cs.CVOct 13, 2026
SaveRobust Logical Anomaly Detection under Capture Shifts via Synthetic Anomalies
Organizations: Graduate School of Informatics, Middle East Technical University, Üniversiteler Mahallesi, Dumlupınar Bulvarı No:1, Ankara, 06800, Türkiye
Abstract
Detectors of logical anomalies, such as missing, extra or wrongly combined parts, usually score a test image by its distance from normal (good) images. They achieve near-perfect image-level AUROC on standard industrial anomaly detection benchmarks such as MVTec LOCO, but their robustness to changes in capture conditions remains largely unexamined. Under a heterogeneous perturbation protocol in which each test image is randomly kept unaltered or subjected to a brightness reduction or a zoom-out transformation, the normal-reference methods we tested lose 9.7 to 34.1 AUROC points, and two of the strongest, CSAD and SALAD, fall from 97.7% and 96.4% to 68.1% and 62.3%. We propose a framework in which a classifier, trained on normal images and on synthetic anomalies, scores a test image without comparing it with normal data. From a short description of each product, a language model derives anomaly types and editing instructions, an image editor applies them to normal images, and the classifier learns to separate the edited images from automatically discovered normal sub-modes. With its full training recipe, the framework reaches 94.9% AUROC on clean images, 1.5 to 2.8 points below SALAD and CSAD, and 94.1% under the perturbation protocol. Trained without brightness or zoom-out augmentation, so that neither test perturbation occurs in training, it still reaches 92.5% (93.2% clean). These results suggest that clean benchmark scores can hide a strong sensitivity to capture conditions, and that learning the decision from both normal images and rule violations may reduce it.
Figures & tables
Figure 1: Clean and heterogeneous AUROC. Five-category mean image-level logical AUROC on MVTec LOCO for clean test images and under the heterogeneous perturbation protocol, in which each test image is randomly kept unaltered or subjected to a brightness reduction or a zoom-out transformation. Our two variants are shown in blue: Ours (full recipe) and Ours (no overlap), which is trained without brightness or zoom-out augmentation and is therefore used to assess robustness. AnomalyCLIP (below the dashed line), a zero-shot method that uses no normal reference, is near chance level even on clean images.
| AnomalyAny, AnomalyVFM, MIRAGE | ComGEN | CSAD, SALAD | Ours | |
|---|---|---|---|---|
| Target anomalies | Structural | Logical and structural | Mainly logical | Logical |
| Knowledge source | Defect lists from an LLM or VLM | Component tokens learned from normal images | Components segmented in normal images | Written description of the category |
| Generation | Diffusion generation, inpainting or editing | Token removal and low-density sampling in a diffusion model | Hand-designed edits of components or maps | Instruction-based editing of normal images into anomaly types |
| Test-time decision | Detector trained on the synthetic data | Segmentation from differences to the nearest normal image | Fused scores, partly normal-reference | Classifier trained on normal images and synthetic anomalies; scores a test image without comparing it with normal data |
Table 1: Synthetic anomalies in related methods and in our framework. Rows follow a synthetic anomaly from its source to the decision at test time. A normal-reference score depends, at least in part, on a comparison with normal training data. Our framework differs from the first column in the anomalies it targets and from the other two in its test-time decision. Columns, from left: AnomalyAny, AnomalyVFM and MIRAGE ( Sun et al., 2025 ; Fučka et al., 2026a ; Hu et al., 2026 ) ; ComGEN ( Tong et al., 2025 ) ; CSAD and SALAD ( Hsieh and Lai, 2024 ; Fučka et al., 2025 ) .
Figure 2: Overview of our framework, with an example from Pushpins. Top: a person describes the category; the language model proposes anomaly types that break the description and writes an editing prompt for each; the image editor applies a prompt to a normal image; curation keeps the valid edits (red ellipse: two pushpins in one compartment). Bottom: the classifier, trained on normal classes and the anomaly types, scores a test image on its own. Arrows show what passes between steps. Dashed orange boxes involve a person; gray boxes are pretrained models. The blue box is the only component used at test time. The edited image is generated by the image editor (Gemini 3 Pro Image), a generative AI model.
Figure 3: Normal images and synthetic anomalies. Each column shows a normal training image, the synthetic anomaly made from it (red ellipse: edited region) and its editing prompt. All five synthetic anomalies passed curation and are in the final training data. The synthetic anomalies are generated by the image editor (Gemini 3 Pro Image), a generative AI model.
