VID-AD: A Dataset for Image-Level Logical Anomaly Detection under Vision-Induced Distraction

Introduction

• Logical anomalies violate global constraints such as quantity, length, type, placement, and relations, while patch-centric detectors [9] [16] struggle under visual distractions.These violations can be visually subtle and spatially dispersed; background changes, illumination shifts, and blur can trigger false positives even when the logical state is unchanged.Source: §1 · pp. 1–2
• VID-AD provides 10,395 images across 10 manufacturing scenarios and five capture conditions (50 one-class tasks), varying appearance while holding each scenario’s logical constraints fixed.Each scenario uses two logical constraints, with anomalies defined by single-constraint or combined violations.Source: §1 · p. 2
• The proposed framework learns logical consistency from normal-image text descriptions using contrastive learning with rewritten negative examples, without real anomalous training images.A VLM produces structured descriptions, and replacement-only edits create attribute-level contradictions while preserving the original text structure.Source: §1 · p. 2
• The method consistently outperforms existing vision-based baselines across all capture conditions, achieving state-of-the-art performance.The introduction also reports that representative vision-based baselines degrade under the VID-AD setting.Source: §1 · p. 2

2.1. VAD Datasets

• Existing VAD benchmarks emphasize visually explicit structural defects, while even MVTec LOCO AD lacks controlled distractors that preserve logical state, motivating VID-AD.Robustness and multi-view datasets address real-world variation but primarily evaluate visual anomalies rather than logical violations under controlled distraction.Source: §2.1 · p. 3

2.2. VAD Methods

• Vision-centric methods detect local defects effectively but can miss global rule violations and become unstable under appearance changes that leave logical state unchanged.The reviewed approaches include feature deviation, reconstruction, distillation, component consistency, and few-shot matching; the proposed framework instead models logic-focused descriptions in language embedding space to target rule violations and resist appearance distractions.Source: §2.2 · p. 3

3.1. Dataset Overview

• VID-AD provides 50 independent one-class tasks across 10 manufacturing scenarios and five capture conditions, holding logical rules fixed while varying visual appearance.Unlike benchmarks focused on structural defects and MVTec LOCO AD, which mixes logical and structural anomalies, VID-AD focuses exclusively on logical anomalies under environmental variation.Source: §3.1 · pp. 3–4

3.2. Scenarios and Logical Anomaly Taxonomy

• The taxonomy covers quantity, length, type, placement, and relation violations, with ten scenarios pairing two constraints and anomalies violating either or both.Normal samples satisfy both constraints; anomaly subsets isolate violations of the first constraint, the second, or both.Source: §3.2 · p. 4

3.3. Capture Conditions for Vision-Induced Distraction

• Five capture conditions—White BG, Cable BG, Mesh BG, Low-light CD, and Blurry CD—vary backgrounds, illumination, or defocus while preserving scenario rules and forming separate benchmark tasks.Source: §3.3 · p. 4

3.4. Benchmark Protocol and Dataset Statistics

• Across 50 one-class tasks, VID-AD contains 10,395 images (2,500 train; 7,895 test), with training restricted to normal samples and tests containing normal and anomalous images.A typical task has 50 training normal images, 50 testing normal images, and approximately 110 testing anomalous images.Source: §3.4 · p. 4

4.1. Problem Setting and Overview

• VID-AD is a one-class benchmark in which models learn logical constraints from normal samples despite significant appearance variations.The challenge is to decouple underlying logical states from vision-induced distractions.Source: §4.1 · p. 5
• A frozen VLM converts images into logic-focused text descriptions guided by scenario-specific prompts, and detection uses these descriptions rather than raw images.This design prioritizes logical consistency over irrelevant pixel-level fluctuations.Source: §4.1 · p. 5
• To train without real anomalies, the method rewrites normal descriptions into logically inconsistent negative texts.The synthesized negatives remain linguistically coherent.Source: §4.1 · p. 5

S S S

• The end-to-end detector uses a frozen VLM to generate logic-focused texts, contrastive fine-tuning of BERT, and k-nearest-neighbor aggregation to score test descriptions against training embeddings.The pipeline uses k=5, with larger normality scores indicating samples closer to the normal training distribution.Source: Figure 3 · p. 6, S S S · p. 6

4.4. Inference and Statistical Ensemble Scoring

• At inference, the frozen VLM and same prompt produce test text, whose fine-tuned embedding is compared with normal training embeddings using k=5 nearest neighbors; higher normality scores indicate closer matches.Lower scores indicate a significant logical deviation from the established training patterns.Source: Figure 3 · p. 6, §4.4 · p. 8

4.2. Vision-to-Text Description and Negative

• Scenario-specific prompts and structured descriptions target logical attributes while filtering appearance noise; each image yields one positive text, and replacement-only edits create fluent negatives with at least one attribute contradiction.The same prompt is used for training and testing, with regionwise descriptions for grid layouts or ordered sequences for relational rules.Source: §4.2 · pp. 6–7

