cs.CVFeb 23, 2025

A Survey on Industrial Anomaly Synthesis

Authors: Yanshu Wang, Xichen Xu, Jinbao Wang, Chengbin Ma, Xiaoning Lei, Guoyang Xie, Guannan Jiang, Zhichao Lu

Organizations: Shanghai Jiao Tong University · Department of Intelligent Manufacturing, CATL · City University of Hong Kong

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

This paper presents a comprehensive review of industrial anomaly synthesis (IAS). Existing surveys on industrial anomalies mainly focus on anomaly detection, while IAS is typically treated as an auxiliary component rather than as an independent topic. However, owing to its increasing importance in data augmentation, downstream model training, and controllable industrial inspection, IAS has become a research direction of growing interest. To address the lack of a dedicated review, we survey a broad range of representative methods and organize them into four paradigms: hand-crafted synthesis, distribution hypothesis-based synthesis, generative model (GM)-based synthesis, and vision-language model (VLM)-based synthesis. We further establish a dedicated taxonomy for IAS, which supports more systematic comparison across methods and offers a clearer view of the field's development. Beyond methodological categorization, we summarize the datasets, benchmarks, and evaluation metrics commonly adopted in IAS, and review recent advances in multimodal anomaly synthesis that remain underexplored in prior surveys. We also provide deployment-oriented comparisons and practical guidance by analyzing input requirements, output forms, controllability, cost, downstream tasks, and practical limitations across IAS subcategories. Overall, this survey provides a structured understanding of existing IAS methods, evaluation settings, practical trade-offs, current limitations, and promising future directions, and is intended to serve as a reference for subsequent research in this area. More resources are available at https://github.com/M-3LAB/awesome-anomaly-synthesis.