Figure 4: Perturbation examples (Juice Bottle). Each row starts from the clean image. Top: brightness reduced by 20%, 40% and 60%. Bottom: zoom-out by 5%, 10% and 15% with edge-replicated borders.
| Hetero- | Drop | Relative | |||
| Method | Reported | Clean | geneous | (points) | drop (%) |
| Ours (no overlap) | – | 93.2 | 92.5 | 0.7 | 1.6 |
| Ours (full recipe) | – | 94.9 | 94.1 | 0.8 | 1.8 |
| CSAD | 96.7 | 97.7 | 68.1 | 29.6 | 62.1 |
| SALAD a | 96.5 | 96.4 | 62.3 | 34.1 | 73.5 |
| LogSAD | 89.3 | 89.2 | 73.9 | 15.3 | 39.0 |
Table 2: Main results. Clean and heterogeneous image-level logical AUROC (%), mean over the five categories.
| Per- | Hetero- | Within-condition | Cross-condition | Worst | |||
| Method | Clean | condition | geneous | drop | drop | condition | |
| Ours (no overlap) | 93.2 | 92.8 | 92.5 | 0.4 | 0.3 | 91.0 | Z15 |
| Ours (full recipe) | 94.9 | 94.1 | 94.1 | 0.8 | 0.0 | 93.0 | Z15 |
| CSAD | 97.7 | 75.0 | 68.1 | 22.7 | 6.9 | 61.9 | Z15 |
| SALAD | 96.4 | 85.6 | 62.3 | 10.8 | 23.3 | 78.0 | B40 |
| LogSAD | 89.2 | 77.5 | 73.9 | 11.7 | 3.6 | 67.6 | Z10 |
Table 3: Decomposition of the drop. Image-level logical AUROC (%), five-category mean. Per-condition: the mean AUROC of the six perturbed conditions, each scored on its own (Table 1 gives each condition). Within-condition drop = clean per-condition: separation lost even within one condition. Cross-condition drop = per-condition heterogeneous: lost mainly because the scores move between conditions. The two add up to the drop of Table 2 (Section 3.6 ); for each normal-reference method, the larger one is in bold. Worst condition: the perturbed condition with the lowest five-category mean (B40: brightness %; Z10: zoom-out 10%; Z15: zoom-out 15%).
Figure 5: Normal images scoring above the median clean anomaly. Each cell gives the share of normal (good) test images whose anomaly score is higher than the median score of the clean anomalous test images, as a five-category mean (%). The clean column is the starting point: a detector that separates clean images well starts near 0%, while PatchCore, which separates them less well, already starts at 12%. A detector whose normal scores do not move with the capture conditions keeps its clean value in every column; a rise means that a brightness reduction or a zoom-out alone moves normal images above the median clean anomaly. AnomalyCLIP is left out: near chance on clean images, it already scores 38% of its normal images above its median clean anomaly.
| Variant | Clean | Heterogeneous |
|---|---|---|
| Ours (full recipe) | 94.9 | 94.1 |
| Without normal sub-mode discovery | 92.9 | 91.7 |
| Ours (no overlap): without test-overlapping augmentations | 93.2 | 92.5 |
Table 4: Ablation. Image-level logical AUROC (%), mean over the five categories; each row below the first removes one component from the full recipe.
Appendix figures & tables16 assets
Supplementary material from the paper’s appendix.
Appendix
| Good | Test | Synthetic | ||||||
|---|---|---|---|---|---|---|---|---|
| Category | Train | Val | Test | logical | Train | Val | Types | |
| Breakfast Box | 351 | 62 | 102 | 83 | 311 | 55 | 1 | 19 |
| Juice Bottle | 335 | 54 | 94 | 142 | 329 | 53 | 4 | 20 |
| Pushpins | 372 | 69 | 138 | 91 | 273 | 51 | 1 | 17 |
| Screw Bag | 360 | 60 | 122 | 137 | 258 | 45 | 1 | 17 |
| Splicing Connectors | 360 | 60 | 119 | 108 | 292 | 48 | 3 | 18 |
Table 1: Data per category. Real MVTec LOCO normal (good) images of the train, validation and test splits and the real logical test anomalies; synthetic anomalies used for training and validation; the number of normal sub-modes ; and the number of the 20 designed anomaly types used in training.
| Setting | Value |
|---|---|
| Backbone | ConvNeXt-L ( Liu et al., 2022 ) , DINOv3 LVD-1689M weights ( timm : convnext_large.dinov3_lvd1689m ; Siméoni et al., 2025 ) |
| Fine-tuning | full network, no frozen layers |
| Head | global average pooling, LayerNorm, linear layer |
| Outputs | normal sub-modes 20 anomaly types (21 outputs; Juice Bottle 24, Splicing Connectors 23); a type not used as a class keeps its output but has no training images |
| Input | resized to , ImageNet mean and standard deviation |
| Optimizer | AdamW ( Loshchilov and Hutter, 2019 ) , , , weight decay 0.05 |
Table 1: Training settings of the classifier (all categories).
