A Machine Learning-Oriented Survey on Tiny Machine Learning,”IEEE Access, vol. 12, pp. 23 406–23 426, 2024

    work page 2024

  35. [35]

    AA-CLIP: Enhancing Zero-Shot Anomaly Detection via Anomaly-Aware CLIP,

    W. Maet al., “AA-CLIP: Enhancing Zero-Shot Anomaly Detection via Anomaly-Aware CLIP,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025

    work page 2025

  36. [36]

    MVTec AD – A Comprehensive Real-World Dataset for Unsupervised Anomaly Detection,

    P. Bergmannet al., “MVTec AD – A Comprehensive Real-World Dataset for Unsupervised Anomaly Detection,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019

    work page 2019

  37. [37]

    ReConPatch: Contrastive Patch Representation Learning for Industrial Anomaly Detection,

    J. Hyunet al., “ReConPatch: Contrastive Patch Representation Learning for Industrial Anomaly Detection,” inIEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2024

    work page 2024

  38. [38]

    Deep learning-based defect detection of metal parts: evaluating current methods in complex conditions,

    S. Jezeket al., “Deep learning-based defect detection of metal parts: evaluating current methods in complex conditions,” inInternational Congress on Ultra Modern Telecommunications and Control Systems and Workshops (ICUMT), 2021, pp. 66–71

    work page 2021

  39. [39]

    SPot-the-Difference Self-supervised Pre-training for Anomaly Detection and Segmentation,

    Y . Zouet al., “SPot-the-Difference Self-supervised Pre-training for Anomaly Detection and Segmentation,” inEuropean Conference on Computer Vision (ECCV), 2022

    work page 2022

  40. [40]

    SuperSimpleNet: Unifying Unsupervised and Super- vised Learning for Fast and Reliable Surface Defect Detection,

    B. Rolihet al., “SuperSimpleNet: Unifying Unsupervised and Super- vised Learning for Fast and Reliable Surface Defect Detection,” in International Conference on Pattern Recognition (ICPR), 2024

    work page 2024