Figures & tables

Figure 1: Publication and citation trends of industrial anomaly detection and industrial anomaly synthesis from 2019 to 2026* (2026* denotes partial-year statistics up to March 30, 2026) . Industrial anomaly detection shows steady growth, while IAS has attracted increasing attention.
Ref.YearPerspectiveIAS Cat.MultimodalDescription
( Chen et al., 2021 )2021Detection0✗Visual inspection review with emphasis on augmentation and practical pipelines.
( Xia et al., 2022 )2022Detection4✗GAN-centric review covering GAN-based techniques related to AD and synthesis.
( Liu et al., 2024 )2024Detection2✗Broad anomaly detection survey; IAS is not a primary focus.
( Cao et al., 2024 )2024Detection4✗Discusses anomaly-related topics and mentions multimodal cues; IAS is not systematically reviewed.
( Mao et al., 2025 )2025Detection1✗Industrial anomaly detection survey covering paradigms, benchmarks, and metrics; includes synthesis-related discussion.
( Li et al., 2025 )2025Detection1✗Industrial visual anomaly detection survey covering learning strategies, generative modeling, and multimodal/VLM-based detection methods.
Table 1: Comparison of previous surveys and our survey.
Figure 2: Publication-source distribution of the 64 representative IAS methods reviewed in this survey. Conference papers constitute 59.4%, primarily from major venues such as CVPR, AAAI, ECCV, and ICCV. Journal articles account for 31.2% and span established venues including IEEE TIP, IEEE TII, IEEE TIM, IEEE TASE, and Pattern Recognition. The remaining 9.4% come from other sources, mainly preprints.
Figure 3: A taxonomy of industrial anomaly synthesis (IAS).
CriterionHand-craftedDistribution hypothesis-basedGM-basedVLM-based
Core mechanismRule-guided synthesis by manually specifying texture, location, or intensity.Distribution-guided synthesis by perturbing normal-sample distributions.Task-specific synthesis with generative models trained or adapted on industrial data.Multimodal semantic synthesis with large-scale pre-trained vision-language models.
Knowledge sourceHuman expertise, industrial priors, and image processing rules.Normal-sample distributions or constrained feature representations.Target-domain industrial data and task-specific anomaly synthesis priors learned during training.Pre-trained vision-language knowledge and industrial data.
Large-scale pre-trainingNot required.Usually not required; frozen feature extractors may be used for representation.Usually not required; synthesis ability mainly comes from target-domain training.Required; generality and semantic knowledge come from large-scale pre-training.
Task adaptationManual adjustment of synthesis rules.Estimation of normal patterns and construction of abnormal deviations.Dedicated training, re-training, fine-tuning, or domain adaptation for specific categories, datasets, or anomaly types.Prompting, instruction tuning, adapters, LoRA, textual inversion, or few-shot reference adaptation.
ModalityUnimodal, based on images.Unimodal, based on images or visual representations.Unimodal, based on images.Multimodal, involving images, text, or other reference cues.
GeneralizationRules need redesign across categories or anomaly types.Assumption-dependent; robustness decreases when normal patterns change.Good performance within the trained domain, but limited under cross-category shifts.Good potential for cross-category, few-shot, and semantically specified anomaly synthesis.
Table 2: Comparison of the four main paradigms for IAS.
DatasetYearVenueData typeSamplesClassesIAS relevanceSite
   (Real/Syn./Mixed)  HDHGMVLM 
MVTec-AD ( Bergmann et al., 2019 )2019CVPRReal5,35415✓✓✓✓link
VisA ( Zou et al., 2022 )2022ECCVReal10,82112✓✓✓✓link
BTAD ( Mishra et al., 2021 )2021ISIEReal2,8303✓✓✓✗link
MPDD ( Jezek et al., 2021 )2022ICUMTReal13466✓✓✓✗link
KSDD ( Tabernik et al., 2020 )2020JIMReal3991✓✓✓✗link
Table 3: Representative datasets commonly used in IAS-related studies. “Data type” indicates whether the benchmark is primarily based on real industrial imagery (Real), fully synthetic data (Syn.), or a mixture of real and synthetic components (Mixed).
MetricLevelFormulaRemarks/usage
Synthesis / Editing Metrics
FID↓\downarrow∥μr−μg∥22+Tr ⁣(Σr+Σg−2(ΣrΣg)1/2)\begin{aligned} \|\mu_{r}-\mu_{g}\|_{2}^{2}+\mathrm{Tr}\!\left(\Sigma_{r}+\Sigma_{g}-2(\Sigma_{r}\Sigma_{g})^{1/2}\right)\end{aligned}Distributional realism between real and generated features (often computed on full images or cropped regions / patches).
Reference-based LPIPS↓\downarrowd(x,y)d(x,y)Perceptual distance between a generated or edited image and a reference image; used for perceptual fidelity, editing consistency, or background preservation.
IC-LPIPS↑\uparrow1K∑k=1K2nk(nk−1)∑i<jnkd(xi(k),xj(k))\frac{1}{K}\sum_{k=1}^{K}\frac{2}{n_{k}(n_{k}-1)}\sum_{i<j}^{n_{k}}d(x_{i}^{(k)},x_{j}^{(k)})Intra-cluster LPIPS used for anomaly diversity in recent IAS evaluations. Generated anomalies are grouped by real anomaly references or clusters, and higher intra-cluster pairwise LPIPS indicates greater diversity.
PSNR↑\uparrow10log⁡10 ⁣(MAX2/MSE)10\log_{10}\!\left(\mathrm{MAX}^{2}/\mathrm{MSE}\right)Pixel-level reconstruction fidelity; often used to assess background preservation in local editing.
SSIM↑\uparrow(2μxμy+c1)(2σxy+c2)(μx2+μy2+c1)(σx2+σy2+c2)\dfrac{(2\mu_{x}\mu_{y}+c_{1})(2\sigma_{xy}+c_{2})}{(\mu_{x}^{2}+\mu_{y}^{2}+c_{1})(\sigma_{x}^{2}+\sigma_{y}^{2}+c_{2})}Structure preservation under synthesis or editing.