| Block | Operators and ranges | |
|---|---|---|
| Flips and rotation | 0.5 each | horizontal flip, vertical flip, rotation by a multiple of |
| Zoom-out* | 0.3 | downscale by 0–15%, edge-replicate padding back to |
| Blur (one of) | 0.5 | Gaussian ( 1–5), defocus (radius 2–3), motion (kernel 3–7), zoom blur (factor up to 1.05) |
| Noise (one of) | 0.5 | per-pixel multiplicative noise (0.7–1.3), Gaussian noise (std 0.01–0.05), JPEG compression, downscaling by 0.5 |
| Photometric* (one of) | 0.5 | brightness ( to of the image mean) and contrast ( ), CLAHE (clip limit 1–4), random tone curve, sharpening |
| Color (one of) | 0.25 | emboss, posterize (4 bits) |
Table 2: Augmentation pipeline (Albumentations), applied in this order to normal images and synthetic anomalies alike. A “one of” block applies one of its members with the given probability. The blocks marked * contain the test perturbations and are removed in the variant without test-overlapping augmentations. Validation and test images are only resized and normalized.
| ID | Anomaly type | Editing prompt |
|---|---|---|
| 1 | The tray is black. | Change the tray color to black. |
| 2 | The tray material is metal. | Change the tray material to metal. |
| 3 | The tray has three compartments. | Add a third compartment to the tray. |
| 4 | The left compartment is wider than the right compartment. | Resize the compartments so the left compartment is wider than the right compartment. |
| 5 | The left compartment contains two fruits. | Remove one fruit from the left compartment so it contains only two whole fruits. |
| 6 | The left compartment contains four fruits. | Add one extra whole fruit to the left compartment so it contains four whole fruits. |
Table 1: Breakfast Box: anomaly types and editing prompts.
Figure 1: Breakfast Box: example pairs. Normal training images (left) and synthetic anomalies made from them (right). Numbers are the type IDs of Table 1 . The synthetic anomalies are generated by the image editor (Gemini 3 Pro Image), a generative AI model (Section 3.2 ).
| ID | Anomaly type | Editing prompt |
|---|---|---|
| 1 | Bottle is opaque. | Make the bottle opaque. |
| 2 | Bottle contains no juice (empty or water). | Remove the juice, leaving the bottle empty. |
| 3 | Liquid level is far below ‘nearly full’. | Lower the liquid level to well below nearly full. |
| 4 | Bottle is completely full with no headspace. | Fill the bottle to the brim, removing all headspace. |
| 5 | Headspace appears at the bottom or side. | Relocate the headspace to the bottom of the bottle. |
| 6 | Bottle has a cap attached. | Add a cap to the bottle. |
Table 2: Juice Bottle: anomaly types and editing prompts.
Figure 2: Juice Bottle: example pairs. Normal training images (left) and synthetic anomalies made from them (right). Numbers are the type IDs of Table 2 . The synthetic anomalies are generated by the image editor (Gemini 3 Pro Image), a generative AI model (Section 3.2 ).
| ID | Anomaly type | Editing prompt |
|---|---|---|
| 1 | All compartments are empty; no pushpins are present in the box. | Remove all pushpins from every compartment so the entire box is empty. |
| 2 | The divider wall between two adjacent compartments is missing, merging them into one larger compartment. | Remove the divider wall between two adjacent compartments so they merge into one larger compartment. |
| 3 | The box has 15 compartments but they are not aligned into straight rows and columns. | Misalign the compartments so they are not in straight rows and columns. |
| 4 | One compartment contains no pushpin. | Remove the pushpin from one compartment leaving it empty. |
| 5 | Two compartments contain no pushpin. | Remove the pushpins from two compartments leaving them empty. |
| 6 | One compartment contains two pushpins. | Add an extra pushpin to one compartment so it contains two pushpins. |
Table 3: Pushpins: anomaly types and editing prompts.
Figure 3: Pushpins: example pairs. Normal training images (left) and synthetic anomalies made from them (right). Numbers are the type IDs of Table 3 . The synthetic anomalies are generated by the image editor (Gemini 3 Pro Image), a generative AI model (Section 3.2 ).
| ID | Anomaly type | Editing prompt |
|---|---|---|
| 1 | The long screw is significantly longer than its standard length. | Make the long screw significantly longer than its standard length. |
| 2 | The long screw is significantly shorter than its standard length. | Make the long screw significantly shorter than its standard length. |
| 3 | The short screw is significantly shorter than its standard length. | Make the short screw significantly shorter than its standard length. |
| 4 | The plastic bag is torn open while all six required components remain inside it. | Tear the plastic bag open while keeping all six components inside the bag. |
| 5 | Both screws are significantly shorter than their standard lengths. | Make both screws significantly shorter than their standard lengths. |
| 6 | Exactly one nut is inside the bag. | Remove one nut so that exactly one nut remains inside the bag. |
Table 4: Screw Bag: anomaly types and editing prompts.