4.3. Contrastive Fine-tuning of Text Encoder

• For each task, pretrained BERT is fine-tuned with dropout-generated positive pairs and contradiction-based negatives, using an InfoNCE/NT-Xent-style loss to pull consistent texts together and repel contradictions.Mean-pooled, L2-normalized embeddings represent the texts; dropout provides two stochastic views of each positive description.Source: §4.3 · p. 7

5.1. Experimental Settings

• The method generates one positive text and one contradiction-based negative text per image using Qwen2-VL-7B-Instruct [44], then fine-tunes BERT independently for each task.Positive generation used γ=0.001 and negative rewriting γ=0.9; both used top-p=0.95 and a 512-token limit. Evaluation reports image-level AUROC, using maximum spatial response for pixel-map methods and aggregated pixel reconstruction error for reconstruction methods.Source: §5.1 · p. 8

5.2. Baselines

• The baselines span feature-space methods PaDiM and PatchCore, reconstruction-based AnoGAN and VAE [40], distillation-based EfficientAD, and segmentation-based CSAD [16].Source: §5.2 · p. 8

5.3. Performance under Different Capture Conditions

• Our method achieved the best AUROC under all five capture conditions, beating the second-best method, CSAD, by 0.132–0.207 AUROC and recording a standard deviation of 0.013.Its best-to-worst condition gap was 0.037, compared with 0.033 for UniVAD, whose absolute accuracy was lower.Source: §5.3 · p. 8, Table 4 · p. 9

5.5. Scenario-wise Robustness and Linguistic Expressibility

• Scenario sensitivity was 0.006 for Fruits and 0.009 for Ropes, but 0.096 for Sticks, where descriptions sometimes omitted rule-critical relations such as relative length.The authors relate robustness to linguistic expressibility: Fruits descriptions preserved rule-relevant attributes across conditions, while Sticks descriptions could omit them. Low sensitivity alone does not guarantee strong detection, but the method combined high AUROC with low sensitivity across multiple scenarios.Source: §5.5 · p. 9, Table 5 · p. 10

5.6. Ablation Study

• Qwen2 7B achieved the best mean AUROC (0.831) and smallest cross-condition variance (Std = 0.013), ahead of Llama3.2 11B (0.822±0.036) and LLaVA 13B (0.701±0.034).The prompt, BERT encoder, contrastive objective, and scoring strategy were held constant across the three VLMs; Llama3.2 11B led under White BG, Cable BG, and Mesh BG, but degraded under Low-light CD and Blurry CD.Source: §5.6 · pp. 9–10

5.4. Qualitative Analysis of Vision-centric Evidence

• CSAD’s anomaly maps varied across capture conditions, with Mesh BG inducing spurious activations and Low-light CD or Blurry CD producing diffuse or weakened localization.The proposed language-based method has no pixel-level anomaly maps, so the paper makes no direct map comparison; its scoring instead depends on consistency of logic-focused descriptions.Source: §5.4 · p. 9, Figure 4 · p. 10

Robustness Under Vision-Induced Distraction

• Language-embedding detection that emphasizes logical attributes over low-level visual features reduces sensitivity to nuisance factors such as background changes and capture degradations.The discussion connects this robustness under vision-induced distraction to representation invariance.Source: §6 · p. 10

Linguistic Expressibility

• Language-based detection is most reliable when rule-critical attributes have stable, complete linguistic grounding, and degrades when relational attributes are expressed inconsistently across capture conditions.Source: §6 · pp. 10–11

Limitations and Future Directions

• Future work could use richer descriptions, reduce prompt dependence, and add structured text such as attribute-value tuples or constraint graphs to model logical rules more explicitly.The current single-text format uses descriptions generated from a fixed prompt; richer strategies may capture finer-grained details for borderline cases, while prompt-free or minimally prompted generation could improve reproducibility and deployment across manufacturing settings.Source: §6 · p. 11

• Representative vision-based baselines degrade under appearance variations, while the proposed method improves performance across all capture conditions, averaging AUROC 0.831 across 50 tasks.Source: §7 · p. 12
• VID-AD is a one-class benchmark comprising 50 tasks across 10 manufacturing scenarios and five capture conditions.Source: §7 · p. 12
• The language-based framework learns global logical consistency from text descriptions through contrastive learning between normal descriptions and semantically perturbed descriptions generated by constrained rewriting.It learns from text descriptions rather than visual features.Source: §7 · p. 12
• Future work will explore reducing prompt dependence to improve usability and robustness in real deployments.Source: §7 · p. 12

Declaration of competing interest

• The authors report no financial or nonfinancial interests, affiliations, or involvement related to the manuscript’s subject matter or materials.All authors certify this declaration.Source: Declaration of competing interest · p. 12