Table 4: Key metrics commonly used in IAS evaluation.
Figure 4: Hand-crafted synthesis. (a) Overview of three representative hand-crafted pipelines, including self-contained synthesis, external-dependent synthesis, and inpainting-based synthesis. These methods synthesize anomaly content through rule-based image manipulation without training a dedicated generative model. (b) Representative qualitative results adapted from ( Xu et al., 2025b ; Xu et al., 2025c ) , including examples of methods such as CutPaste ( Li et al., 2021 ) and DRAEM ( Zavrtanik et al., 2021a ) .
FamilyResource burdenPractical ceilingDeployment / limitation analysis
 TrainInferMemoryControlRealismSuitable scenarioPractical limitation
Self-containedN/ALowLowMediumLowRapid lightweight augmentationLimited diversity and synthetic artifacts
External-dependentN/ALowMediumMediumMediumBroad appearance diversificationBoundary inconsistency and semantic mismatch
Inpainting-basedN/ALowLowMediumLowReconstruction-oriented pipelinesMissing/corrupted appearance bias
Table 5: Practical profile of hand-crafted IAS
Figure 5: Distribution hypothesis-based synthesis. (a) Overview of the distribution hypothesis-based paradigm. It constructs abnormal samples or representations by introducing controlled deviations around the distribution of normal data, including prior-dependent strategies guided by explicit geometric assumptions and data-driven strategies based on direct perturbation, replacement, or sampling. (b) Representative qualitative illustrations adapted from ( Chen et al., 2024b ) .
FamilyResource burdenPractical ceilingDeployment / limitation analysis
 TrainInferMemoryControlRealismSuitable scenarioPractical limitation
Prior-dependentMediumLowMediumMediumLowDetection under structured normal distributionsUnstable normal-geometry assumption
Data-drivenMediumLowMediumLowLowDetection/localization with feature perturbationStrong reliance on feature quality
Table 6: Practical profile of distribution hypothesis-based IAS
Figure 6: GM-based synthesis. (a) Overview of the GM-based paradigm. It employs deep GMs, such as GANs and diffusion models, for IAS, including full-image synthesis, full-image translation, and local anomaly synthesis with region-wise editing. (b) Representative qualitative results adapted from SDGAN ( Niu et al., 2020 ) , RealNet ( Zhang et al., 2024 ) , and SARD ( Wang et al., 2025 ) .
FamilyResource burdenPractical ceilingDeployment / limitation analysis
 TrainInferMemoryControlRealismSuitable scenarioPractical limitation
Full-image synthesisMediumMediumMediumLowHighWhole-image anomaly dataset expansionMode collapse and domain overfitting
Full-image translationMediumMediumMediumLowHighNormal-to-abnormal source translationWeak spatial controllability
Local anomaly synthesisMediumHighHighHighHighSpatially controlled segmentation augmentationBoundary artifacts and mask dependence
Table 7: Practical profile of GM-based IAS
Figure 7: VLM-based synthesis. (a) Overview of the VLM-based paradigm. It leverages large-scale pre-trained vision–language or multimodal diffusion backbones together with text prompts, reference images, spatial conditions, and other auxiliary cues for IAS. Single-stage methods produce the final abnormal output through one unified dominant synthesis process, whereas multi-stage methods explicitly introduce multiple synthesis stages or coordinated sub-pipelines, such as mask synthesis, anomaly-part synthesis, or progressive refinement. (b) Representative qualitative results adapted from FAST ( Xu et al., 2025b ) , AnomalyDiffusion ( Hu et al., 2024b ) , and SeaS ( Dai et al., 2025 ) .
FamilyResource burdenPractical ceilingDeployment / limitation analysis
 TrainInferMemoryControlRealismSuitable scenarioPractical limitation
Single-stage VLM-basedMediumHighHighMediumHighMultimodal-conditioned anomaly synthesisSemantic misalignment and weak masks
Multi-stage VLM-basedHighHighHighHighHighStructured synthesis with aligned supervisionError propagation and pipeline complexity
Table 8: Practical profile of VLM-based IAS
ParadigmSubcategoryReal abnormal dataSpatial annotationText / multimodal promptReference inputSynthesis levelOutput form
Hand-crafted SynthesisSelf-contained–Optional––PixelAbnormal image + mask
 External-dependent–Optional–RequiredPixelAbnormal image + mask
 Inpainting-based–Optional––PixelAbnormal image + mask
Distribution hypothesis-based SynthesisPrior-dependent––––LatentAbnormal feature
 Data-driven––––LatentAbnormal feature
GM-based SynthesisFull-image synthesisRequiredOptional––Latent + pixelAbnormal image
Table 9: Input requirements and output forms of IAS subcategories.
ParadigmSubcategoryCostControllabilityDownstream tasksKey trade-off
Hand-crafted SynthesisSelf-containedLowMediumC/D/SGood source-background consistency Limited anomaly diversity
 External-dependentLowMediumC/D/SRicher anomaly appearance Higher risk of local inconsistency
 Inpainting-basedLowMediumC/D/SSimple masking-based synthesis Anomalies may look missing or corrupted
Distribution hypothesis-based SynthesisPrior-dependentMediumMediumC/DClear latent-space structure Limited flexibility for complex abnormal distributions
 Data-drivenMediumLowC/DAdaptive feature perturbation Strong reliance on feature quality
GM-based SynthesisFull-image synthesisMediumLowC/DHigh generation freedom Weaker preservation of source-image structure
Table 10: Deployment-oriented comparison of IAS subcategories.
More in this topic