Figure 4: Screw Bag: example pairs. Normal training images (left) and synthetic anomalies made from them (right). Numbers are the type IDs of Table 4 . The synthetic anomalies are generated by the image editor (Gemini 3 Pro Image), a generative AI model (Section 3.2 ).
| ID | Anomaly type | Editing prompt |
|---|---|---|
| 1 | Only one splicing connector is present. | Remove the right splicing connector so only one connector remains. |
| 2 | Two cables are present between the connectors. | Add a second cable connecting the two connectors. |
| 3 | No cable is present between the connectors. | Remove the cable so no cable is present between the connectors. |
| 4 | The cable is a color that is not used in normal samples. | Change the cable to a color that is not used in normal samples. |
| 5 | The two connectors are different types (different number of ports/levers). | Replace one connector with a connector that has a different number of ports/levers. |
| 6 | The cable color does not match the connector type. | Change the cable to a color that does not match the connector type. |
Table 5: Splicing Connectors: anomaly types and editing prompts.
Figure 5: Splicing Connectors: example pairs. Normal training images (left) and synthetic anomalies made from them (right). Numbers are the type IDs of Table 5 . The synthetic anomalies are generated by the image editor (Gemini 3 Pro Image), a generative AI model (Section 3.2 ).
| Category | Candidate | Gates: silhouette / smallest cluster / ARI | Chosen |
|---|---|---|---|
| Juice Bottle | 4 | : 0.53 / 11% / 0.94 pass | 4 |
| Splicing Connectors | 4 | : 0.20 / 14% / 0.94 fail (silhouette); : 0.28 / 33% / 1.00 pass | 3 |
| Screw Bag | 2 | : 0.31 / 11% / 0.66 fail (ARI) | 1 |
| Breakfast Box | 1 | single candidate, no gate | 1 |
| Pushpins | 1 | single candidate, no gate | 1 |
Table 1: Normal sub-mode discovery per category. Only normal training images are used. The candidate is chosen by 5-fold cross-validated log-likelihood with the one-standard-error rule; a candidate must pass three gates (silhouette , smallest cluster , median bootstrap ARI ), otherwise is tried.
| Method | Clean | B20 | B40 | B60 | Z5 | Z10 | Z15 |
|---|---|---|---|---|---|---|---|
| Ours (no overlap) | 93.2 | 93.2 | 93.4 | 93.8 | 92.8 | 92.4 | 91.0 |
| Ours (full recipe) | 94.9 | 94.8 | 94.7 | 94.4 | 94.1 | 93.6 | 93.0 |
| CSAD | 97.7 | 92.2 | 82.0 | 74.4 | 74.7 | 64.8 | 61.9 |
| SALAD | 96.4 | 89.4 | 78.0 | 80.8 | 94.7 | 87.3 | 83.4 |
| LogSAD | 89.2 | 87.5 | 84.0 | 80.0 | 75.7 | 67.6 | 70.5 |
| PatchCore | 75.4 | 75.4 | 74.3 | 72.7 | 67.2 | 62.8 | 57.3 |
Table 1: Results per condition. Image-level logical AUROC (%), each condition scored alone, mean over the five categories. B20/B40/B60: brightness reduced by 20/40/60%; Z5/Z10/Z15: zoom-out by 5/10/15% with edge replication.
| Breakfast Box | Juice Bottle | Pushpins | Screw Bag | Splicing Connectors | ||||||
| Method | Cl. | Het. | Cl. | Het. | Cl. | Het. | Cl. | Het. | Cl. | Het. |
| Ours (no overlap) | 92.8 | 93.3 | 92.7 | 92.3 | 99.5 | 99.0 | 91.3 | 90.1 | 89.5 | 87.8 |
| Ours (full recipe) | 97.1 | 96.4 | 97.0 | 96.4 | 100.0 | 99.8 | 90.1 | 87.7 | 90.3 | 90.2 |
| CSAD | 95.8 | 60.0 | 95.5 | 71.2 | 99.7 | 66.2 | 99.9 | 87.3 | 97.4 | 56.1 |
| SALAD | 92.3 | 61.0 | 99.8 | 66.0 | 100.0 | 68.0 | 94.0 | 60.4 | 96.1 | 55.8 |
| LogSAD | 97.1 | 70.4 | 94.6 | 81.7 | 84.0 | 82.7 | 79.9 | 62.5 | 90.2 | 72.4 |
Table 1: Results per category. Image-level logical AUROC (%): clean (Cl.) and heterogeneous (Het.). Ours (no overlap): trained without test-overlapping augmentations.