Synthetic Anomaly Generation

Synthetic anomaly generation creates controlled abnormal examples to train, test, and evaluate detectors when real anomalies are scarce.

See all 65 papers

Explore similar work

CardsList
Apr 9, 2026

AnomalyAgent: Agentic Industrial Anomaly Synthesis via Tool-Augmented Reinforcement Learning

Industrial anomaly generation is a crucial method for alleviating the data scarcity problem in anomaly detection tasks. Most existing anomaly synthesis methods rely on single-step generation mechanisms, lacking complex reasoning and iterative optimization capabilities, making it difficult to generate anomaly samples with high semantic realism. We propose AnomalyAgent, an anomaly synthesis agent with self-reflection, knowledge retrieval, and iterative refinement capabilities, aiming to generate realistic and diverse anomalies. Specifically, AnomalyAgent is equipped with five tools: Prompt Generation (PG), Image Generation (IG), Quality Evaluation (QE), Knowledge Retrieval (KR), and Mask Generation (MG), enabling closed-loop optimization. To improve decision-making and self-reflection, we construct structured trajectories from real anomaly images and design a two-stage training framework: supervised fine-tuning followed by reinforcement learning. This process is driven by a three-part reward mechanism: (1) task rewards to supervise the quality and location rationality of generated anomalies; (2) reflection rewards to train the model's ability to improve anomaly synthesis prompt; (3) behavioral rewards to ensure adherence to the trajectory. On the MVTec-AD dataset, AnomalyAgent achieves IS/IC-L of 2.10/0.33 for anomaly generation, 57.0% classification accuracy using ResNet34, and 99.3%/74.2% AP at the image/pixel level using a simple UNet, surpassing all zero-shot SOTA methods. The code is publicly available at https://github.com/sssjmmm/AnomalyAgent.

Jiaming Su, Tengchao Yang, Ruikang Zhang +3

Shanghai Jiao Tong University Shanghai, China · Tongji University Shanghai, China · Fudan University Shanghai, China

Synthetic Anomaly GenerationTool-Using AgentsIndustrial Anomaly DetectionRL for Generative Models
AskPDF
Sep 29, 2026

FLASH: A "Generate Once, Synthesize Many" Framework for Synthetic Anomaly Generation in Industrial Anomaly Detection

Synthetic anomaly generation helps expand industrial anomaly datasets when real defects are scarce or unavailable. Existing approaches lie at two extremes: procedural approaches are fast but struggle to represent complex anomalies, while generative approaches produce diverse defects but require costly per-sample generation. We present FLASH, a framework that decouples defect generation from anomaly synthesis under a ``generate once, synthesize many'' paradigm. Given only normal images, FLASH uses Vision-Language Model (VLM) guidance and an image-generation model to produce a small set of defect images, from which it extracts, validates, and banks reusable defect patches. For synthesis of anomalous images, Object Boundary Suppression (OBS) first identifies the probable foreground object-aware region of the host image, while Multi-Resolution Spectral Pyramid (MRSP) noise generates diverse, size-controllable masks that determine the defect location and spatial extent. It then composes a large and diverse synthetic anomalous image set by localizing the defect region, sampling size-controllable placement masks and seamlessly blending retrieved defects onto new defect-free images without further need for image generation. Experiments on the MVTec AD 2 dataset show that FLASH-generated anomalies nearly close the calibration gap on real defects, reaching 78.1% image-level F1 against an 83.6% real-anomaly upper bound and providing the most consistent calibration transfer across detectors among procedural and generative alternatives. Moreover, FLASH synthesizes anomalies more than 11.95x faster than per-sample generative approaches.

Abhay Kumar Das, Rajesh Gangireddy, Ashwin Vaidya +1

Silicon University, Bhubaneswar, India · Intel

Synthetic Data AugmentationIndustrial Anomaly DetectionSynthetic Anomaly Generation
AskPDF
Aug 8, 2026

UniScale: Arbitrary-Scale Industrial Anomaly Generation

Industrial anomaly inspection faces a major challenge due to the lack of real-world anomaly samples. While generative models are used to create anomaly data, existing methods still struggle when handling small-scale anomalies. This failure occurs because extreme downsampling in diffusion models causes the information of small anomalies to be lost in the latent space. To address this, we introduce UniScale, a unified training and inference framework for high-fidelity industrial anomaly generation across arbitrary scales. During training, we introduce an Error-Suppressed Multi-Scale Training (EMT) strategy, which enables the model to learn the rich location-aware textures of anomalies, while suppressing upsampling-induced interpolation errors in texture acquisition, ensuring the model is capable of learning small-scale anomalies, while remaining effective for regular scale anomalies. For inference, we propose Generation-then-Fusion Denoising. It decouples anomaly generation from background integration, preventing small anomalies from being overwhelmed. Extensive experiments demonstrate that our method outperforms state-of-the-art competitors in both anomaly generation quality and downstream detection performance. It achieves a relative IS(a) improvement of 45.86% (from 1.81 to 2.64) on VisA and 37.70% (from 1.22 to 1.68) on MVTec AD 2, while also improving the downstream pixel-level IoU by 4.22% on VisA and AUROC by 6.55% on MVTec AD 2. Code is available at https://github.com/HUST-SLOW/UniScale.

Shilei Zeng, Linxin Guan, Xurui Li +2

School of Electronic Information and Communications, Huazhong University of Science and Technology

Industrial Anomaly DetectionSynthetic Anomaly GenerationDiffusion Models
AskPDF